fix: prevent in-place mutation of metadata dicts in _create_memory and _create_procedural_memory

Addresses #2648. The metadata dict passed by callers was mutated in-place
inside _create_memory and _create_procedural_memory (both sync and async),
causing unsafe behavior under concurrent/async usage.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -1,41 +1,56 @@
-name: 🐛 Bug Report
-description: Create a report to help us reproduce and fix the bug
+name: Bug Report
+description: Report a bug in mem0
+labels: ["bug"]
 
 body:
-- type: markdown
-  attributes:
-    value: >
-      #### Before submitting a bug, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/embedchain/embedchain/issues?q=is%3Aissue+sort%3Acreated-desc+).
-- type: textarea
-  attributes:
-    label: 🐛 Describe the bug
-    description: |
-      Please provide a clear and concise description of what the bug is.
+  - type: dropdown
+    id: component
+    attributes:
+      label: Component
+      description: Which part of mem0 is affected?
+      options:
+        - Core / Python SDK
+        - TypeScript SDK
+        - Vector Store (Qdrant, PGVector, Redis, Chroma, etc.)
+        - Graph Memory (Neo4j, Memgraph, etc.)
+        - Ollama / Local Models
+        - OpenMemory (MCP Server)
+        - OpenClaw
+        - REST API
+        - Other
+    validations:
+      required: true
 
-      If relevant, add a minimal example so that we can reproduce the error by running the code. It is very important for the snippet to be as succinct (minimal) as possible, so please take time to trim down any irrelevant code to help us debug efficiently. We are going to copy-paste your code and we expect to get the same result as you did: avoid any external data, and include the relevant imports, etc. For example:
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      value: |
+        ### Summary
 
-      ```python
-      # All necessary imports at the beginning
-      import embedchain as ec
-      # Your code goes here
+        A clear summary of the bug.
 
+        ### Steps to Reproduce
 
-      ```
+        ```python
+        from mem0 import Memory
 
-      Please also paste or describe the results you observe instead of the expected results. If you observe an error, please paste the error message including the **full** traceback of the exception. It may be relevant to wrap error messages in ```` ```triple quotes blocks``` ````.
-    placeholder: |
-      A clear and concise description of what the bug is.
+        m = Memory()
+        # Your code here...
+        ```
 
-      ```python
-      Sample code to reproduce the problem
-      ```
+        ### Expected Behavior
 
-      ```
-      The error message you got, with the full traceback.
-      ````
-  validations:
-    required: true
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
+        What you expected to happen.
+
+        ### Actual Behavior
+
+        What actually happened. Paste the full error traceback if applicable.
+
+        ### Environment
+
+        - mem0 version:
+        - Python/Node version:
+        - OS:
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -1,8 +1,8 @@
 blank_issues_enabled: true
 contact_links:
-  - name: 1-on-1 Session
-    url: https://cal.com/taranjeetio/ec
-    about: Speak directly with Taranjeet, the founder, to discuss issues, share feedback, or explore improvements for Embedchain
-  - name: Discord
+  - name: Discord Community
     url: https://discord.gg/6PzXDgEjG5
-    about: General community discussions
+    about: Ask questions and discuss with the community
+  - name: Documentation
+    url: https://docs.mem0.ai
+    about: Read the official mem0 documentation

.github/ISSUE_TEMPLATE/documentation_issue.yml

@@ -1,11 +1,23 @@
-name: Documentation
-description: Report an issue related to the Embedchain docs.
-title: "DOC: <Please write a comprehensive title after the 'DOC: ' prefix>"
+name: Documentation Issue
+description: Report an issue or suggest an improvement to the mem0 docs
+labels: ["documentation"]
 
 body:
-- type: textarea
-  attributes:
-    label: "Issue with current documentation:"
-    description: >
-      Please make sure to leave a reference to the document/code you're
-      referring to.
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      value: |
+        ### Page
+
+        Link to the docs page: https://docs.mem0.ai/...
+
+        ### What's Wrong or Missing
+
+        Describe what's incorrect, unclear, or missing.
+
+        ### Suggested Fix
+
+        How should the docs be improved?
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -1,23 +1,42 @@
-name: 🚀 Feature request
-description: Submit a proposal/request for a new Embedchain feature
+name: Feature Request
+description: Suggest a new feature or improvement for mem0
+labels: ["enhancement"]
 
 body:
-- type: textarea
-  id: feature-request
-  attributes:
-    label: 🚀 The feature
-    description: >
-      A clear and concise description of the feature proposal
-  validations:
-    required: true
-- type: textarea
-  attributes:
-    label: Motivation, pitch
-    description: >
-      Please outline the motivation for the proposal. Is your feature request related to a specific problem? e.g., *"I'm working on X and would like Y to be possible"*. If this is related to another GitHub issue, please link here too.
-  validations:
-    required: true
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
+  - type: dropdown
+    id: component
+    attributes:
+      label: Component
+      description: Which part of mem0 does this relate to?
+      options:
+        - Core / Python SDK
+        - TypeScript SDK
+        - Vector Store (Qdrant, PGVector, Redis, Chroma, etc.)
+        - Graph Memory (Neo4j, Memgraph, etc.)
+        - Ollama / Local Models
+        - OpenMemory (MCP Server)
+        - OpenClaw
+        - REST API
+        - Benchmarks / Evals
+        - Other
+    validations:
+      required: true
+
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      value: |
+        ### Use Case
+
+        What problem are you trying to solve?
+
+        ### Proposed Solution
+
+        How should this work? Include API examples or pseudocode if helpful.
+
+        ### Alternatives Considered
+
+        Any workarounds you've tried or other approaches considered.
+    validations:
+      required: true

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,41 +1,38 @@
+## Linked Issue
+
+Closes #<!-- issue number -->
+
 ## Description
 
-Please include a summary of the change and which issue is fixed. Please also include relevant motivation and context. List any dependencies that are required for this change.
+<!-- What does this PR do? Why is it needed? -->
 
-Fixes # (issue)
+## Type of Change
 
-## Type of change
-
-Please delete options that are not relevant.
-
-- [ ] Bug fix (non-breaking change which fixes an issue)
-- [ ] New feature (non-breaking change which adds functionality)
-- [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
-- [ ] Refactor (does not change functionality, e.g. code style improvements, linting)
+- [ ] Bug fix (non-breaking change that fixes an issue)
+- [ ] New feature (non-breaking change that adds functionality)
+- [ ] Breaking change (fix or feature that would cause existing functionality to change)
+- [ ] Refactor (no functional changes)
 - [ ] Documentation update
 
-## How Has This Been Tested?
+## Breaking Changes
 
-Please describe the tests that you ran to verify your changes. Provide instructions so we can reproduce. Please also list any relevant details for your test configuration
+<!-- If this is a breaking change, describe what breaks and the migration path. Delete this section if not applicable. -->
 
-Please delete options that are not relevant.
+N/A
 
-- [ ] Unit Test
-- [ ] Test Script (please provide)
+## Test Coverage
 
-## Checklist:
+- [ ] I added/updated unit tests
+- [ ] I added/updated integration tests
+- [ ] I tested manually (describe below)
+- [ ] No tests needed (explain why)
 
-- [ ] My code follows the style guidelines of this project
-- [ ] I have performed a self-review of my own code
-- [ ] I have commented my code, particularly in hard-to-understand areas
-- [ ] I have made corresponding changes to the documentation
-- [ ] My changes generate no new warnings
-- [ ] I have added tests that prove my fix is effective or that my feature works
-- [ ] New and existing unit tests pass locally with my changes
-- [ ] Any dependent changes have been merged and published in downstream modules
-- [ ] I have checked my code and corrected any misspellings
+<!-- Describe how you tested this, or link to CI results. -->
 
-## Maintainer Checklist
+## Checklist
 
-- [ ] closes #xxxx (Replace xxxx with the GitHub issue number)
-- [ ] Made sure Checks passed
+- [ ] My code follows the project's style guidelines
+- [ ] I have performed a self-review of my code
+- [ ] I have added tests that prove my fix/feature works
+- [ ] New and existing tests pass locally
+- [ ] I have updated documentation if needed

.github/advanced-issue-labeler.yml

@@ -0,0 +1,20 @@
+# Maps dropdown selections to GitHub labels
+# Used by the advanced-issue-labeler GitHub Action
+
+component:
+  - label: "sdk-python"
+    matcher: "Core / Python SDK"
+  - label: "sdk-typescript"
+    matcher: "TypeScript SDK"
+  - label: "vector-store"
+    matcher: "Vector Store"
+  - label: "graph-memory"
+    matcher: "Graph Memory"
+  - label: "ollama"
+    matcher: "Ollama"
+  - label: "openmemory"
+    matcher: "OpenMemory"
+  - label: "openclaw"
+    matcher: "OpenClaw"
+  - label: "rest-api"
+    matcher: "REST API"

.github/workflows/issue-labeler.yml

@@ -0,0 +1,39 @@
+name: Auto-label issues
+
+on:
+  issues:
+    types: [opened]
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  label:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: stefanbuck/github-issue-parser@v3
+        id: issue-parser
+        with:
+          template-path: .github/ISSUE_TEMPLATE/bug_report.yml
+
+      - uses: redhat-plumbers-in-action/advanced-issue-labeler@v3
+        with:
+          issue-form: ${{ steps.issue-parser.outputs.jsonString }}
+          section: component
+          token: ${{ secrets.GITHUB_TOKEN }}
+          config-path: .github/advanced-issue-labeler.yml
+
+      - uses: stefanbuck/github-issue-parser@v3
+        id: feature-parser
+        if: contains(github.event.issue.labels.*.name, 'enhancement')
+        with:
+          template-path: .github/ISSUE_TEMPLATE/feature_request.yml
+
+      - uses: redhat-plumbers-in-action/advanced-issue-labeler@v3
+        if: contains(github.event.issue.labels.*.name, 'enhancement')
+        with:
+          issue-form: ${{ steps.feature-parser.outputs.jsonString }}
+          section: component
+          token: ${{ secrets.GITHUB_TOKEN }}
+          config-path: .github/advanced-issue-labeler.yml

.github/workflows/stale.yml

@@ -0,0 +1,48 @@
+name: Close stale issues
+
+on:
+  schedule:
+    - cron: '0 0 * * *'
+  workflow_dispatch:
+
+permissions:
+  issues: write
+  pull-requests: write
+
+jobs:
+  stale:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/stale@v9
+        with:
+          # Issue settings
+          days-before-issue-stale: 90
+          days-before-issue-close: 14
+          stale-issue-label: 'stale'
+          stale-issue-message: >
+            This issue has been automatically marked as stale because it has not
+            had any activity in 90 days. It will be closed in 14 days if no
+            further activity occurs. If this is still relevant, please leave a
+            comment or remove the `stale` label.
+          close-issue-message: >
+            This issue has been closed due to inactivity. If this is still
+            relevant, feel free to reopen it or create a new issue.
+
+          # PR settings — mark stale but never auto-close
+          days-before-pr-stale: 90
+          days-before-pr-close: -1
+          stale-pr-label: 'stale'
+          stale-pr-message: >
+            This pull request has been automatically marked as stale because it
+            has not had any activity in 90 days. Please update your branch and
+            address any review comments, or it may be closed in the future.
+
+          # Exempt these labels from stale processing
+          exempt-issue-labels: 'P0-critical,P1-high,good first issue,security'
+          exempt-pr-labels: 'P0-critical,P1-high'
+
+          # Remove stale label when there is new activity
+          remove-stale-when-updated: true
+
+          # Process up to 100 issues per run to stay within API limits
+          operations-per-run: 100

LLM.md

@@ -273,7 +273,7 @@ config = MemoryConfig(
 
 ### Supported Providers
 
-#### LLM Providers (19 supported)
+#### LLM Providers (20 supported)
 - **openai** - OpenAI GPT models (default)
 - **anthropic** - Claude models
 - **gemini** - Google Gemini
@@ -284,6 +284,7 @@ config = MemoryConfig(
 - **azure_openai** - Azure OpenAI
 - **litellm** - LiteLLM proxy
 - **deepseek** - DeepSeek models
+- **minimax** - MiniMax models
 - **xai** - xAI models
 - **sarvam** - Sarvam AI
 - **lmstudio** - LM Studio local server

docs/components/llms/models/minimax.mdx

@@ -0,0 +1,56 @@
+---
+title: MiniMax
+description: "Configure MiniMax as an LLM provider in Mem0 with API key setup and optional custom endpoint configuration."
+---
+
+To use MiniMax LLM models, you have to set the `MINIMAX_API_KEY` environment variable. You can also optionally set `MINIMAX_API_BASE` if you need to use a different API endpoint (defaults to "https://api.minimax.io/v1").
+
+## Usage
+
+```python
+import os
+from mem0 import Memory
+
+os.environ["MINIMAX_API_KEY"] = "your-api-key"
+os.environ["OPENAI_API_KEY"] = "your-api-key" # for embedder model
+
+config = {
+    "llm": {
+        "provider": "minimax",
+        "config": {
+            "model": "MiniMax-M2.7",  # default model
+            "temperature": 0.2,
+            "max_tokens": 2000,
+            "top_p": 1.0
+        }
+    }
+}
+
+m = Memory.from_config(config)
+messages = [
+    {"role": "user", "content": "I'm planning to watch a movie tonight. Any recommendations?"},
+    {"role": "assistant", "content": "How about thriller movies? They can be quite engaging."},
+    {"role": "user", "content": "I'm not a big fan of thriller movies but I love sci-fi movies."},
+    {"role": "assistant", "content": "Got it! I'll avoid thriller recommendations and suggest sci-fi movies in the future."}
+]
+m.add(messages, user_id="alice", metadata={"category": "movies"})
+```
+
+You can also configure the API base URL in the config:
+
+```python
+config = {
+    "llm": {
+        "provider": "minimax",
+        "config": {
+            "model": "MiniMax-M2.7",
+            "minimax_base_url": "https://your-custom-endpoint.com",
+            "api_key": "your-api-key"  # alternatively to using environment variable
+        }
+    }
+}
+```
+
+## Config
+
+All available parameters for the `minimax` config are present in [Master List of All Params in Config](../config).

docs/components/llms/overview.mdx

@@ -31,6 +31,7 @@ See the list of supported LLMs below.
   <Card title="Google AI" href="/components/llms/models/google_AI" />
   <Card title="AWS bedrock" href="/components/llms/models/aws_bedrock" />
   <Card title="DeepSeek" href="/components/llms/models/deepseek" />
+  <Card title="MiniMax" href="/components/llms/models/minimax" />
   <Card title="xAI" href="/components/llms/models/xAI" />
   <Card title="Sarvam AI" href="/components/llms/models/sarvam" />
   <Card title="LM Studio" href="/components/llms/models/lmstudio" />

docs/components/vectordbs/dbs/databricks.mdx

@@ -17,8 +17,10 @@ config = {
             "workspace_url": "https://your-workspace.databricks.com",
             "access_token": "your-access-token",
             "endpoint_name": "your-vector-search-endpoint",
-            "index_name": "catalog.schema.index_name",
-            "source_table_name": "catalog.schema.source_table",
+            "catalog": "your_catalog",
+            "schema": "your_schema",
+            "table_name": "your_table",
+            "collection_name": "your_index_name",
             "embedding_dimension": 1536
         }
     }
@@ -42,17 +44,22 @@ Here are the parameters available for configuring Databricks Vector Search:
 | --- | --- | --- |
 | `workspace_url` | The URL of your Databricks workspace | **Required** |
 | `access_token` | Personal Access Token for authentication | `None` |
-| `service_principal_client_id` | Service principal client ID (alternative to access_token) | `None` |
-| `service_principal_client_secret` | Service principal client secret (required with client_id) | `None` |
+| `client_id` | Service principal client ID (alternative to access_token) | `None` |
+| `client_secret` | Service principal client secret (required with client_id) | `None` |
+| `azure_client_id` | Azure AD application client ID (for Azure Databricks) | `None` |
+| `azure_client_secret` | Azure AD application client secret (for Azure Databricks) | `None` |
 | `endpoint_name` | Name of the Vector Search endpoint | **Required** |
-| `index_name` | Name of the vector index (Unity Catalog format: catalog.schema.index) | **Required** |
-| `source_table_name` | Name of the source Delta table (Unity Catalog format: catalog.schema.table) | **Required** |
-| `embedding_dimension` | Dimension of self-managed embeddings | `1536` |
-| `embedding_source_column` | Column name for text when using Databricks-computed embeddings | `None` |
+| `catalog` | Unity Catalog catalog name | **Required** |
+| `schema` | Unity Catalog schema name | **Required** |
+| `table_name` | Source Delta table name | **Required** |
+| `collection_name` | Vector search index name | `mem0` |
+| `index_type` | Index type: `DELTA_SYNC` or `DIRECT_ACCESS` | `DELTA_SYNC` |
 | `embedding_model_endpoint_name` | Databricks serving endpoint for embeddings | `None` |
-| `embedding_vector_column` | Column name for self-managed embedding vectors | `embedding` |
+| `embedding_dimension` | Dimension of self-managed embeddings | `1536` |
 | `endpoint_type` | Type of endpoint (`STANDARD` or `STORAGE_OPTIMIZED`) | `STANDARD` |
-| `sync_computed_embeddings` | Whether to sync computed embeddings automatically | `True` |
+| `pipeline_type` | Sync pipeline type: `TRIGGERED` or `CONTINUOUS` | `TRIGGERED` |
+| `warehouse_name` | Databricks SQL warehouse name (if using SQL warehouse) | `None` |
+| `query_type` | Query type: `ANN` or `HYBRID` | `ANN` |
 
 ### Authentication
 
@@ -65,11 +72,13 @@ config = {
         "provider": "databricks",
         "config": {
             "workspace_url": "https://your-workspace.databricks.com",
-            "service_principal_client_id": "your-service-principal-id",
-            "service_principal_client_secret": "your-service-principal-secret",
+            "client_id": "your-service-principal-id",
+            "client_secret": "your-service-principal-secret",
             "endpoint_name": "your-endpoint",
-            "index_name": "catalog.schema.index_name",
-            "source_table_name": "catalog.schema.source_table"
+            "catalog": "your_catalog",
+            "schema": "your_schema",
+            "table_name": "your_table",
+            "collection_name": "your_index_name",
         }
     }
 }
@@ -84,8 +93,10 @@ config = {
             "workspace_url": "https://your-workspace.databricks.com",
             "access_token": "your-personal-access-token",
             "endpoint_name": "your-endpoint",
-            "index_name": "catalog.schema.index_name",
-            "source_table_name": "catalog.schema.source_table"
+            "catalog": "your_catalog",
+            "schema": "your_schema",
+            "table_name": "your_table",
+            "collection_name": "your_index_name",
         }
     }
 }
@@ -103,7 +114,6 @@ config = {
         "config": {
             # ... authentication config ...
             "embedding_dimension": 768,  # Match your embedding model
-            "embedding_vector_column": "embedding"
         }
     }
 }
@@ -118,7 +128,6 @@ config = {
         "provider": "databricks",
         "config": {
             # ... authentication config ...
-            "embedding_source_column": "text",
             "embedding_model_endpoint_name": "e5-small-v2"
         }
     }
@@ -127,8 +136,8 @@ config = {
 
 ### Important Notes
 
-- **Delta Sync Index**: This implementation uses Delta Sync Index, which automatically syncs with your source Delta table. Direct vector insertion/deletion/update operations will log warnings as they're not supported with Delta Sync.
-- **Unity Catalog**: Both the source table and index must be in Unity Catalog format (`catalog.schema.table_name`).
+- **Index Types**: This implementation supports both `DELTA_SYNC` (auto-syncs with source Delta table) and `DIRECT_ACCESS` (manage vectors directly) index types.
+- **Unity Catalog**: The source table and index are created under the specified `catalog.schema` namespace.
 - **Endpoint Auto-Creation**: If the specified endpoint doesn't exist, it will be created automatically.
 - **Index Auto-Creation**: If the specified index doesn't exist, it will be created automatically with the provided configuration.
 - **Filter Support**: Supports filtering by metadata fields, with different syntax for STANDARD vs STORAGE_OPTIMIZED endpoints.

docs/components/vectordbs/dbs/pgvector.mdx

@@ -48,7 +48,7 @@ const config = {
       password: '123',
       host: '127.0.0.1',
       port: 5432,
-      dbname: 'vector_store', // Optional, defaults to 'postgres'
+      dbname: 'vector_store', // Optional; TypeScript OSS defaults to `vector_store` when omitted
       diskann: false, // Optional, requires pgvectorscale extension
       hnsw: false, // Optional, for HNSW indexing
     },
@@ -85,6 +85,8 @@ Here are the parameters available for configuring pgvector:
 | `connection_string` | PostgreSQL connection string (overrides individual connection parameters) | `None` |
 | `connection_pool` | psycopg2 connection pool object (overrides connection string and individual parameters) | `None` |
 
+**Note (TypeScript OSS):** If you omit `dbname`, the TypeScript client uses the database name `vector_store`. Python defaults to `postgres` for `dbname`, as in the table above.
+
 **Note**: The connection parameters have the following priority:
 1. `connection_pool` (highest priority)
 2. `connection_string`

docs/components/vectordbs/dbs/turbopuffer.mdx

@@ -0,0 +1,79 @@
+---
+title: "Turbopuffer"
+description: "Use Turbopuffer as a serverless vector database in Mem0 for low-latency search at scale with native metadata filtering."
+---
+[Turbopuffer](https://turbopuffer.com) is a serverless vector database optimized for low-latency search at scale. It offers cost-effective vector storage with native metadata filtering.
+
+### Usage
+
+```python
+import os
+from mem0 import Memory
+
+os.environ["OPENAI_API_KEY"] = "sk-xx"
+os.environ["TURBOPUFFER_API_KEY"] = "tpuf_xxxxxxxxxxxx"
+
+config = {
+    "vector_store": {
+        "provider": "turbopuffer",
+        "config": {
+            "collection_name": "movie_preferences",
+            "embedding_model_dims": 1536,
+            "region": "gcp-us-central1",
+        }
+    }
+}
+
+m = Memory.from_config(config)
+
+messages = [
+    {"role": "user", "content": "I'm planning to watch a movie tonight. Any recommendations?"},
+    {"role": "assistant", "content": "How about thriller movies? They can be quite engaging."},
+    {"role": "user", "content": "I'm not a big fan of thrillers but I love sci-fi."},
+    {"role": "assistant", "content": "Got it! I'll suggest sci-fi movies instead."}
+]
+
+m.add(messages, user_id="alice", metadata={"category": "movies"})
+
+# Search memories
+results = m.search(query="sci-fi recommendations", user_id="alice")
+```
+
+### Config
+
+Here are the parameters available for configuring Turbopuffer:
+
+| Parameter | Description | Default Value |
+| --- | --- | --- |
+| `collection_name` | Name of the namespace/collection | `mem0` |
+| `embedding_model_dims` | Dimensions of the embedding model (must match your chosen embedding model) | `1536` |
+| `api_key` | Turbopuffer API key | Environment variable: `TURBOPUFFER_API_KEY` |
+| `region` | Turbopuffer region | `gcp-us-central1` |
+| `distance_metric` | Distance metric for vector similarity (`cosine_distance` or `euclidean_squared`) | `cosine_distance` |
+| `batch_size` | Batch size for bulk operations | `100` |
+| `extra_params` | Additional parameters for the Turbopuffer client | `None` |
+
+### Regions
+
+| Region | Location |
+| --- | --- |
+| `gcp-us-central1` | Iowa, USA (Default) |
+| `aws-us-west-2` | Oregon, USA |
+
+### Config Example
+
+```python
+config = {
+    "vector_store": {
+        "provider": "turbopuffer",
+        "config": {
+            "collection_name": "my_memories",
+            "embedding_model_dims": 1536,
+            "api_key": "tpuf_xxxxxxxxxxxx",
+            "region": "aws-us-west-2",
+            "distance_metric": "cosine_distance",
+            "batch_size": 200,
+        }
+    }
+}
+```

docs/components/vectordbs/overview.mdx

@@ -33,6 +33,7 @@ See the list of supported vector databases below.
   <Card title="LangChain" href="/components/vectordbs/dbs/langchain"></Card>
   <Card title="Amazon S3 Vectors" href="/components/vectordbs/dbs/s3_vectors"></Card>
   <Card title="Databricks" href="/components/vectordbs/dbs/databricks"></Card>
+  <Card title="Turbopuffer" href="/components/vectordbs/dbs/turbopuffer"></Card>
 </CardGroup>
 
 ## Usage

docs/docs.json

@@ -40,6 +40,7 @@
                     "icon": "rocket",
                     "pages": [
                       "platform/overview",
+                      "vibecoding",
                       "platform/mem0-mcp",
                       "platform/platform-vs-oss",
                       "platform/quickstart"
@@ -142,6 +143,7 @@
                     "icon": "rocket",
                     "pages": [
                       "open-source/overview",
+                      "vibecoding",
                       "open-source/python-quickstart",
                       "open-source/node-quickstart"
                     ]
@@ -188,6 +190,7 @@
                               "components/llms/models/google_AI",
                               "components/llms/models/aws_bedrock",
                               "components/llms/models/deepseek",
+                              "components/llms/models/minimax",
                               "components/llms/models/xAI",
                               "components/llms/models/sarvam",
                               "components/llms/models/lmstudio",
@@ -230,7 +233,8 @@
                               "components/vectordbs/dbs/cassandra",
                               "components/vectordbs/dbs/s3_vectors",
                               "components/vectordbs/dbs/databricks",
-                              "components/vectordbs/dbs/neptune_analytics"
+                              "components/vectordbs/dbs/neptune_analytics",
+                              "components/vectordbs/dbs/turbopuffer"
                             ]
                           }
                         ]
@@ -300,6 +304,7 @@
                     "icon": "square-terminal",
                     "pages": [
                       "openmemory/overview",
+                      "vibecoding",
                       "openmemory/quickstart",
                       "openmemory/integrations"
                     ]

docs/open-source/features/graph-memory.mdx

@@ -35,7 +35,7 @@ graph LR
 Mem0’s extraction LLM identifies entities, relationships, and timestamps from the conversation payload you send to `memory.add`.
 </Step>
 <Step title="Store vectors and edges together">
-Embeddings land in your configured vector database while nodes and edges flow into a Bolt-compatible graph backend (Neo4j, Memgraph, Neptune, or Kuzu).
+Embeddings land in your configured vector database while nodes and edges flow into a graph backend (Neo4j, Memgraph, Neptune, Kuzu, or Apache AGE).
 </Step>
 <Step title="Expose graph context at search time">
 `memory.search` performs vector similarity (optionally reranked by your configured reranker) and returns the results list. Graph Memory runs in parallel and adds related entities in the `relations` array—it does not reorder the vector hits automatically.
@@ -264,7 +264,7 @@ Monitor graph growth, especially on free tiers, by periodically cleaning dormant
 
 ## Decision Points
 
-- Select the graph store that fits your deployment (managed Aura vs. self-hosted Neo4j vs. AWS Neptune vs. local Kuzu).
+- Select the graph store that fits your deployment (managed Aura vs. self-hosted Neo4j vs. AWS Neptune vs. local Kuzu vs. Apache AGE on PostgreSQL).
 - Decide when to enable graph writes per request; routine conversations may stay vector-only to save latency.
 - Set a policy for pruning stale relationships so your graph stays fast and affordable.
 
@@ -381,6 +381,42 @@ config = {
 
 Kuzu will clear its state when using `:memory:` once the process exits. See the [Kuzu documentation](https://kuzudb.com/docs/) for advanced settings.
   </Accordion>
+  <Accordion title="Apache AGE (PostgreSQL extension)">
+    [Apache AGE](https://age.apache.org/) adds graph database capabilities to PostgreSQL, letting you run Cypher queries alongside SQL on the same server. Start AGE via Docker, then configure Mem0:
+
+```bash
+docker run --name age-postgres \
+  -e POSTGRES_DB=mem0_db \
+  -e POSTGRES_USER=mem0_user \
+  -e POSTGRES_PASSWORD=mem0_pass \
+  -p 5432:5432 \
+  -d apache/age
+```
+
+```python
+from mem0 import Memory
+
+config = {
+    "graph_store": {
+        "provider": "apache_age",
+        "config": {
+            "host": "localhost",
+            "port": 5432,
+            "database": "mem0_db",
+            "username": "mem0_user",
+            "password": "mem0_pass",
+            "graph_name": "mem0_graph",
+        },
+    },
+}
+
+m = Memory.from_config(config_dict=config)
+```
+
+Apache AGE does not have a built-in vector index, so similarity search is computed client-side. This works well for moderate graph sizes; for very large graphs consider pairing AGE with pgvector for the vector store.
+
+Reference: [Apache AGE documentation](https://age.apache.org/age-manual/master/index.html).
+  </Accordion>
 </AccordionGroup>
 
 <CardGroup cols={2}>

docs/open-source/features/rest-api.mdx

@@ -14,7 +14,7 @@ The Mem0 REST API server exposes every OSS memory operation over HTTP. Run it al
 </Info>
 
 <Warning>
-  Add your own authentication and HTTPS before exposing the server to anything beyond your internal network. The default image does not include auth.
+  Enable API key authentication (see below) and HTTPS before exposing the server to anything beyond your internal network.
 </Warning>
 
 ---
@@ -22,6 +22,7 @@ The Mem0 REST API server exposes every OSS memory operation over HTTP. Run it al
 ## Feature
 
 - **CRUD endpoints:** Create, retrieve, search, update, delete, and reset memories by `user_id`, `agent_id`, or `run_id`.
+- **API key authentication:** Optionally secure all endpoints with a shared API key via the `X-API-Key` header.
 - **Status health check:** Access base routes to confirm the server is online.
 - **OpenAPI explorer:** Visit `/docs` for interactive testing and schema reference.
 
@@ -91,6 +92,41 @@ uvicorn main:app --reload
 
 ---
 
+## Authentication
+
+The server supports optional API key authentication. When the `ADMIN_API_KEY` environment variable is set, every endpoint requires a valid `X-API-Key` header. The `/` redirect, `/docs`, and `/openapi.json` routes remain open so you can always reach the interactive API explorer.
+
+| `ADMIN_API_KEY` value | Behavior |
+|---|---|
+| Not set / empty | All endpoints are open (no auth) |
+| Any non-empty string | Requests must include `X-API-Key: <your-key>` |
+
+### Enable authentication
+
+Add the key to your `.env` file:
+
+```bash
+ADMIN_API_KEY=your-secret-api-key
+```
+
+Then include the header in every request:
+
+```bash
+curl -X POST http://localhost:8000/memories \
+  -H "Content-Type: application/json" \
+  -H "X-API-Key: your-secret-api-key" \
+  -d '{
+    "messages": [{"role": "user", "content": "I love pizza."}],
+    "user_id": "alice"
+  }'
+```
+
+<Warning>
+  The server logs a warning at startup when `ADMIN_API_KEY` is not set. Always set it in production.
+</Warning>
+
+---
+
 ## See it in action
 
 ### Create and search memories via HTTP
@@ -137,7 +173,7 @@ curl "http://localhost:8000/memories/search?user_id=alice&query=vegetable"
 
 ## Best practices
 
-1. **Add authentication:** Protect endpoints with API gateways, proxies, or custom FastAPI middleware.
+1. **Enable authentication:** Set `ADMIN_API_KEY` to secure all endpoints, or use an API gateway for more advanced schemes.
 2. **Use HTTPS:** Terminate TLS at your load balancer or reverse proxy.
 3. **Monitor uptime:** Track request rates, latency, and error codes per endpoint.
 4. **Version configs:** Keep environment files and Docker Compose definitions in source control.

docs/openmemory/integrations.mdx

@@ -47,6 +47,7 @@ Replace `<client-name>` with the desired client name and `<user-id>` with the va
 | Witsy       | `npx install-mcp http://localhost:8765/mcp/witsy/sse/<user-id> --client witsy` |
 | Enconvo     | `npx install-mcp http://localhost:8765/mcp/enconvo/sse/<user-id> --client enconvo` |
 | Augment     | `npx install-mcp http://localhost:8765/mcp/augment/sse/<user-id> --client augment` |
+| Claude Code | `claude mcp add -s user -t sse OpenMemory http://localhost:8765/mcp/openmemory/sse/<user-id>` |
 
 ### What This Does
 

docs/openmemory/quickstart.mdx

@@ -183,6 +183,11 @@ You can configure the MCP client using the following command (replace `username`
 npx @openmemory/install local "http://localhost:8765/mcp/cursor/sse/username" --client cursor
 ```
 
+**For Claude Code (CLI):**
+```bash
+claude mcp add -s user -t sse OpenMemory "http://localhost:8765/mcp/openmemory/sse/username"
+```
+
 The OpenMemory dashboard will be available at http://localhost:3000. From here, you can view and manage your memories and check connection status with your MCP clients.
 
 Once set up, OpenMemory runs locally on your machine, ensuring all your AI memories remain private and secure while being accessible across any compatible MCP client.

docs/templates/cookbook_template.mdx

@@ -166,13 +166,13 @@ Call out the most common mistake or edge case for this layer.
     title="[Related cookbook / deep dive]"
     description="[Why this pairs well with the current guide]"
     icon="arrow-right"
-    href="/[related-link]"
+    href="#related-link"
   />
   <Card
     title="[Next cookbook in journey]"
     description="[Set expectation for the next step]"
     icon="rocket"
-    href="/[next-link]"
+    href="#next-link"
   />
 </CardGroup>
 ```

docs/templates/migration_guide_template.mdx

@@ -145,13 +145,13 @@ npm install mem0ai@[version]
     title="[Deep dive reference]"
     description="[Why this reference matters post-migration]"
     icon="book"
-    href="/[reference-link]"
+    href="#reference-link"
   />
   <Card
     title="[Applied example or next step]"
     description="[What readers can build now]"
     icon="rocket"
-    href="/[example-link]"
+    href="#example-link"
   />
 </CardGroup>
 ```

docs/vibecoding.mdx

@@ -0,0 +1,181 @@
+---
+title: "Vibecoding with Mem0"
+sidebarTitle: "Vibecoding"
+description: "Agent skills, starter prompts, and setup for building with Mem0 using AI coding tools."
+icon: "wand-magic-sparkles"
+---
+
+These docs are designed to be easily consumable by LLMs. Each page has a button that lets you copy the page as Markdown or paste directly into ChatGPT, Claude, or any AI coding tool.
+
+We follow the llms.txt standard:
+
+- [llms.txt](https://docs.mem0.ai/llms.txt)
+
+<CardGroup cols={2}>
+  <Card title="Get an API Key" icon="key" href="https://app.mem0.ai">
+    Sign up for Mem0 Platform and start building
+  </Card>
+  <Card title="Quickstart" icon="rocket" href="/platform/quickstart">
+    Store your first memory in under 5 minutes
+  </Card>
+</CardGroup>
+
+## Agent Skills
+
+Teach your coding assistant how to build with Mem0:
+
+```bash
+npx skills add https://github.com/mem0ai/mem0 --skill mem0
+```
+
+Works with Claude Code, Cursor, Windsurf, and any assistant that supports skills. Once installed, your assistant understands Mem0's full API, framework integrations, and common patterns.
+
+## Claude Code Plugin
+
+The [OpenMemory plugin](https://github.com/mem0ai/claude-code-plugin) gives Claude Code **persistent memory across sessions, projects, and teams** — automatically.
+
+<Steps>
+<Step title="Get your API key">
+Sign up at [app.openmemory.dev](https://app.openmemory.dev).
+</Step>
+
+<Step title="Install the plugin">
+```bash
+/plugin add mem0ai/claude-code-plugin
+```
+</Step>
+
+<Step title="Set your environment variable">
+```bash
+export OPENMEMORY_API_KEY="your-key-here"
+```
+</Step>
+
+<Step title="Start coding">
+The plugin activates automatically. It captures decisions at session end, preserves context during compaction, and retrieves relevant memories at session start.
+</Step>
+</Steps>
+
+## MCP Server Setup
+
+Connect Cursor, Windsurf, Claude Desktop, or any MCP-compatible client to Mem0.
+
+<Tabs>
+<Tab title="OpenMemory Hosted">
+Sign up at [app.openmemory.dev](https://app.openmemory.dev), then pick your client:
+
+<CodeGroup>
+```bash Claude Desktop
+npx @openmemory/install --client claude --env OPENMEMORY_API_KEY=your-key
+```
+
+```bash Cursor
+npx @openmemory/install --client cursor --env OPENMEMORY_API_KEY=your-key
+```
+
+```bash Windsurf
+npx @openmemory/install --client windsurf --env OPENMEMORY_API_KEY=your-key
+```
+</CodeGroup>
+
+For full setup options, see [OpenMemory Quickstart](/openmemory/quickstart).
+</Tab>
+
+<Tab title="Mem0 Platform MCP">
+Get your API key from [app.mem0.ai](https://app.mem0.ai), then add to your MCP config:
+
+```json
+{
+  "mcpServers": {
+    "mem0": {
+      "command": "uvx",
+      "args": ["mem0-mcp-server"],
+      "env": {
+        "MEM0_API_KEY": "m0-...",
+        "MEM0_DEFAULT_USER_ID": "your-handle"
+      }
+    }
+  }
+}
+```
+
+For Docker, Smithery, and advanced options, see [Mem0 MCP Setup](/platform/mem0-mcp).
+</Tab>
+</Tabs>
+
+## Universal Starter Prompt
+
+Copy this into any AI tool to start building with Mem0:
+
+```text
+I want to start building with Mem0 — a self-improving memory layer for LLM
+applications that gives agents persistent context across sessions.
+
+## Mem0 Resources
+
+**Documentation:**
+- Main docs: https://docs.mem0.ai
+- Platform Quickstart: https://docs.mem0.ai/platform/quickstart
+- OSS Python Quickstart: https://docs.mem0.ai/open-source/python-quickstart
+- OSS Node.js Quickstart: https://docs.mem0.ai/open-source/node-quickstart
+- API Reference: https://docs.mem0.ai/api-reference
+- Full LLM-friendly docs: https://docs.mem0.ai/llms.txt
+
+**Code & Examples:**
+- Core repo: https://github.com/mem0ai/mem0
+- Python SDK: pip install mem0ai
+- TypeScript SDK: npm install mem0ai
+- Cookbooks: https://docs.mem0.ai/cookbooks/overview
+
+**What Mem0 Does:**
+Mem0 is a memory layer for AI apps — managed (Mem0 Platform) or self-hosted
+(Open Source). It stores, retrieves, and manages user memories so agents
+remember preferences, learn from interactions, and personalize over time.
+Sub-50ms retrieval. Dual storage: vector embeddings + graph databases.
+
+**Architecture Overview:**
+- Memory is scoped by user_id, agent_id, or run_id
+- Core operations: add, search, update, delete
+- Memory types: factual (preferences, facts), episodic (past interactions),
+  semantic (concept relationships), working (session state)
+- Integration pattern: retrieve relevant memories → generate response → store
+  new memories
+
+**Quick Usage (Python Platform):**
+  from mem0 import MemoryClient
+  client = MemoryClient(api_key="m0-xxx")
+  client.add("I prefer dark mode and use VS Code.", user_id="user1")
+  results = client.search("What editor do they use?", user_id="user1")
+
+**Quick Usage (JavaScript Platform):**
+  import MemoryClient from 'mem0ai';
+  const client = new MemoryClient({ apiKey: 'm0-xxx' });
+  await client.add([{ role: "user", content: "I prefer dark mode." }], { user_id: "user1" });
+  const results = await client.search("What editor?", { user_id: "user1" });
+
+**Quick Usage (Python Open Source):**
+  from mem0 import Memory
+  m = Memory()
+  m.add("I prefer dark mode and use VS Code.", user_id="user1")
+  results = m.search("What editor do they use?", user_id="user1")
+
+Help me integrate Mem0 into my project. Start by asking what I'm building,
+what language/framework I'm using, and whether I want managed or self-hosted.
+```
+
+## Go Deeper
+
+<CardGroup cols={2}>
+  <Card title="Platform Quickstart" icon="cloud" href="/platform/quickstart">
+    Get started with the managed API
+  </Card>
+  <Card title="Open Source" icon="code-branch" href="/open-source/overview">
+    Self-host with full control
+  </Card>
+  <Card title="Cookbooks" icon="book" href="/cookbooks/overview">
+    Production-ready tutorials and examples
+  </Card>
+  <Card title="API Reference" icon="code" href="/api-reference">
+    Explore every REST endpoint
+  </Card>
+</CardGroup>

examples/nemoclaw/install-mem0-plugin.sh

@@ -0,0 +1,419 @@
+#!/usr/bin/env bash
+# ─────────────────────────────────────────────────────────────────────────────
+# install-mem0-plugin.sh
+#
+# Installs and configures the @mem0/openclaw-mem0 plugin for an existing
+# NemoClaw sandbox. Assumes NemoClaw is already installed and onboarded.
+#
+# Usage:
+#   chmod +x install-mem0-plugin.sh && ./install-mem0-plugin.sh
+#
+# Requirements:
+#   - NemoClaw installed and onboarded (sandbox in Ready state)
+#   - Mem0 API key (from app.mem0.ai)
+# ─────────────────────────────────────────────────────────────────────────────
+
+set -euo pipefail
+
+# ── Colors and formatting ────────────────────────────────────────────────────
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+CYAN='\033[0;36m'
+BOLD='\033[1m'
+DIM='\033[2m'
+NC='\033[0m'
+
+info()    { echo -e "${BLUE}[INFO]${NC}  $*"; }
+success() { echo -e "${GREEN}[OK]${NC}    $*"; }
+warn()    { echo -e "${YELLOW}[WARN]${NC}  $*"; }
+error()   { echo -e "${RED}[ERROR]${NC} $*"; }
+step()    { echo -e "\n${BOLD}${CYAN}── $* ──${NC}\n"; }
+ask()     { echo -en "${BOLD}$*${NC}"; }
+
+die() {
+  error "$*"
+  exit 1
+}
+
+# ── Defaults ─────────────────────────────────────────────────────────────────
+
+MEM0_USER_ID="${MEM0_USER_ID:-default}"
+PLUGIN_PKG="@mem0/openclaw-mem0"
+CONTAINER_NAME="nemoclaw-dev"
+
+# ── Detect platform ──────────────────────────────────────────────────────────
+
+OS_TYPE="$(uname -s)"
+IS_MACOS=false
+IS_LINUX=false
+
+case "$OS_TYPE" in
+  Darwin) IS_MACOS=true ;;
+  Linux)  IS_LINUX=true ;;
+  *)      die "Unsupported OS: $OS_TYPE" ;;
+esac
+
+# ── Helper functions ─────────────────────────────────────────────────────────
+
+check_command() {
+  command -v "$1" &>/dev/null
+}
+
+ensure_nvm() {
+  export NVM_DIR="${NVM_DIR:-$HOME/.nvm}"
+  if [[ -s "$NVM_DIR/nvm.sh" ]]; then
+    source "$NVM_DIR/nvm.sh"
+  fi
+}
+
+ensure_path() {
+  for p in "$HOME/.local/bin" "$HOME/.nvm/versions/node/"*/bin; do
+    if [[ -d "$p" ]] && [[ ":$PATH:" != *":$p:"* ]]; then
+      export PATH="$p:$PATH"
+    fi
+  done
+}
+
+# Wrapper to run a command natively (Linux) or in the container (macOS)
+run_cmd() {
+  if $IS_MACOS; then
+    docker exec "$CONTAINER_NAME" bash -c "export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && $*"
+  else
+    eval "$@"
+  fi
+}
+
+# ── Banner ───────────────────────────────────────────────────────────────────
+
+echo ""
+echo -e "${BOLD}${CYAN}╔══════════════════════════════════════════════════════════════╗${NC}"
+echo -e "${BOLD}${CYAN}║         Mem0 Plugin Installer for NemoClaw                   ║${NC}"
+echo -e "${BOLD}${CYAN}║         Long-term memory for your OpenClaw agent             ║${NC}"
+echo -e "${BOLD}${CYAN}╚══════════════════════════════════════════════════════════════╝${NC}"
+echo ""
+
+if $IS_MACOS; then
+  info "Platform: macOS (using Docker container '$CONTAINER_NAME')"
+else
+  info "Platform: Linux"
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PRE-CHECK: Verify NemoClaw is installed and sandbox is ready
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Pre-check: Verifying NemoClaw setup"
+
+if $IS_LINUX; then
+  ensure_nvm
+  ensure_path
+fi
+
+# Verify nemoclaw/openshell are available
+if $IS_MACOS; then
+  if ! docker ps --format '{{.Names}}' | grep -q "^${CONTAINER_NAME}$"; then
+    die "Docker container '$CONTAINER_NAME' is not running. Start it with: docker start $CONTAINER_NAME"
+  fi
+  if ! run_cmd "command -v nemoclaw" &>/dev/null; then
+    die "NemoClaw is not installed in container '$CONTAINER_NAME'. Run the full setup script first."
+  fi
+else
+  if ! check_command openshell; then
+    die "openshell not found. Is NemoClaw installed and onboarded? Try: source ~/.bashrc"
+  fi
+fi
+
+# Detect sandbox
+SANDBOX_NAME=$(run_cmd "openshell sandbox list 2>/dev/null" | awk 'NR>1 && $1!="" {print $1; exit}' || true)
+
+if [[ -z "$SANDBOX_NAME" ]]; then
+  die "No sandbox found. Run 'nemoclaw onboard' first."
+fi
+
+# Verify sandbox is ready
+SANDBOX_PHASE=$(run_cmd "openshell sandbox get '$SANDBOX_NAME' 2>/dev/null" | grep -i "phase" | awk '{print $NF}' || true)
+if [[ "$SANDBOX_PHASE" != "Ready" ]]; then
+  warn "Sandbox '$SANDBOX_NAME' is not in Ready state (current: ${SANDBOX_PHASE:-unknown})."
+  warn "Waiting up to 2 minutes..."
+  WAIT_OK=false
+  for i in $(seq 1 24); do
+    SANDBOX_PHASE=$(run_cmd "openshell sandbox get '$SANDBOX_NAME' 2>/dev/null" | grep -i "phase" | awk '{print $NF}' || true)
+    if [[ "$SANDBOX_PHASE" == "Ready" ]]; then
+      WAIT_OK=true
+      break
+    fi
+    sleep 5
+  done
+  if ! $WAIT_OK; then
+    die "Sandbox '$SANDBOX_NAME' did not become Ready. Run: openshell sandbox list"
+  fi
+fi
+
+success "NemoClaw installed"
+success "Sandbox '$SANDBOX_NAME' is ready"
+
+# ═════════════════════════════════════════════════════════════════════════════
+# STEP 1: Install Mem0 Plugin
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Step 1: Installing Mem0 plugin ($PLUGIN_PKG)"
+
+# Helper: run a command inside the sandbox non-interactively via piped stdin
+sandbox_exec() {
+  local cmd="$1"
+  if $IS_MACOS; then
+    printf '%s\nexit\n' "$cmd" | docker exec -i "$CONTAINER_NAME" bash -c \
+      "export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && openshell sandbox connect '$SANDBOX_NAME'" 2>&1
+  else
+    printf '%s\nexit\n' "$cmd" | openshell sandbox connect "$SANDBOX_NAME" 2>&1
+  fi
+}
+
+# Check if plugin is already installed
+PLUGIN_EXISTS=false
+if run_cmd "openshell sandbox download '$SANDBOX_NAME' /sandbox/.openclaw/extensions/openclaw-mem0/package.json /tmp/_mem0_plugin_check.json" &>/dev/null; then
+  if [[ -f /tmp/_mem0_plugin_check.json ]] || run_cmd "test -f /tmp/_mem0_plugin_check.json" &>/dev/null; then
+    PLUGIN_EXISTS=true
+  fi
+fi
+rm -f /tmp/_mem0_plugin_check.json 2>/dev/null || true
+run_cmd "rm -f /tmp/_mem0_plugin_check.json" 2>/dev/null || true
+
+if $PLUGIN_EXISTS; then
+  success "Mem0 plugin already installed in sandbox"
+else
+  info "Downloading $PLUGIN_PKG..."
+
+  # Download and build outside the sandbox (on host or in container)
+  run_cmd "cd /tmp && rm -rf openclaw-mem0-full mem0-openclaw-mem0-*.tgz openclaw-mem0-full.tgz"
+
+  if ! run_cmd "cd /tmp && npm pack '$PLUGIN_PKG' 2>/dev/null"; then
+    die "Failed to download $PLUGIN_PKG from npm. Check your internet connection."
+  fi
+
+  success "Downloaded plugin"
+
+  info "Installing plugin dependencies..."
+  run_cmd "mkdir -p /tmp/openclaw-mem0-full && cd /tmp/openclaw-mem0-full && tar xzf /tmp/mem0-openclaw-mem0-*.tgz --strip-components=1 && npm install --omit=dev 2>&1 | tail -3"
+
+  success "Dependencies installed"
+
+  info "Uploading plugin to sandbox..."
+  run_cmd "cd /tmp && tar czf openclaw-mem0-full.tgz -C openclaw-mem0-full ."
+
+  if ! run_cmd "openshell sandbox upload '$SANDBOX_NAME' /tmp/openclaw-mem0-full.tgz /sandbox/openclaw-mem0-full.tgz 2>&1"; then
+    die "Failed to upload plugin to sandbox. Check: openshell sandbox list"
+  fi
+
+  success "Plugin uploaded"
+
+  info "Extracting plugin inside sandbox..."
+  sandbox_exec "mkdir -p ~/.openclaw/extensions/openclaw-mem0 && tar xzf /sandbox/openclaw-mem0-full.tgz/openclaw-mem0-full.tgz -C ~/.openclaw/extensions/openclaw-mem0 2>/dev/null || tar xzf /sandbox/openclaw-mem0-full.tgz -C ~/.openclaw/extensions/openclaw-mem0 2>/dev/null && echo EXTRACT_OK" >/dev/null 2>&1 || true
+
+  # Verify
+  VERIFY_OK=false
+  if run_cmd "openshell sandbox download '$SANDBOX_NAME' /sandbox/.openclaw/extensions/openclaw-mem0/package.json /tmp/_mem0_verify.json" &>/dev/null; then
+    VERIFY_OK=true
+  fi
+  rm -f /tmp/_mem0_verify.json 2>/dev/null || true
+  run_cmd "rm -f /tmp/_mem0_verify.json" 2>/dev/null || true
+  if $VERIFY_OK; then
+    success "Plugin extracted inside sandbox"
+  else
+    warn "Could not verify plugin extraction. You may need to extract manually."
+    if $IS_MACOS; then
+      echo "  docker exec -it $CONTAINER_NAME bash"
+      echo "  export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh"
+    fi
+    echo "  nemoclaw $SANDBOX_NAME connect"
+    echo "  mkdir -p ~/.openclaw/extensions/openclaw-mem0"
+    echo "  tar xzf /sandbox/openclaw-mem0-full.tgz/openclaw-mem0-full.tgz -C ~/.openclaw/extensions/openclaw-mem0"
+  fi
+
+  # Clean up
+  run_cmd "rm -rf /tmp/openclaw-mem0-full /tmp/mem0-openclaw-mem0-*.tgz /tmp/openclaw-mem0-full.tgz" 2>/dev/null || true
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# STEP 2: Update Network Policy
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Step 2: Updating network policy to allow api.mem0.ai and telemetry"
+
+# Find baseline policy
+BASELINE_POLICY=$(run_cmd "find / -path '*/nemoclaw-blueprint/policies/openclaw-sandbox.yaml' 2>/dev/null | head -1" || true)
+
+if [[ -z "$BASELINE_POLICY" ]]; then
+  die "Cannot find NemoClaw baseline policy file. Is NemoClaw installed?"
+fi
+
+info "Baseline policy: $BASELINE_POLICY"
+
+# Check if mem0_api already exists
+HAS_MEM0=$(run_cmd "grep -c mem0_api '$BASELINE_POLICY' 2>/dev/null" || echo "0")
+
+if [[ "$HAS_MEM0" != "0" ]]; then
+  success "mem0_api already in baseline policy"
+else
+  info "Adding api.mem0.ai to network policy..."
+fi
+
+# Create custom policy with mem0_api + telemetry
+run_cmd "node -e \"
+const fs = require('fs');
+let c = fs.readFileSync('$BASELINE_POLICY', 'utf8');
+const mem0Block = '\\n  mem0_api:\\n    name: mem0_api\\n    endpoints:\\n      - host: api.mem0.ai\\n        port: 443\\n        access: full\\n    binaries:\\n      - { path: /usr/local/bin/node }\\n      - { path: /usr/local/bin/openclaw }\\n';
+const telemetryBlock = '\\n  mem0_telemetry:\\n    name: mem0_telemetry\\n    endpoints:\\n      - host: us.i.posthog.com\\n        port: 443\\n        access: full\\n    binaries:\\n      - { path: /usr/local/bin/node }\\n      - { path: /usr/local/bin/openclaw }\\n';
+if (!c.includes('mem0_api')) {
+  if (c.includes('# ── Messaging')) {
+    c = c.replace('  # ── Messaging', mem0Block + '\\n  # ── Messaging');
+  } else {
+    c += mem0Block;
+  }
+}
+if (!c.includes('mem0_telemetry')) {
+  if (c.includes('mem0_api:')) {
+    c = c.replace('  mem0_api:', telemetryBlock + '\\n  mem0_api:');
+  } else {
+    c += telemetryBlock;
+  }
+}
+fs.writeFileSync('/tmp/nemoclaw-mem0-policy.yaml', c);
+console.log('ok');
+\""
+
+success "Custom policy file created"
+
+# Apply the policy
+info "Applying network policy..."
+if ! run_cmd "openshell policy set '$SANDBOX_NAME' --policy /tmp/nemoclaw-mem0-policy.yaml --wait 2>&1"; then
+  error "Failed to apply network policy."
+  echo ""
+  echo "  If you see 'sandbox not found', re-run: nemoclaw onboard"
+  echo "  Then re-run this script."
+  echo ""
+  die "Network policy update failed."
+fi
+
+success "Network policy applied — api.mem0.ai and telemetry allowed"
+
+# ═════════════════════════════════════════════════════════════════════════════
+# STEP 3: Configure Plugin
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Step 3: Configuring Mem0 plugin"
+
+echo ""
+echo -e "  ${DIM}Get your Mem0 API key from: https://app.mem0.ai${NC}"
+echo -e "  ${DIM}The key starts with 'm0-'${NC}"
+echo ""
+ask "Enter your Mem0 API key: "
+read -r MEM0_API_KEY
+
+if [[ -z "$MEM0_API_KEY" ]]; then
+  die "Mem0 API key is required."
+fi
+
+if [[ ! "$MEM0_API_KEY" =~ ^m0- ]]; then
+  warn "Key doesn't start with 'm0-'. Make sure this is correct."
+fi
+
+echo ""
+echo -e "  ${DIM}The user ID scopes all memories. Pick any unique identifier.${NC}"
+echo -e "  ${DIM}Examples: alice, user_123, your-email@example.com${NC}"
+echo ""
+ask "Enter user ID [$MEM0_USER_ID]: "
+read -r custom_user_id
+MEM0_USER_ID="${custom_user_id:-$MEM0_USER_ID}"
+
+info "Configuring plugin inside sandbox..."
+
+CONFIG_SCRIPT="openclaw config set plugins.slots.memory openclaw-mem0 2>&1 | tail -1 && \
+openclaw config set plugins.entries.openclaw-mem0.enabled true 2>&1 | tail -1 && \
+openclaw config set plugins.entries.openclaw-mem0.config.apiKey '$MEM0_API_KEY' 2>&1 | tail -1 && \
+openclaw config set plugins.entries.openclaw-mem0.config.userId '$MEM0_USER_ID' 2>&1 | tail -1 && \
+echo SETUP_DONE"
+
+CONFIG_OUTPUT=$(sandbox_exec "$CONFIG_SCRIPT" || true)
+
+if echo "$CONFIG_OUTPUT" | grep -q "SETUP_DONE"; then
+  success "Plugin configured (mode: platform, user: $MEM0_USER_ID)"
+else
+  warn "Could not verify config. You may need to configure manually:"
+  echo ""
+  if $IS_MACOS; then
+    echo "  docker exec -it $CONTAINER_NAME bash"
+    echo "  export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh"
+  fi
+  echo "  nemoclaw $SANDBOX_NAME connect"
+  echo "  openclaw config set plugins.slots.memory openclaw-mem0"
+  echo "  openclaw config set plugins.entries.openclaw-mem0.enabled true"
+  echo "  openclaw config set plugins.entries.openclaw-mem0.config.apiKey \"$MEM0_API_KEY\""
+  echo "  openclaw config set plugins.entries.openclaw-mem0.config.userId \"$MEM0_USER_ID\""
+  echo ""
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# Done
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Verification"
+
+echo ""
+echo -e "${BOLD}${GREEN}╔══════════════════════════════════════════════════════════════╗${NC}"
+echo -e "${BOLD}${GREEN}║                    Setup Complete!                           ║${NC}"
+echo -e "${BOLD}${GREEN}╚══════════════════════════════════════════════════════════════╝${NC}"
+echo ""
+echo -e "  ${BOLD}Sandbox:${NC}   $SANDBOX_NAME"
+echo -e "  ${BOLD}Plugin:${NC}    @mem0/openclaw-mem0 (platform mode)"
+echo -e "  ${BOLD}User ID:${NC}   $MEM0_USER_ID"
+if $IS_MACOS; then
+  echo -e "  ${BOLD}Container:${NC} $CONTAINER_NAME"
+fi
+echo ""
+echo -e "  ${BOLD}${CYAN}Next steps:${NC}"
+echo ""
+
+if $IS_MACOS; then
+  echo -e "  1. Open a shell in the container:"
+  echo ""
+  echo -e "     ${DIM}docker exec -it $CONTAINER_NAME bash${NC}"
+  echo -e "     ${DIM}export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh${NC}"
+  echo ""
+  echo -e "  2. Connect to the sandbox and start the gateway:"
+  echo ""
+  echo -e "     ${DIM}nemoclaw $SANDBOX_NAME connect${NC}"
+  echo -e "     ${DIM}nemoclaw-start${NC}"
+else
+  echo -e "  1. Connect to the sandbox and start the gateway:"
+  echo ""
+  echo -e "     ${DIM}source ~/.bashrc${NC}"
+  echo -e "     ${DIM}nemoclaw $SANDBOX_NAME connect${NC}"
+  echo -e "     ${DIM}nemoclaw-start${NC}"
+fi
+echo ""
+echo -e "  Then verify the plugin loaded (look for 'openclaw-mem0: registered'):"
+echo ""
+echo -e "     ${DIM}openclaw plugins list${NC}"
+echo ""
+echo -e "  Test auto-capture (storing memories):"
+echo ""
+echo -e "     ${DIM}openclaw agent --agent main --local -m \"My name is Alice\" --session-id test1${NC}"
+echo ""
+echo -e "  Test auto-recall (new session, memories should appear):"
+echo ""
+echo -e "     ${DIM}openclaw agent --agent main --local -m \"What do you know about me?\" --session-id test2${NC}"
+echo ""
+echo -e "  Or use the interactive TUI:"
+echo ""
+echo -e "     ${DIM}openclaw tui${NC}"
+echo ""
+echo -e "  ${YELLOW}Note:${NC} You may see 'Telemetry event capture failed' errors."
+echo -e "  These are harmless and do not affect memory functionality."
+echo ""
+echo -e "  ${BOLD}Documentation:${NC} https://docs.mem0.ai"
+echo -e "  ${BOLD}Plugin source:${NC} https://www.npmjs.com/package/@mem0/openclaw-mem0"
+echo ""

examples/nemoclaw/quickstart.md

@@ -0,0 +1,272 @@
+# Mem0 Plugin for NemoClaw — Quickstart
+
+Add persistent long-term memory to your [NemoClaw](https://docs.nvidia.com/nemoclaw/latest/get-started/quickstart.html) OpenClaw agent using the `@mem0/openclaw-mem0` plugin.
+
+> **Note:** This plugin requires **Mem0 Platform mode** (i.e., a Mem0 API key from [app.mem0.ai](https://app.mem0.ai)). Open-source mode is not supported in NemoClaw sandboxes because the sandbox proxy blocks `/v1/embeddings` requests required by the open-source backend. See [Known Limitations](#known-limitations) for details.
+
+## Prerequisites
+
+| Resource | Recommended | Minimum |
+|----------|-------------|---------|
+| CPU      | 4+ vCPU     | 2 vCPU  |
+| RAM      | 16 GB       | 8 GB    |
+| Disk     | 40 GB free  | 20 GB free |
+
+**Accounts required:**
+
+- **NVIDIA** — sign up at [build.nvidia.com](https://build.nvidia.com), generate an API key at [build.nvidia.com/settings/api-keys](https://build.nvidia.com/settings/api-keys) (starts with `nvapi-`)
+- **Mem0** — sign up at [app.mem0.ai](https://app.mem0.ai), generate an API key from the dashboard (starts with `m0-`)
+
+**Supported platforms:** Ubuntu 22.04+, macOS (via Docker), Windows (WSL 2 + Docker)
+
+---
+
+## Choose Your Path
+
+### Option A: Full Setup (NemoClaw + Mem0 Plugin)
+
+Use this if you **don't have NemoClaw installed yet**. The script handles everything: Docker, Node.js, NemoClaw installation, onboarding, Mem0 plugin installation, network policy, and configuration.
+
+```bash
+# Download
+curl -fsSL https://raw.githubusercontent.com/mem0ai/mem0/main/examples/nemoclaw/setup-mem0-nemoclaw.sh -o setup-mem0-nemoclaw.sh
+
+# Run
+chmod +x setup-mem0-nemoclaw.sh
+./setup-mem0-nemoclaw.sh
+```
+
+The script runs through 7 phases:
+
+| Phase | What it does | User input |
+|-------|-------------|------------|
+| 1 | Prerequisites (Docker, RAM, disk) | None (automatic) |
+| 2 | Install NemoClaw | None (automatic) |
+| 3 | NemoClaw onboarding (sandbox + k3s) | Sandbox name, NVIDIA API key |
+| 4 | Install Mem0 plugin into sandbox | None (automatic) |
+| 5 | Update network policy for `api.mem0.ai` | None (automatic) |
+| 6 | Configure plugin | Mem0 API key, user ID |
+| 7 | Verification | None |
+
+### Option B: Plugin Only (NemoClaw Already Installed)
+
+Use this if you **already have NemoClaw installed and onboarded** with a sandbox in `Ready` state.
+
+```bash
+# Download
+curl -fsSL https://raw.githubusercontent.com/mem0ai/mem0/main/examples/nemoclaw/install-mem0-plugin.sh -o install-mem0-plugin.sh
+
+# Run
+chmod +x install-mem0-plugin.sh
+./install-mem0-plugin.sh
+```
+
+The script auto-detects your sandbox and runs 3 steps:
+
+| Step | What it does | User input |
+|------|-------------|------------|
+| 1 | Install Mem0 plugin into sandbox | None (automatic) |
+| 2 | Update network policy for `api.mem0.ai` | None (automatic) |
+| 3 | Configure plugin | Mem0 API key, user ID |
+
+---
+
+## Verify the Installation
+
+After either script completes, connect to the sandbox and start the gateway:
+
+```bash
+source ~/.bashrc
+nemoclaw <sandbox-name> connect
+nemoclaw-start
+```
+
+Look for this line in the startup output:
+
+```
+openclaw-mem0: registered (mode: platform, user: your-user-id, graph: false, autoRecall: true, autoCapture: true)
+```
+
+You can also verify with:
+
+```bash
+openclaw plugins list
+```
+
+The `Memory (Mem0)` plugin should show status **loaded**.
+
+## Test the Integration
+
+All test commands run **inside the sandbox**.
+
+### Test 1: Auto-capture (storing memories)
+
+```bash
+openclaw agent --agent main --local \
+  -m "My name is Alice and I work on distributed systems" \
+  --session-id test1
+```
+
+Look for: `openclaw-mem0: auto-captured 1 memories`
+
+### Test 2: Auto-recall (retrieving memories across sessions)
+
+Start a **new session** (different `--session-id`):
+
+```bash
+openclaw agent --agent main --local \
+  -m "What do you know about me?" \
+  --session-id test2
+```
+
+Look for: `openclaw-mem0: injecting 1 memories into context (1 long-term, 0 session)`
+
+The agent should respond with information from the previous session ("Alice", "distributed systems").
+
+### Test 3: Interactive TUI
+
+```bash
+openclaw tui
+```
+
+Send messages and the plugin will automatically capture and recall memories in the background.
+
+---
+
+## Plugin Configuration Reference
+
+All options are set inside the sandbox via:
+
+```bash
+openclaw config set plugins.entries.openclaw-mem0.config.<key> <value>
+```
+
+| Key | Type | Default | Description |
+|-----|------|---------|-------------|
+| `mode` | `"platform"` \| `"open-source"` | `"platform"` | Backend mode |
+| `apiKey` | string | — | Mem0 API key (starts with `m0-`) |
+| `userId` | string | `"default"` | Unique identifier for the user |
+| `autoRecall` | boolean | `true` | Inject memories before each agent turn |
+| `autoCapture` | boolean | `true` | Store facts after each agent turn |
+| `topK` | number | `5` | Max memories per recall |
+| `searchThreshold` | number | `0.3` | Min similarity score (0-1) |
+| `orgId` | string | — | Mem0 organization ID |
+| `projectId` | string | — | Mem0 project ID |
+| `enableGraph` | boolean | `false` | Enable entity graph for relationships |
+| `customInstructions` | string | — | Rules for what Mem0 should store/exclude |
+
+## Agent Memory Tools
+
+Once the plugin is active, the agent can use these tools during conversations:
+
+| Tool | Description |
+|------|-------------|
+| `memory_search` | Search memories by natural language |
+| `memory_list` | List all stored memories for a user |
+| `memory_store` | Explicitly save a fact |
+| `memory_get` | Retrieve a memory by ID |
+| `memory_forget` | Delete by ID or by query |
+
+## CLI Commands
+
+Run these inside the sandbox:
+
+```bash
+# Search all memories (long-term + session)
+openclaw mem0 search "what languages does the user know"
+
+# Search only long-term memories
+openclaw mem0 search "user preferences" --scope long-term
+
+# Search only session memories
+openclaw mem0 search "current task" --scope session
+
+# Memory stats
+openclaw mem0 stats
+```
+
+---
+
+## Troubleshooting
+
+For detailed troubleshooting steps, see the [troubleshooting guide](troubleshooting-guide.pdf) included in this directory.
+
+### Common Issues
+
+#### `npm tar TAR_ENTRY_ERROR ENOENT` during NemoClaw installation
+
+This is a known npm bug where concurrent tar extraction races cause `ENOENT` errors on deeply nested packages. The `setup-mem0-nemoclaw.sh` script works around this by cloning the NemoClaw repo and installing with `--maxsockets=1` to serialize downloads. If you installed NemoClaw manually and hit this error:
+
+```bash
+git clone --depth 1 https://github.com/NVIDIA/NemoClaw.git ~/.nemoclaw-src
+cd ~/.nemoclaw-src
+npm install --maxsockets=1
+npm link
+```
+
+#### `sandbox not found` during onboarding step 7
+
+The gateway restarted during onboarding and lost the sandbox state. Re-run `nemoclaw onboard`. When prompted that the sandbox already exists, choose `y` to recreate it.
+
+#### `npm error 403 Forbidden` when installing plugin inside sandbox
+
+The OpenShell gateway's TLS proxy blocks scoped npm packages (the `%2f` in the URL). Use the manual installation method (Method 2 in the scripts) which downloads outside the sandbox and uploads the tarball.
+
+#### `capture failed: Connection error` or `recall failed: Connection error`
+
+The network policy is not applied or missing the `mem0_api` entry. Re-run the plugin install script or manually apply the policy:
+
+```bash
+openshell policy set <sandbox-name> --policy /tmp/nemoclaw-mem0-policy.yaml --wait
+```
+
+#### `Telemetry event capture failed: TypeError: fetch failed`
+
+This is harmless. The Mem0 SDK's telemetry endpoint (`us.i.posthog.com`) is blocked by the sandbox proxy. It does not affect memory functionality.
+
+#### Plugin shows `disabled` in `openclaw plugins list`
+
+The memory slot is not set to `openclaw-mem0`. Inside the sandbox, run:
+
+```bash
+openclaw config set plugins.slots.memory openclaw-mem0
+```
+
+#### `K8s namespace not ready` on Ubuntu 24.04
+
+Ubuntu 24.04 defaults to cgroup v2 which causes k3s (used by NemoClaw) to fail. Apply the cgroup fix:
+
+```bash
+sudo python3 -c "
+import json, os
+p = '/etc/docker/daemon.json'
+c = json.load(open(p)) if os.path.exists(p) else {}
+c['default-cgroupns-mode'] = 'host'
+json.dump(c, open(p, 'w'), indent=2)
+"
+sudo systemctl restart docker
+```
+
+Then re-run `nemoclaw onboard`.
+
+### Known Limitations
+
+**Open-source mode is not supported in NemoClaw sandboxes.** The Mem0 open-source mode requires calling `/v1/embeddings` on an external LLM provider. NemoClaw's sandbox proxy intercepts all OpenAI-compatible API requests but only allows `/v1/chat/completions` through. Use **platform mode** instead.
+
+---
+
+## Files in This Directory
+
+| File | Description |
+|------|-------------|
+| `setup-mem0-nemoclaw.sh` | Full setup script (NemoClaw + Mem0 plugin) |
+| `install-mem0-plugin.sh` | Plugin-only install script (NemoClaw already set up) |
+| `troubleshooting-guide.pdf` | Detailed setup and troubleshooting guide |
+| `quickstart.md` | This file |
+
+## Links
+
+- [Mem0 Documentation](https://docs.mem0.ai)
+- [NemoClaw Documentation](https://docs.nvidia.com/nemoclaw/latest/get-started/quickstart.html)
+- [`@mem0/openclaw-mem0` on npm](https://www.npmjs.com/package/@mem0/openclaw-mem0)
+- [Mem0 Dashboard](https://app.mem0.ai)

examples/nemoclaw/setup-mem0-nemoclaw.sh

@@ -0,0 +1,864 @@
+#!/usr/bin/env bash
+# ─────────────────────────────────────────────────────────────────────────────
+# setup-mem0-nemoclaw.sh
+#
+# One-script setup for NemoClaw + Mem0 OpenClaw plugin.
+# Supports Ubuntu servers (native) and macOS (via Docker container).
+#
+# Usage:
+#   curl -fsSL https://raw.githubusercontent.com/mem0ai/mem0/main/scripts/setup-mem0-nemoclaw.sh | bash
+#   # or
+#   chmod +x setup-mem0-nemoclaw.sh && ./setup-mem0-nemoclaw.sh
+#
+# Requirements:
+#   - Ubuntu 22.04+ OR macOS with Docker Desktop OR Windows with WSL 2 + Docker Desktop
+#   - 8 GB RAM minimum (16 GB recommended)
+#   - 40 GB free disk (20 GB minimum)
+#   - NVIDIA API key (from build.nvidia.com)
+#   - Mem0 API key (from app.mem0.ai)
+# ─────────────────────────────────────────────────────────────────────────────
+
+set -euo pipefail
+
+# ── Colors and formatting ────────────────────────────────────────────────────
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+CYAN='\033[0;36m'
+BOLD='\033[1m'
+DIM='\033[2m'
+NC='\033[0m'
+
+info()    { echo -e "${BLUE}[INFO]${NC}  $*"; }
+success() { echo -e "${GREEN}[OK]${NC}    $*"; }
+warn()    { echo -e "${YELLOW}[WARN]${NC}  $*"; }
+error()   { echo -e "${RED}[ERROR]${NC} $*"; }
+step()    { echo -e "\n${BOLD}${CYAN}── $* ──${NC}\n"; }
+ask()     { echo -en "${BOLD}$*${NC}"; }
+
+die() {
+  error "$*"
+  exit 1
+}
+
+# ── Defaults ─────────────────────────────────────────────────────────────────
+
+SANDBOX_NAME="${SANDBOX_NAME:-my-assistant}"
+MEM0_USER_ID="${MEM0_USER_ID:-default}"
+MIN_RAM_MB=6000
+MIN_DISK_MB=15000
+PLUGIN_PKG="@mem0/openclaw-mem0"
+CONTAINER_NAME="nemoclaw-dev"
+
+# ── Detect platform ─────────────────────────────────────────────────────────
+
+OS_TYPE="$(uname -s)"
+IS_MACOS=false
+IS_LINUX=false
+IS_WSL=false
+
+case "$OS_TYPE" in
+  Darwin) IS_MACOS=true ;;
+  Linux)
+    IS_LINUX=true
+    # Detect WSL (Windows Subsystem for Linux)
+    if grep -qi "microsoft\|wsl" /proc/version 2>/dev/null; then
+      IS_WSL=true
+    fi
+    ;;
+  MINGW*|MSYS*|CYGWIN*)
+    error "This script cannot run in Git Bash, MSYS2, or Cygwin."
+    echo ""
+    echo "  NemoClaw requires a full Linux environment. On Windows, use WSL 2:"
+    echo ""
+    echo "  1. Open PowerShell as Administrator and run:"
+    echo "     wsl --install -d Ubuntu-24.04"
+    echo ""
+    echo "  2. Restart your computer when prompted"
+    echo ""
+    echo "  3. Open 'Ubuntu' from the Start menu (this opens a WSL 2 shell)"
+    echo ""
+    echo "  4. Install Docker Desktop for Windows:"
+    echo "     https://www.docker.com/products/docker-desktop/"
+    echo "     Enable 'Use the WSL 2 based engine' in Docker Desktop settings"
+    echo "     Enable 'Ubuntu-24.04' under Resources → WSL Integration"
+    echo ""
+    echo "  5. In the Ubuntu WSL 2 shell, re-run this script:"
+    echo "     curl -fsSL https://raw.githubusercontent.com/mem0ai/mem0/main/scripts/setup-mem0-nemoclaw.sh | bash"
+    echo ""
+    die "Please use WSL 2 instead."
+    ;;
+  *)
+    die "Unsupported OS: $OS_TYPE. This script supports Linux (Ubuntu), macOS, and Windows (WSL 2)."
+    ;;
+esac
+
+# ── Helper functions ─────────────────────────────────────────────────────────
+
+check_command() {
+  command -v "$1" &>/dev/null
+}
+
+get_ram_mb() {
+  if $IS_MACOS; then
+    sysctl -n hw.memsize 2>/dev/null | awk '{printf "%.0f", $1/1024/1024}' || echo "0"
+  else
+    free -m 2>/dev/null | awk '/^Mem:/ {print $2}' || echo "0"
+  fi
+}
+
+get_disk_mb() {
+  if $IS_MACOS; then
+    df -m / 2>/dev/null | awk 'NR==2 {print $4}' || echo "0"
+  else
+    df -m / 2>/dev/null | awk 'NR==2 {print $4}' || echo "0"
+  fi
+}
+
+ensure_nvm() {
+  export NVM_DIR="${NVM_DIR:-$HOME/.nvm}"
+  if [[ -s "$NVM_DIR/nvm.sh" ]]; then
+    source "$NVM_DIR/nvm.sh"
+  fi
+}
+
+ensure_path() {
+  for p in "$HOME/.local/bin" "$HOME/.nvm/versions/node/"*/bin; do
+    if [[ -d "$p" ]] && [[ ":$PATH:" != *":$p:"* ]]; then
+      export PATH="$p:$PATH"
+    fi
+  done
+}
+
+wait_for_sandbox() {
+  local name="$1"
+  local timeout=120
+  local elapsed=0
+  while (( elapsed < timeout )); do
+    local phase
+    phase=$($RUN_CMD openshell sandbox get "$name" 2>/dev/null | grep -i "phase" | awk '{print $NF}' || true)
+    if [[ "$phase" == "Ready" ]]; then
+      return 0
+    fi
+    sleep 5
+    elapsed=$((elapsed + 5))
+  done
+  return 1
+}
+
+# ── macOS: run command inside the container ──────────────────────────────────
+# On macOS, NemoClaw runs inside a Docker container. All nemoclaw/openshell
+# commands must be exec'd into the container. On Linux, they run natively.
+
+setup_run_cmd() {
+  if $IS_MACOS; then
+    RUN_CMD="docker exec $CONTAINER_NAME bash -c 'export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && "
+    RUN_CMD_SUFFIX="'"
+    RUN_CMD_IT="docker exec -it $CONTAINER_NAME bash -c 'export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && "
+    RUN_CMD_IT_SUFFIX="'"
+  else
+    RUN_CMD=""
+    RUN_CMD_SUFFIX=""
+    RUN_CMD_IT=""
+    RUN_CMD_IT_SUFFIX=""
+  fi
+}
+
+# Wrapper to run a command natively (Linux) or in the container (macOS)
+run_cmd() {
+  if $IS_MACOS; then
+    docker exec "$CONTAINER_NAME" bash -c "export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && $*"
+  else
+    eval "$@"
+  fi
+}
+
+run_cmd_it() {
+  if $IS_MACOS; then
+    docker exec -it "$CONTAINER_NAME" bash -c "export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && $*"
+  else
+    eval "$@"
+  fi
+}
+
+# ── Banner ───────────────────────────────────────────────────────────────────
+
+echo ""
+echo -e "${BOLD}${CYAN}╔══════════════════════════════════════════════════════════════╗${NC}"
+echo -e "${BOLD}${CYAN}║         Mem0 Plugin Setup for NemoClaw                       ║${NC}"
+echo -e "${BOLD}${CYAN}║         Long-term memory for your OpenClaw agent             ║${NC}"
+echo -e "${BOLD}${CYAN}╚══════════════════════════════════════════════════════════════╝${NC}"
+echo ""
+
+if $IS_MACOS; then
+  info "Platform: macOS (will use Docker container)"
+elif $IS_WSL; then
+  info "Platform: Windows (WSL 2)"
+else
+  info "Platform: Linux"
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 1: Prerequisites
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Phase 1: Checking prerequisites"
+
+# ── OS check ─────────────────────────────────────────────────────────────────
+
+if $IS_LINUX; then
+  if [[ -f /etc/os-release ]]; then
+    source /etc/os-release
+    if [[ "$ID" != "ubuntu" ]]; then
+      warn "Detected OS: $PRETTY_NAME (not Ubuntu). This script is tested on Ubuntu 22.04+."
+      ask "Continue anyway? [y/N]: "
+      read -r cont
+      [[ "$cont" =~ ^[Yy]$ ]] || exit 0
+    else
+      success "OS: $PRETTY_NAME"
+    fi
+  else
+    warn "Cannot detect Linux distribution. Proceeding anyway."
+  fi
+elif $IS_MACOS; then
+  MACOS_VERSION=$(sw_vers -productVersion 2>/dev/null || echo "unknown")
+  success "OS: macOS $MACOS_VERSION"
+fi
+
+# ── RAM check ────────────────────────────────────────────────────────────────
+
+RAM_MB=$(get_ram_mb)
+if (( RAM_MB < MIN_RAM_MB )); then
+  error "Insufficient RAM: ${RAM_MB} MB available, ${MIN_RAM_MB} MB required."
+  echo ""
+  echo "  NemoClaw needs at least 8 GB RAM (16 GB recommended)."
+  if $IS_WSL; then
+    echo ""
+    echo "  WSL 2 may have limited memory. Increase it:"
+    echo "    1. Create/edit %USERPROFILE%\\.wslconfig in Windows"
+    echo "    2. Add:"
+    echo "       [wsl2]"
+    echo "       memory=8GB"
+    echo "    3. Restart WSL:  wsl --shutdown  (from PowerShell)"
+  elif $IS_LINUX; then
+    echo ""
+    echo "  If running on AWS EC2:"
+    echo "    1. Stop the instance"
+    echo "    2. Change instance type to t3.large (8 GB) or t3.xlarge (16 GB)"
+    echo "    3. Start the instance and re-run this script"
+  fi
+  echo ""
+  die "Aborting due to insufficient RAM."
+else
+  success "RAM: ${RAM_MB} MB available"
+fi
+
+# ── Disk check ───────────────────────────────────────────────────────────────
+
+DISK_MB=$(get_disk_mb)
+if (( DISK_MB < MIN_DISK_MB )); then
+  error "Insufficient disk space: ${DISK_MB} MB available, ${MIN_DISK_MB} MB required."
+  echo ""
+  echo "  NemoClaw needs at least 20 GB free disk (40 GB recommended)."
+  echo ""
+  echo "  Quick fix:"
+  echo "    docker system prune -a -f    # Remove unused Docker data"
+  if $IS_LINUX; then
+    echo ""
+    echo "  If running on AWS EC2, expand the EBS volume:"
+    echo "    1. Go to AWS Console → EC2 → Volumes"
+    echo "    2. Select the volume, Actions → Modify Volume → increase size"
+    echo "    3. Then run:"
+    echo "       sudo growpart /dev/xvda 1"
+    echo "       sudo resize2fs /dev/xvda1"
+  fi
+  echo ""
+  die "Aborting due to insufficient disk space."
+else
+  success "Disk: ${DISK_MB} MB available"
+fi
+
+# ── Docker check ─────────────────────────────────────────────────────────────
+
+if ! check_command docker; then
+  if $IS_MACOS; then
+    error "Docker not found."
+    echo ""
+    echo "  Install Docker Desktop for macOS:"
+    echo "    https://www.docker.com/products/docker-desktop/"
+    echo ""
+    echo "  After installing, start Docker Desktop and re-run this script."
+    echo ""
+    die "Docker Desktop is required on macOS."
+  elif $IS_WSL; then
+    error "Docker not found inside WSL."
+    echo ""
+    echo "  Docker Desktop must be installed on Windows with WSL 2 integration enabled:"
+    echo ""
+    echo "  1. Install Docker Desktop for Windows:"
+    echo "     https://www.docker.com/products/docker-desktop/"
+    echo ""
+    echo "  2. Open Docker Desktop → Settings → General:"
+    echo "     ✓ Enable 'Use the WSL 2 based engine'"
+    echo ""
+    echo "  3. Open Docker Desktop → Settings → Resources → WSL Integration:"
+    echo "     ✓ Enable integration with your Ubuntu distribution"
+    echo ""
+    echo "  4. Click 'Apply & restart', then re-run this script in WSL."
+    echo ""
+    die "Docker Desktop WSL 2 integration is required."
+  else
+    info "Docker not found. Installing Docker..."
+    sudo apt-get update -qq
+    sudo apt-get install -y -qq docker.io >/dev/null 2>&1
+    sudo systemctl enable --now docker
+    sudo usermod -aG docker "$USER"
+    success "Docker installed"
+  fi
+fi
+
+if $IS_LINUX; then
+  # Ensure Docker daemon is running
+  if ! sudo docker info &>/dev/null; then
+    info "Starting Docker daemon..."
+    sudo systemctl start docker
+    sleep 2
+  fi
+
+  # Handle Docker group permissions
+  if ! docker info &>/dev/null; then
+    if id -nG "$USER" | grep -qw docker || grep -q "^docker:.*\b${USER}\b" /etc/group; then
+      info "Activating docker group for current session..."
+      exec sg docker -c "$0 $*"
+    else
+      info "Adding $USER to docker group..."
+      sudo usermod -aG docker "$USER"
+      info "Activating docker group for current session..."
+      exec sg docker -c "$0 $*"
+    fi
+  fi
+else
+  # macOS: just verify Docker is responding
+  if ! docker info &>/dev/null; then
+    error "Docker is not running."
+    echo ""
+    echo "  Start Docker Desktop and wait for it to be ready, then re-run this script."
+    echo ""
+    die "Docker Desktop is not running."
+  fi
+fi
+
+success "Docker: running ($(docker --version | awk '{print $3}' | tr -d ','))"
+
+# ── Linux-only: cgroup v2 fix for Ubuntu 24.04 ──────────────────────────────
+
+if $IS_LINUX && [[ "${VERSION_ID:-}" == "24.04" ]]; then
+  DAEMON_JSON="/etc/docker/daemon.json"
+  NEEDS_CGROUP_FIX=false
+
+  if [[ ! -f "$DAEMON_JSON" ]]; then
+    NEEDS_CGROUP_FIX=true
+  elif ! grep -q '"default-cgroupns-mode"' "$DAEMON_JSON" 2>/dev/null; then
+    NEEDS_CGROUP_FIX=true
+  fi
+
+  if $NEEDS_CGROUP_FIX; then
+    warn "Ubuntu 24.04 detected — applying cgroup v2 fix for Docker."
+    warn "This prevents 'K8s namespace not ready' errors during onboarding."
+    sudo python3 -c "
+import json, os
+p = '$DAEMON_JSON'
+c = json.load(open(p)) if os.path.exists(p) else {}
+c['default-cgroupns-mode'] = 'host'
+json.dump(c, open(p, 'w'), indent=2)
+"
+    sudo systemctl restart docker
+    success "Docker cgroup v2 fix applied"
+  else
+    success "Docker cgroup v2: already configured"
+  fi
+fi
+
+# ── Linux-only: Swap check ──────────────────────────────────────────────────
+
+if $IS_LINUX; then
+  SWAP_MB=$(free -m 2>/dev/null | awk '/^Swap:/ {print $2}' || echo "0")
+  if (( RAM_MB < 12000 && SWAP_MB < 2000 )); then
+    warn "Low RAM (${RAM_MB} MB) and low swap. Adding 4 GB swap to prevent OOM kills."
+    if [[ ! -f /swapfile ]]; then
+      sudo fallocate -l 4G /swapfile
+      sudo chmod 600 /swapfile
+      sudo mkswap /swapfile >/dev/null
+      sudo swapon /swapfile
+      success "4 GB swap enabled"
+    else
+      success "Swap file already exists"
+    fi
+  fi
+fi
+
+# ── macOS-only: Create Docker container ──────────────────────────────────────
+
+if $IS_MACOS; then
+  step "Phase 1b: Setting up NemoClaw Docker container"
+
+  if docker ps --format '{{.Names}}' | grep -q "^${CONTAINER_NAME}$"; then
+    success "Container '$CONTAINER_NAME' is already running"
+  elif docker ps -a --format '{{.Names}}' | grep -q "^${CONTAINER_NAME}$"; then
+    info "Starting existing container '$CONTAINER_NAME'..."
+    docker start "$CONTAINER_NAME"
+    success "Container started"
+  else
+    info "Creating Ubuntu container '$CONTAINER_NAME' for NemoClaw..."
+    echo -e "  ${DIM}This container runs NemoClaw with --network host and Docker socket access.${NC}"
+
+    docker run -d \
+      --name "$CONTAINER_NAME" \
+      --privileged \
+      --network host \
+      -v /var/run/docker.sock:/var/run/docker.sock \
+      ubuntu:24.04 sleep infinity
+
+    success "Container '$CONTAINER_NAME' created"
+
+    info "Installing dependencies inside container..."
+    docker exec "$CONTAINER_NAME" bash -c "apt-get update -qq && apt-get install -y -qq curl git docker.io >/dev/null 2>&1"
+    success "Dependencies installed"
+  fi
+
+  # Check if NemoClaw is installed in the container
+  NEMO_IN_CONTAINER=$(docker exec "$CONTAINER_NAME" bash -c "export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh 2>/dev/null && command -v nemoclaw" 2>/dev/null || true)
+  if [[ -z "$NEMO_IN_CONTAINER" ]]; then
+    info "Installing NemoClaw inside container (this takes a few minutes)..."
+    # The NVIDIA installer triggers npm tar race conditions. Workaround: clone
+    # and install with --maxsockets=1 to serialize downloads.
+    docker exec -it "$CONTAINER_NAME" bash -c "
+      export NVM_DIR=/root/.nvm &&
+      if [ ! -s \"\$NVM_DIR/nvm.sh\" ]; then
+        curl -fsSL https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash &&
+        source \"\$NVM_DIR/nvm.sh\" &&
+        nvm install 22
+      else
+        source \"\$NVM_DIR/nvm.sh\"
+      fi &&
+      git clone --depth 1 https://github.com/NVIDIA/NemoClaw.git /root/.nemoclaw-src &&
+      cd /root/.nemoclaw-src &&
+      npm install --maxsockets=1 &&
+      npm link
+    "
+    success "NemoClaw installed in container"
+  else
+    success "NemoClaw already installed in container"
+  fi
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 2: Install NemoClaw (Linux-only, macOS handled above)
+# ═════════════════════════════════════════════════════════════════════════════
+
+if $IS_LINUX; then
+  step "Phase 2: Installing NemoClaw"
+
+  ensure_nvm
+  ensure_path
+
+  if check_command nemoclaw; then
+    success "NemoClaw already installed: $(nemoclaw --version 2>/dev/null || echo 'found')"
+  else
+    info "Installing NemoClaw (this takes a few minutes)..."
+
+    # Install Node.js via nvm if not available
+    if ! check_command node; then
+      info "Node.js not found — installing via nvm..."
+      export NVM_DIR="${NVM_DIR:-$HOME/.nvm}"
+      curl -fsSL https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash
+      source "$NVM_DIR/nvm.sh"
+      nvm install 22
+      ensure_nvm
+      ensure_path
+    fi
+
+    # The NVIDIA installer (curl | bash) triggers npm tar race conditions
+    # causing ENOENT errors on deeply nested packages. Workaround: clone the
+    # repo and install with --maxsockets=1 to serialize downloads.
+    NEMOCLAW_DIR="$HOME/.nemoclaw-src"
+    rm -rf "$NEMOCLAW_DIR"
+    info "Cloning NemoClaw from GitHub..."
+    git clone --depth 1 https://github.com/NVIDIA/NemoClaw.git "$NEMOCLAW_DIR"
+    cd "$NEMOCLAW_DIR"
+    info "Installing dependencies (serialized to avoid tar race)..."
+    npm install --maxsockets=1
+    npm link
+    cd - >/dev/null
+
+    ensure_nvm
+    ensure_path
+    source "$HOME/.bashrc" 2>/dev/null || true
+    hash -r 2>/dev/null || true
+
+    if ! check_command nemoclaw; then
+      # npm link may place the binary outside the current PATH; find and add it
+      NEMOCLAW_BIN=$(find "$HOME/.nvm" -name nemoclaw \( -type f -o -type l \) -path "*/bin/*" 2>/dev/null | head -1)
+      if [[ -n "$NEMOCLAW_BIN" ]]; then
+        export PATH="$(dirname "$NEMOCLAW_BIN"):$PATH"
+      fi
+    fi
+
+    if ! check_command nemoclaw; then
+      die "NemoClaw installation failed. Try: source ~/.bashrc && nemoclaw --help"
+    fi
+
+    success "NemoClaw installed"
+  fi
+
+  ensure_path
+  if ! check_command openshell; then
+    warn "openshell not on PATH yet — it will be installed during onboarding."
+  fi
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 3: NemoClaw Onboarding
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Phase 3: NemoClaw Onboarding"
+
+# Check if a sandbox already exists
+EXISTING_SANDBOX=""
+if $IS_MACOS; then
+  EXISTING_SANDBOX=$(run_cmd "openshell sandbox list 2>/dev/null" | awk 'NR>1 && $1!="" {print $1; exit}' || true)
+else
+  ensure_path
+  if check_command openshell; then
+    EXISTING_SANDBOX=$(openshell sandbox list 2>/dev/null | awk 'NR>1 && $1!="" {print $1; exit}' || true)
+  fi
+fi
+
+if [[ -n "$EXISTING_SANDBOX" ]]; then
+  success "Sandbox already exists: $EXISTING_SANDBOX"
+  SANDBOX_NAME="$EXISTING_SANDBOX"
+  ask "Use existing sandbox '$SANDBOX_NAME'? [Y/n]: "
+  read -r use_existing
+  if [[ "$use_existing" =~ ^[Nn]$ ]]; then
+    ask "Enter sandbox name [my-assistant]: "
+    read -r custom_name
+    SANDBOX_NAME="${custom_name:-my-assistant}"
+    info "Running NemoClaw onboarding..."
+    echo -e "  ${DIM}You'll need your NVIDIA API key (nvapi-...) from build.nvidia.com${NC}"
+    echo ""
+    run_cmd_it "nemoclaw onboard"
+    if $IS_LINUX; then ensure_path; fi
+  fi
+else
+  info "No existing sandbox found. Running NemoClaw onboarding..."
+  echo ""
+  echo -e "  ${DIM}The onboarding wizard will guide you through 7 steps:${NC}"
+  echo -e "  ${DIM}  1. Preflight checks (automatic)${NC}"
+  echo -e "  ${DIM}  2. Start gateway (automatic, takes 1-2 min)${NC}"
+  echo -e "  ${DIM}  3. Sandbox name — enter a name (e.g. my-assistant)${NC}"
+  echo -e "  ${DIM}  4. NVIDIA API key — paste your nvapi-... key${NC}"
+  echo -e "  ${DIM}  5. Inference provider (automatic)${NC}"
+  echo -e "  ${DIM}  6. OpenClaw setup (automatic)${NC}"
+  echo -e "  ${DIM}  7. Policy presets — type Y to apply pypi and npm${NC}"
+  echo ""
+  ask "Press Enter to start onboarding..."
+  read -r
+
+  run_cmd_it "nemoclaw onboard"
+
+  if $IS_LINUX; then
+    ensure_nvm
+    ensure_path
+  fi
+
+  # Detect sandbox name
+  SANDBOX_NAME=$(run_cmd "openshell sandbox list 2>/dev/null" | awk 'NR>1 && $1!="" {print $1; exit}' || echo "$SANDBOX_NAME")
+fi
+
+# Verify sandbox exists
+SANDBOX_PHASE=$(run_cmd "openshell sandbox get '$SANDBOX_NAME' 2>/dev/null" | grep -i "phase" | awk '{print $NF}' || true)
+if [[ "$SANDBOX_PHASE" != "Ready" ]]; then
+  warn "Sandbox '$SANDBOX_NAME' is not in Ready state (current: ${SANDBOX_PHASE:-unknown})."
+  warn "Waiting up to 2 minutes..."
+  WAIT_OK=false
+  for i in $(seq 1 24); do
+    SANDBOX_PHASE=$(run_cmd "openshell sandbox get '$SANDBOX_NAME' 2>/dev/null" | grep -i "phase" | awk '{print $NF}' || true)
+    if [[ "$SANDBOX_PHASE" == "Ready" ]]; then
+      WAIT_OK=true
+      break
+    fi
+    sleep 5
+  done
+  if ! $WAIT_OK; then
+    die "Sandbox '$SANDBOX_NAME' did not become Ready. Run: openshell sandbox list"
+  fi
+fi
+
+success "Sandbox '$SANDBOX_NAME' is ready"
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 4: Install Mem0 Plugin
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Phase 4: Installing Mem0 plugin ($PLUGIN_PKG)"
+
+# Helper: run a command inside the sandbox non-interactively via piped stdin
+sandbox_exec() {
+  local cmd="$1"
+  if $IS_MACOS; then
+    printf '%s\nexit\n' "$cmd" | docker exec -i "$CONTAINER_NAME" bash -c \
+      "export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh && openshell sandbox connect '$SANDBOX_NAME'" 2>&1
+  else
+    printf '%s\nexit\n' "$cmd" | openshell sandbox connect "$SANDBOX_NAME" 2>&1
+  fi
+}
+
+# Check if plugin is already installed by trying to download a known file from the sandbox.
+# We avoid sandbox_exec for the check because piped stdin to openshell sandbox connect
+# produces shell prompt noise that causes false positives with grep.
+PLUGIN_EXISTS=false
+if run_cmd "openshell sandbox download '$SANDBOX_NAME' /sandbox/.openclaw/extensions/openclaw-mem0/package.json /tmp/_mem0_plugin_check.json" &>/dev/null; then
+  if [[ -f /tmp/_mem0_plugin_check.json ]] || run_cmd "test -f /tmp/_mem0_plugin_check.json" &>/dev/null; then
+    PLUGIN_EXISTS=true
+  fi
+fi
+rm -f /tmp/_mem0_plugin_check.json 2>/dev/null || true
+run_cmd "rm -f /tmp/_mem0_plugin_check.json" 2>/dev/null || true
+
+if $PLUGIN_EXISTS; then
+  success "Mem0 plugin already installed in sandbox"
+else
+  info "Downloading $PLUGIN_PKG..."
+
+  # Download and build outside the sandbox (on host or in container)
+  run_cmd "cd /tmp && rm -rf openclaw-mem0-full mem0-openclaw-mem0-*.tgz openclaw-mem0-full.tgz"
+
+  if ! run_cmd "cd /tmp && npm pack '$PLUGIN_PKG' 2>/dev/null"; then
+    die "Failed to download $PLUGIN_PKG from npm. Check your internet connection."
+  fi
+
+  success "Downloaded plugin"
+
+  info "Installing plugin dependencies..."
+  run_cmd "mkdir -p /tmp/openclaw-mem0-full && cd /tmp/openclaw-mem0-full && tar xzf /tmp/mem0-openclaw-mem0-*.tgz --strip-components=1 && npm install --omit=dev 2>&1 | tail -3"
+
+  success "Dependencies installed"
+
+  info "Uploading plugin to sandbox..."
+  run_cmd "cd /tmp && tar czf openclaw-mem0-full.tgz -C openclaw-mem0-full ."
+
+  if ! run_cmd "openshell sandbox upload '$SANDBOX_NAME' /tmp/openclaw-mem0-full.tgz /sandbox/openclaw-mem0-full.tgz 2>&1"; then
+    die "Failed to upload plugin to sandbox. Check: openshell sandbox list"
+  fi
+
+  success "Plugin uploaded"
+
+  info "Extracting plugin inside sandbox..."
+  sandbox_exec "mkdir -p ~/.openclaw/extensions/openclaw-mem0 && tar xzf /sandbox/openclaw-mem0-full.tgz/openclaw-mem0-full.tgz -C ~/.openclaw/extensions/openclaw-mem0 2>/dev/null || tar xzf /sandbox/openclaw-mem0-full.tgz -C ~/.openclaw/extensions/openclaw-mem0 2>/dev/null && echo EXTRACT_OK" >/dev/null 2>&1 || true
+
+  # Verify
+  VERIFY_OK=false
+  if run_cmd "openshell sandbox download '$SANDBOX_NAME' /sandbox/.openclaw/extensions/openclaw-mem0/package.json /tmp/_mem0_verify.json" &>/dev/null; then
+    VERIFY_OK=true
+  fi
+  rm -f /tmp/_mem0_verify.json 2>/dev/null || true
+  run_cmd "rm -f /tmp/_mem0_verify.json" 2>/dev/null || true
+  if $VERIFY_OK; then
+    success "Plugin extracted inside sandbox"
+  else
+    warn "Could not verify plugin extraction. You may need to extract manually."
+    if $IS_MACOS; then
+      echo "  docker exec -it $CONTAINER_NAME bash"
+      echo "  export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh"
+    fi
+    echo "  nemoclaw $SANDBOX_NAME connect"
+    echo "  mkdir -p ~/.openclaw/extensions/openclaw-mem0"
+    echo "  tar xzf /sandbox/openclaw-mem0-full.tgz/openclaw-mem0-full.tgz -C ~/.openclaw/extensions/openclaw-mem0"
+  fi
+
+  # Clean up
+  run_cmd "rm -rf /tmp/openclaw-mem0-full /tmp/mem0-openclaw-mem0-*.tgz /tmp/openclaw-mem0-full.tgz" 2>/dev/null || true
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 5: Update Network Policy
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Phase 5: Updating network policy to allow api.mem0.ai and telemetry"
+
+# Find baseline policy
+BASELINE_POLICY=$(run_cmd "find / -path '*/nemoclaw-blueprint/policies/openclaw-sandbox.yaml' 2>/dev/null | head -1" || true)
+
+if [[ -z "$BASELINE_POLICY" ]]; then
+  die "Cannot find NemoClaw baseline policy file. Is NemoClaw installed?"
+fi
+
+info "Baseline policy: $BASELINE_POLICY"
+
+# Check if mem0_api already exists
+HAS_MEM0=$(run_cmd "grep -c mem0_api '$BASELINE_POLICY' 2>/dev/null" || echo "0")
+
+if [[ "$HAS_MEM0" != "0" ]]; then
+  success "mem0_api already in baseline policy"
+else
+  info "Adding api.mem0.ai to network policy..."
+fi
+
+# Create custom policy with mem0_api + telemetry — use node for reliable cross-platform YAML editing
+run_cmd "node -e \"
+const fs = require('fs');
+let c = fs.readFileSync('$BASELINE_POLICY', 'utf8');
+const mem0Block = '\\n  mem0_api:\\n    name: mem0_api\\n    endpoints:\\n      - host: api.mem0.ai\\n        port: 443\\n        access: full\\n    binaries:\\n      - { path: /usr/local/bin/node }\\n      - { path: /usr/local/bin/openclaw }\\n';
+const telemetryBlock = '\\n  mem0_telemetry:\\n    name: mem0_telemetry\\n    endpoints:\\n      - host: us.i.posthog.com\\n        port: 443\\n        access: full\\n    binaries:\\n      - { path: /usr/local/bin/node }\\n      - { path: /usr/local/bin/openclaw }\\n';
+if (!c.includes('mem0_api')) {
+  if (c.includes('# ── Messaging')) {
+    c = c.replace('  # ── Messaging', mem0Block + '\\n  # ── Messaging');
+  } else {
+    c += mem0Block;
+  }
+}
+if (!c.includes('mem0_telemetry')) {
+  if (c.includes('mem0_api:')) {
+    c = c.replace('  mem0_api:', telemetryBlock + '\\n  mem0_api:');
+  } else {
+    c += telemetryBlock;
+  }
+}
+fs.writeFileSync('/tmp/nemoclaw-mem0-policy.yaml', c);
+console.log('ok');
+\""
+
+success "Custom policy file created"
+
+# Apply the policy
+info "Applying network policy..."
+if ! run_cmd "openshell policy set '$SANDBOX_NAME' --policy /tmp/nemoclaw-mem0-policy.yaml --wait 2>&1"; then
+  error "Failed to apply network policy."
+  echo ""
+  echo "  If you see 'sandbox not found', re-run: nemoclaw onboard"
+  echo "  Then re-run this script."
+  echo ""
+  die "Network policy update failed."
+fi
+
+success "Network policy applied — api.mem0.ai and telemetry allowed"
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 6: Configure Plugin
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Phase 6: Configuring Mem0 plugin"
+
+echo ""
+echo -e "  ${DIM}Get your Mem0 API key from: https://app.mem0.ai${NC}"
+echo -e "  ${DIM}The key starts with 'm0-'${NC}"
+echo ""
+ask "Enter your Mem0 API key: "
+read -r MEM0_API_KEY
+
+if [[ -z "$MEM0_API_KEY" ]]; then
+  die "Mem0 API key is required."
+fi
+
+if [[ ! "$MEM0_API_KEY" =~ ^m0- ]]; then
+  warn "Key doesn't start with 'm0-'. Make sure this is correct."
+fi
+
+echo ""
+echo -e "  ${DIM}The user ID scopes all memories. Pick any unique identifier.${NC}"
+echo -e "  ${DIM}Examples: alice, user_123, your-email@example.com${NC}"
+echo ""
+ask "Enter user ID [$MEM0_USER_ID]: "
+read -r custom_user_id
+MEM0_USER_ID="${custom_user_id:-$MEM0_USER_ID}"
+
+info "Configuring plugin inside sandbox..."
+
+CONFIG_SCRIPT="openclaw config set plugins.slots.memory openclaw-mem0 2>&1 | tail -1 && \
+openclaw config set plugins.entries.openclaw-mem0.enabled true 2>&1 | tail -1 && \
+openclaw config set plugins.entries.openclaw-mem0.config.apiKey '$MEM0_API_KEY' 2>&1 | tail -1 && \
+openclaw config set plugins.entries.openclaw-mem0.config.userId '$MEM0_USER_ID' 2>&1 | tail -1 && \
+echo SETUP_DONE"
+
+CONFIG_OUTPUT=$(sandbox_exec "$CONFIG_SCRIPT" || true)
+
+if echo "$CONFIG_OUTPUT" | grep -q "SETUP_DONE"; then
+  success "Plugin configured (mode: platform, user: $MEM0_USER_ID)"
+else
+  warn "Could not verify config. You may need to configure manually:"
+  echo ""
+  if $IS_MACOS; then
+    echo "  docker exec -it $CONTAINER_NAME bash"
+    echo "  export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh"
+  fi
+  echo "  nemoclaw $SANDBOX_NAME connect"
+  echo "  openclaw config set plugins.slots.memory openclaw-mem0"
+  echo "  openclaw config set plugins.entries.openclaw-mem0.enabled true"
+  echo "  openclaw config set plugins.entries.openclaw-mem0.config.apiKey \"$MEM0_API_KEY\""
+  echo "  openclaw config set plugins.entries.openclaw-mem0.config.userId \"$MEM0_USER_ID\""
+  echo ""
+fi
+
+# ═════════════════════════════════════════════════════════════════════════════
+# PHASE 7: Verify & Test
+# ═════════════════════════════════════════════════════════════════════════════
+
+step "Phase 7: Verification"
+
+echo ""
+echo -e "${BOLD}${GREEN}╔══════════════════════════════════════════════════════════════╗${NC}"
+echo -e "${BOLD}${GREEN}║                    Setup Complete!                           ║${NC}"
+echo -e "${BOLD}${GREEN}╚══════════════════════════════════════════════════════════════╝${NC}"
+echo ""
+echo -e "  ${BOLD}Sandbox:${NC}   $SANDBOX_NAME"
+echo -e "  ${BOLD}Plugin:${NC}    @mem0/openclaw-mem0 (platform mode)"
+echo -e "  ${BOLD}User ID:${NC}   $MEM0_USER_ID"
+if $IS_MACOS; then
+  echo -e "  ${BOLD}Container:${NC} $CONTAINER_NAME"
+fi
+echo ""
+echo -e "  ${BOLD}${CYAN}Next steps:${NC}"
+echo ""
+
+if $IS_MACOS; then
+  echo -e "  1. Open a shell in the container:"
+  echo ""
+  echo -e "     ${DIM}docker exec -it $CONTAINER_NAME bash${NC}"
+  echo -e "     ${DIM}export NVM_DIR=/root/.nvm && source /root/.nvm/nvm.sh${NC}"
+  echo ""
+  echo -e "  2. Connect to the sandbox and start the gateway:"
+  echo ""
+  echo -e "     ${DIM}nemoclaw $SANDBOX_NAME connect${NC}"
+  echo -e "     ${DIM}nemoclaw-start${NC}"
+else
+  echo -e "  1. Connect to the sandbox and start the gateway:"
+  echo ""
+  echo -e "     ${DIM}source ~/.bashrc${NC}"
+  echo -e "     ${DIM}nemoclaw $SANDBOX_NAME connect${NC}"
+  echo -e "     ${DIM}nemoclaw-start${NC}"
+fi
+echo ""
+echo -e "  Then verify the plugin loaded (look for 'openclaw-mem0: registered'):"
+echo ""
+echo -e "     ${DIM}openclaw plugins list${NC}"
+echo ""
+echo -e "  Test auto-capture (storing memories):"
+echo ""
+echo -e "     ${DIM}openclaw agent --agent main --local -m \"My name is Alice\" --session-id test1${NC}"
+echo ""
+echo -e "  Test auto-recall (new session, memories should appear):"
+echo ""
+echo -e "     ${DIM}openclaw agent --agent main --local -m \"What do you know about me?\" --session-id test2${NC}"
+echo ""
+echo -e "  Or use the interactive TUI:"
+echo ""
+echo -e "     ${DIM}openclaw tui${NC}"
+echo ""
+echo -e "  ${YELLOW}Note:${NC} You may see 'Telemetry event capture failed' errors."
+echo -e "  These are harmless and do not affect memory functionality."
+echo ""
+echo -e "  ${BOLD}Documentation:${NC} https://docs.mem0.ai"
+echo -e "  ${BOLD}Plugin source:${NC} https://www.npmjs.com/package/@mem0/openclaw-mem0"
+echo ""

mem0-ts/src/oss/src/config/defaults.ts

@@ -34,12 +34,6 @@ export const DEFAULT_MEMORY_CONFIG: MemoryConfig = {
       username: process.env.NEO4J_USERNAME || "neo4j",
       password: process.env.NEO4J_PASSWORD || "password",
     },
-    llm: {
-      provider: "openai",
-      config: {
-        model: "gpt-4-turbo-preview",
-      },
-    },
   },
   historyStore: {
     provider: "sqlite",

mem0-ts/src/oss/src/index.ts

@@ -26,4 +26,5 @@ export * from "./vector_stores/supabase";
 export * from "./vector_stores/langchain";
 export * from "./vector_stores/vectorize";
 export * from "./vector_stores/azure_ai_search";
+export * from "./vector_stores/pgvector";
 export * from "./utils/factory";

mem0-ts/src/oss/src/memory/graph_memory.ts

@@ -80,18 +80,18 @@ export class MemoryGraph {
     );
 
     this.llmProvider = "openai";
+    let llmConfig = this.config.llm.config;
+
     if (this.config.llm?.provider) {
       this.llmProvider = this.config.llm.provider;
     }
     if (this.config.graphStore?.llm?.provider) {
       this.llmProvider = this.config.graphStore.llm.provider;
+      llmConfig = this.config.graphStore.llm.config ?? llmConfig;
     }
 
-    this.llm = LLMFactory.create(this.llmProvider, this.config.llm.config);
-    this.structuredLlm = LLMFactory.create(
-      this.llmProvider,
-      this.config.llm.config,
-    );
+    this.llm = LLMFactory.create(this.llmProvider, llmConfig);
+    this.structuredLlm = LLMFactory.create(this.llmProvider, llmConfig);
     this.threshold = 0.7;
   }
 

mem0-ts/src/oss/src/prompts/index.ts

@@ -31,6 +31,7 @@ export const MemoryUpdateSchema = z.object({
         old_memory: z
           .string()
           .optional()
+          .nullable()
           .describe(
             "The previous content of the memory item if the event was UPDATE.",
           ),

mem0-ts/src/oss/src/utils/factory.ts

@@ -34,6 +34,7 @@ import { LangchainLLM } from "../llms/langchain";
 import { LangchainEmbedder } from "../embeddings/langchain";
 import { LangchainVectorStore } from "../vector_stores/langchain";
 import { AzureAISearch } from "../vector_stores/azure_ai_search";
+import { PGVector } from "../vector_stores/pgvector";
 
 export class EmbedderFactory {
   static create(provider: string, config: EmbeddingConfig): Embedder {
@@ -104,6 +105,8 @@ export class VectorStoreFactory {
         return new VectorizeDB(config as any);
       case "azure-ai-search":
         return new AzureAISearch(config as any);
+      case "pgvector":
+        return new PGVector(config as any);
       default:
         throw new Error(`Unsupported vector store provider: ${provider}`);
     }

mem0-ts/src/oss/src/vector_stores/pgvector.ts

@@ -1,4 +1,4 @@
-import { Client, Pool } from "pg";
+import { Client } from "pg";
 import { VectorStore } from "./base";
 import { SearchFilters, VectorStoreConfig, VectorStoreResult } from "../types";
 
@@ -35,6 +35,7 @@ export class PGVector implements VectorStore {
       host: config.host,
       port: config.port,
     });
+    this.initialize().catch(console.error);
   }
 
   async initialize(): Promise<void> {

mem0-ts/src/oss/src/vector_stores/redis.ts

@@ -522,7 +522,7 @@ export class RedisDB implements VectorStore {
 
       return {
         id: vectorId,
-        payload,
+        payload: toCamelCase(payload),
       };
     } catch (error) {
       console.error("Error getting vector:", error);

mem0-ts/src/oss/tests/config-manager.test.ts

@@ -349,6 +349,78 @@ describe("ConfigManager", () => {
   });
 });
 
+// ─────────────────────────────────────────────────────────────────────────
+// Graph store LLM config propagation (issue #3425)
+// ─────────────────────────────────────────────────────────────────────────
+describe("mergeConfig - graph store LLM config (issue #3425)", () => {
+  const baseEmbedder = {
+    provider: "openai",
+    config: { apiKey: "test-key" },
+  };
+  const baseVectorStore = {
+    provider: "memory",
+    config: { collectionName: "test" },
+  };
+  const graphStoreNeo4j = {
+    provider: "neo4j",
+    config: {
+      url: "neo4j://localhost:7687",
+      username: "neo4j",
+      password: "password",
+    },
+  };
+
+  it("should NOT have a default graphStore.llm — root llm should be the fallback", () => {
+    const config = ConfigManager.mergeConfig({
+      embedder: baseEmbedder,
+      vectorStore: baseVectorStore,
+      llm: {
+        provider: "anthropic",
+        config: { model: "claude-sonnet-4-20250514" },
+      },
+      graphStore: graphStoreNeo4j,
+    });
+
+    // graphStore should NOT have its own llm after merge
+    expect(config.graphStore?.llm).toBeUndefined();
+    // root llm should be anthropic
+    expect(config.llm.provider).toBe("anthropic");
+    expect(config.llm.config.model).toBe("claude-sonnet-4-20250514");
+  });
+
+  it("should preserve explicit graphStore.llm when user provides it", () => {
+    const config = ConfigManager.mergeConfig({
+      embedder: baseEmbedder,
+      vectorStore: baseVectorStore,
+      llm: {
+        provider: "anthropic",
+        config: { model: "claude-sonnet-4-20250514" },
+      },
+      graphStore: {
+        ...graphStoreNeo4j,
+        llm: { provider: "openai", config: { model: "gpt-4o" } },
+      },
+    });
+
+    // graphStore should have its own llm
+    expect(config.graphStore?.llm?.provider).toBe("openai");
+    expect(config.graphStore?.llm?.config).toEqual({ model: "gpt-4o" });
+    // root llm should still be anthropic
+    expect(config.llm.provider).toBe("anthropic");
+  });
+
+  it("should not have graphStore.llm when user does not provide one", () => {
+    const config = ConfigManager.mergeConfig({
+      embedder: baseEmbedder,
+      vectorStore: baseVectorStore,
+      llm: { provider: "openai", config: { model: "gpt-4o" } },
+    });
+
+    // Default graphStore should not have llm
+    expect(config.graphStore?.llm).toBeUndefined();
+  });
+});
+
 // ─────────────────────────────────────────────────────────────────────────
 // Memory class – LM Studio end-to-end flow (mocked factories)
 // ─────────────────────────────────────────────────────────────────────────

mem0-ts/src/oss/tests/factory.unit.test.ts

@@ -118,6 +118,11 @@ jest.mock("../src/vector_stores/azure_ai_search", () => ({
     .fn()
     .mockImplementation((config) => ({ type: "azure-ai-search", config })),
 }));
+jest.mock("../src/vector_stores/pgvector", () => ({
+  PGVector: jest
+    .fn()
+    .mockImplementation((config) => ({ type: "pgvector", config })),
+}));
 jest.mock("../src/storage/SupabaseHistoryManager", () => ({
   SupabaseHistoryManager: jest
     .fn()
@@ -236,6 +241,7 @@ describe("VectorStoreFactory", () => {
     ["langchain"],
     ["vectorize"],
     ["azure-ai-search"],
+    ["pgvector"],
   ])("creates vector store for provider '%s'", (provider) => {
     expect(() =>
       VectorStoreFactory.create(provider, dummyVSConfig),

mem0-ts/src/oss/tests/graph-memory-parsing.test.ts

@@ -511,6 +511,137 @@ describe("Prompt construction — all json_object sites include 'json'", () => {
 // 5. Edge cases – malformed entity fields in _removeSpacesFromEntities
 // ═══════════════════════════════════════════════════════════════════════════
 
+// ═══════════════════════════════════════════════════════════════════════════
+// 5a. LLM config propagation — graph store uses correct provider & config
+//     Regression test for https://github.com/mem0ai/mem0/issues/3425
+// ═══════════════════════════════════════════════════════════════════════════
+
+describe("LLM config propagation to graph store (issue #3425)", () => {
+  const { LLMFactory } = require("../src/utils/factory");
+
+  beforeEach(() => {
+    (LLMFactory.create as jest.Mock).mockClear();
+  });
+
+  it("uses root llm config when no graphStore.llm is provided", () => {
+    const config = {
+      graphStore: {
+        config: {
+          url: "bolt://localhost:7687",
+          username: "neo4j",
+          password: "test",
+        },
+      },
+      embedder: { provider: "openai", config: {} },
+      llm: {
+        provider: "anthropic",
+        config: { model: "claude-sonnet-4-20250514", apiKey: "sk-ant-test" },
+      },
+    } as any;
+
+    new MemoryGraph(config);
+
+    expect(LLMFactory.create).toHaveBeenCalledWith("anthropic", {
+      model: "claude-sonnet-4-20250514",
+      apiKey: "sk-ant-test",
+    });
+    // Both llm and structuredLlm should use the same config
+    expect(LLMFactory.create).toHaveBeenCalledTimes(2);
+    expect(LLMFactory.create).toHaveBeenNthCalledWith(1, "anthropic", {
+      model: "claude-sonnet-4-20250514",
+      apiKey: "sk-ant-test",
+    });
+    expect(LLMFactory.create).toHaveBeenNthCalledWith(2, "anthropic", {
+      model: "claude-sonnet-4-20250514",
+      apiKey: "sk-ant-test",
+    });
+  });
+
+  it("uses graphStore.llm config when provided, overriding root llm", () => {
+    const config = {
+      graphStore: {
+        config: {
+          url: "bolt://localhost:7687",
+          username: "neo4j",
+          password: "test",
+        },
+        llm: {
+          provider: "openai",
+          config: { model: "gpt-4o", apiKey: "sk-openai-test" },
+        },
+      },
+      embedder: { provider: "openai", config: {} },
+      llm: {
+        provider: "anthropic",
+        config: { model: "claude-sonnet-4-20250514", apiKey: "sk-ant-test" },
+      },
+    } as any;
+
+    new MemoryGraph(config);
+
+    // Should use graphStore.llm, NOT root llm
+    expect(LLMFactory.create).toHaveBeenNthCalledWith(1, "openai", {
+      model: "gpt-4o",
+      apiKey: "sk-openai-test",
+    });
+    expect(LLMFactory.create).toHaveBeenNthCalledWith(2, "openai", {
+      model: "gpt-4o",
+      apiKey: "sk-openai-test",
+    });
+  });
+
+  it("falls back to root llm config when graphStore.llm.config is undefined", () => {
+    // Note: in practice, Zod schema requires config when graphStore.llm is
+    // present. This tests the defensive fallback in MemoryGraph itself.
+    const config = {
+      graphStore: {
+        config: {
+          url: "bolt://localhost:7687",
+          username: "neo4j",
+          password: "test",
+        },
+        llm: {
+          provider: "openai",
+          // config explicitly undefined
+          config: undefined,
+        },
+      },
+      embedder: { provider: "openai", config: {} },
+      llm: {
+        provider: "anthropic",
+        config: { model: "claude-sonnet-4-20250514" },
+      },
+    } as any;
+
+    new MemoryGraph(config);
+
+    // Provider from graphStore.llm, but config falls back to root llm.config
+    expect(LLMFactory.create).toHaveBeenNthCalledWith(1, "openai", {
+      model: "claude-sonnet-4-20250514",
+    });
+  });
+
+  it("defaults to openai when neither root nor graphStore llm provider is set", () => {
+    const config = {
+      graphStore: {
+        config: {
+          url: "bolt://localhost:7687",
+          username: "neo4j",
+          password: "test",
+        },
+      },
+      embedder: { provider: "openai", config: {} },
+      llm: { config: { model: "gpt-4" } },
+    } as any;
+
+    new MemoryGraph(config);
+
+    expect(LLMFactory.create).toHaveBeenNthCalledWith(1, "openai", {
+      model: "gpt-4",
+    });
+  });
+});
+
 describe("_removeSpacesFromEntities (via _establishNodesRelationsFromData)", () => {
   it("normalises spaces and case in entity source/relationship/destination", async () => {
     mockGenerateResponse.mockResolvedValueOnce({

mem0/configs/llms/anthropic.py

@@ -16,7 +16,7 @@ class AnthropicConfig(BaseLlmConfig):
         temperature: float = 0.1,
         api_key: Optional[str] = None,
         max_tokens: int = 2000,
-        top_p: float = 0.1,
+        top_p: Optional[float] = None,
         top_k: int = 1,
         enable_vision: bool = False,
         vision_details: Optional[str] = "auto",
@@ -32,7 +32,7 @@ class AnthropicConfig(BaseLlmConfig):
             temperature: Controls randomness, defaults to 0.1
             api_key: Anthropic API key, defaults to None
             max_tokens: Maximum tokens to generate, defaults to 2000
-            top_p: Nucleus sampling parameter, defaults to 0.1
+            top_p: Nucleus sampling parameter, defaults to None (omitted to avoid conflict with temperature)
             top_k: Top-k sampling parameter, defaults to 1
             enable_vision: Enable vision capabilities, defaults to False
             vision_details: Vision detail level, defaults to "auto"

mem0/configs/llms/minimax.py

@@ -0,0 +1,56 @@
+from typing import Optional
+
+from mem0.configs.llms.base import BaseLlmConfig
+
+
+class MinimaxConfig(BaseLlmConfig):
+    """
+    Configuration class for MiniMax-specific parameters.
+    Inherits from BaseLlmConfig and adds MiniMax-specific settings.
+    """
+
+    def __init__(
+        self,
+        # Base parameters
+        model: Optional[str] = None,
+        temperature: float = 0.1,
+        api_key: Optional[str] = None,
+        max_tokens: int = 2000,
+        top_p: float = 0.1,
+        top_k: int = 1,
+        enable_vision: bool = False,
+        vision_details: Optional[str] = "auto",
+        http_client_proxies: Optional[dict] = None,
+        # MiniMax-specific parameters
+        minimax_base_url: Optional[str] = None,
+    ):
+        """
+        Initialize MiniMax configuration.
+
+        Args:
+            model: MiniMax model to use, defaults to None
+            temperature: Controls randomness, defaults to 0.1
+            api_key: MiniMax API key, defaults to None
+            max_tokens: Maximum tokens to generate, defaults to 2000
+            top_p: Nucleus sampling parameter, defaults to 0.1
+            top_k: Top-k sampling parameter, defaults to 1
+            enable_vision: Enable vision capabilities, defaults to False
+            vision_details: Vision detail level, defaults to "auto"
+            http_client_proxies: HTTP client proxy settings, defaults to None
+            minimax_base_url: MiniMax API base URL, defaults to None
+        """
+        # Initialize base parameters
+        super().__init__(
+            model=model,
+            temperature=temperature,
+            api_key=api_key,
+            max_tokens=max_tokens,
+            top_p=top_p,
+            top_k=top_k,
+            enable_vision=enable_vision,
+            vision_details=vision_details,
+            http_client_proxies=http_client_proxies,
+        )
+
+        # MiniMax-specific parameters
+        self.minimax_base_url = minimax_base_url

mem0/configs/vector_stores/chroma.py

@@ -32,8 +32,8 @@ class ChromaDbConfig(BaseModel):
             values.pop("path", None)
             return values
         
-        # Check if local/server configuration is provided (excluding default tmp path for cloud config)
-        local_config = bool(path and path != "/tmp/chroma") or bool(host and port)
+        # Check if local/server configuration is provided
+        local_config = bool(path) or bool(host and port)
         
         if not cloud_config and not local_config:
             raise ValueError("Either ChromaDB Cloud configuration (api_key, tenant) or local configuration (path or host/port) must be provided.")

mem0/configs/vector_stores/qdrant.py

@@ -16,7 +16,7 @@ class QdrantConfig(BaseModel):
     path: Optional[str] = Field("/tmp/qdrant", description="Path for local Qdrant database")
     url: Optional[str] = Field(None, description="Full URL for Qdrant server")
     api_key: Optional[str] = Field(None, description="API key for Qdrant server")
-    on_disk: Optional[bool] = Field(False, description="Enables persistent storage")
+    on_disk: Optional[bool] = Field(False,description="Enables persistent storage. Vectors are kept on disk (True) or in memory (False). Does not delete the local database path.")
 
     @model_validator(mode="before")
     @classmethod

mem0/configs/vector_stores/turbopuffer.py

@@ -0,0 +1,45 @@
+import os
+from typing import Any, Dict, Optional
+
+from pydantic import BaseModel, ConfigDict, Field, model_validator
+
+
+class TurbopufferConfig(BaseModel):
+    collection_name: str = Field("mem0", description="Name of the namespace/collection")
+    embedding_model_dims: int = Field(1536, description="Dimensions of the embedding model")
+    api_key: Optional[str] = Field(None, description="API key for Turbopuffer")
+    region: str = Field("gcp-us-central1", description="Turbopuffer region (e.g., 'gcp-us-central1', 'aws-us-west-2')")
+    distance_metric: str = Field(
+        "cosine_distance",
+        description="Distance metric for vector similarity ('cosine_distance' or 'euclidean_squared')",
+    )
+    batch_size: int = Field(100, description="Batch size for bulk operations")
+    extra_params: Optional[Dict[str, Any]] = Field(
+        None,
+        description="Additional parameters for Turbopuffer client",
+    )
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_api_key(cls, values: Dict[str, Any]) -> Dict[str, Any]:
+        api_key = values.get("api_key")
+        if not api_key and "TURBOPUFFER_API_KEY" not in os.environ:
+            raise ValueError(
+                "Either 'api_key' must be provided or TURBOPUFFER_API_KEY environment variable must be set."
+            )
+        return values
+
+    @model_validator(mode="before")
+    @classmethod
+    def validate_extra_fields(cls, values: Dict[str, Any]) -> Dict[str, Any]:
+        allowed_fields = set(cls.model_fields.keys())
+        input_fields = set(values.keys())
+        extra_fields = input_fields - allowed_fields
+        if extra_fields:
+            raise ValueError(
+                f"Extra fields not allowed: {', '.join(extra_fields)}. "
+                f"Please input only the following fields: {', '.join(allowed_fields)}"
+            )
+        return values
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)

mem0/graphs/configs.py

@@ -77,12 +77,32 @@ class KuzuConfig(BaseModel):
     db: Optional[str] = Field(":memory:", description="Path to a Kuzu database file")
 
 
+class ApacheAgeConfig(BaseModel):
+    host: Optional[str] = Field("localhost", description="PostgreSQL server hostname")
+    port: Optional[int] = Field(5432, description="PostgreSQL server port")
+    database: Optional[str] = Field(None, description="PostgreSQL database name")
+    username: Optional[str] = Field(None, description="PostgreSQL username")
+    password: Optional[str] = Field(None, description="PostgreSQL password")
+    graph_name: Optional[str] = Field("mem0_graph", description="Name of the Apache AGE graph")
+
+    @model_validator(mode="before")
+    def check_required_fields(cls, values):
+        database, username, password = (
+            values.get("database"),
+            values.get("username"),
+            values.get("password"),
+        )
+        if not database or not username or not password:
+            raise ValueError("Please provide 'database', 'username' and 'password'.")
+        return values
+
+
 class GraphStoreConfig(BaseModel):
     provider: str = Field(
-        description="Provider of the data store (e.g., 'neo4j', 'memgraph', 'neptune', 'kuzu')",
+        description="Provider of the data store (e.g., 'neo4j', 'memgraph', 'neptune', 'kuzu', 'apache_age')",
         default="neo4j",
     )
-    config: Union[Neo4jConfig, MemgraphConfig, NeptuneConfig, KuzuConfig] = Field(
+    config: Union[Neo4jConfig, MemgraphConfig, NeptuneConfig, KuzuConfig, ApacheAgeConfig] = Field(
         description="Configuration for the specific data store", default=None
     )
     llm: Optional[LlmConfig] = Field(description="LLM configuration for querying the graph store", default=None)
@@ -110,5 +130,7 @@ class GraphStoreConfig(BaseModel):
             return NeptuneConfig(**v.model_dump())
         elif provider == "kuzu":
             return KuzuConfig(**v.model_dump())
+        elif provider == "apache_age":
+            return ApacheAgeConfig(**v.model_dump())
         else:
             raise ValueError(f"Unsupported graph store provider: {provider}")

mem0/graphs/neptune/base.py

@@ -409,6 +409,28 @@ class NeptuneBase(ABC):
         """
         pass
 
+    def delete(self, data, filters):
+        """
+        Delete graph entities associated with the given memory text.
+
+        Extracts entities and relationships from the memory text using the same
+        pipeline as add(), then deletes the matching relationships in the graph.
+
+        Args:
+            data (str): The memory text whose graph entities should be removed.
+            filters (dict): Scope filters (user_id, agent_id, run_id).
+        """
+        try:
+            entity_type_map = self._retrieve_nodes_from_data(data, filters)
+            if not entity_type_map:
+                logger.debug("No entities found in memory text, skipping graph cleanup")
+                return
+            to_be_deleted = self._establish_nodes_relations_from_data(data, filters, entity_type_map)
+            if to_be_deleted:
+                self._delete_entities(to_be_deleted, filters["user_id"])
+        except Exception as e:
+            logger.error(f"Error during graph cleanup for memory delete: {e}")
+
     def delete_all(self, filters):
         cypher, params = self._delete_all_cypher(filters)
         self.graph.query(cypher, params=params)

mem0/llms/anthropic.py

@@ -40,6 +40,31 @@ class AnthropicLLM(LLMBase):
         api_key = self.config.api_key or os.getenv("ANTHROPIC_API_KEY")
         self.client = anthropic.Anthropic(api_key=api_key)
 
+    def _get_common_params(self, **kwargs) -> Dict:
+        """Get common parameters, avoiding sending both temperature and top_p together.
+
+        Anthropic rejects requests that include both temperature and top_p.
+        When both are set, we keep temperature and drop top_p.
+        """
+        params = {}
+
+        if self.config.max_tokens is not None:
+            params["max_tokens"] = self.config.max_tokens
+
+        has_temperature = self.config.temperature is not None
+        has_top_p = self.config.top_p is not None
+
+        if has_temperature and has_top_p:
+            # Anthropic forbids both; prefer temperature
+            params["temperature"] = self.config.temperature
+        elif has_temperature:
+            params["temperature"] = self.config.temperature
+        elif has_top_p:
+            params["top_p"] = self.config.top_p
+
+        params.update(kwargs)
+        return params
+
     def generate_response(
         self,
         messages: List[Dict[str, str]],

mem0/llms/azure_openai_structured.py

@@ -1,3 +1,4 @@
+import json
 import os
 from typing import Dict, List, Optional
 
@@ -6,6 +7,7 @@ from openai import AzureOpenAI
 
 from mem0.configs.llms.base import BaseLlmConfig
 from mem0.llms.base import LLMBase
+from mem0.memory.utils import extract_json
 
 SCOPE = "https://cognitiveservices.azure.com/.default"
 
@@ -83,9 +85,35 @@ class AzureOpenAIStructuredLLM(LLMBase):
             params["tools"] = tools
             params["tool_choice"] = tool_choice
 
-        if tools:
-            params["tools"] = tools
-            params["tool_choice"] = tool_choice
-
         response = self.client.chat.completions.create(**params)
         return self._parse_response(response, tools)
+
+    def _parse_response(self, response, tools):
+        """
+        Process the response based on whether tools are used or not.
+
+        Args:
+            response: The raw response from API.
+            tools: The list of tools provided in the request.
+
+        Returns:
+            str or dict: The processed response.
+        """
+        if tools:
+            processed_response = {
+                "content": response.choices[0].message.content,
+                "tool_calls": [],
+            }
+
+            if response.choices[0].message.tool_calls:
+                for tool_call in response.choices[0].message.tool_calls:
+                    processed_response["tool_calls"].append(
+                        {
+                            "name": tool_call.function.name,
+                            "arguments": json.loads(extract_json(tool_call.function.arguments)),
+                        }
+                    )
+
+            return processed_response
+        else:
+            return response.choices[0].message.content

mem0/llms/configs.py

@@ -23,6 +23,7 @@ class LlmConfig(BaseModel):
             "azure_openai_structured",
             "gemini",
             "deepseek",
+            "minimax",
             "xai",
             "sarvam",
             "lmstudio",

mem0/llms/gemini.py

@@ -32,22 +32,25 @@ class GeminiLLM(LLMBase):
         Returns:
             str or dict: The processed response.
         """
+        # Get parts safely — content can be None when Gemini blocks the response
+        candidate = response.candidates[0] if response.candidates else None
+        parts = candidate.content.parts if candidate and candidate.content else None
+
         if tools:
             processed_response = {
                 "content": None,
                 "tool_calls": [],
             }
 
-            # Extract content from the first candidate
-            if response.candidates and response.candidates[0].content.parts:
-                for part in response.candidates[0].content.parts:
+            if parts:
+                # Extract content from the first candidate
+                for part in parts:
                     if hasattr(part, "text") and part.text:
                         processed_response["content"] = part.text
                         break
 
-            # Extract function calls
-            if response.candidates and response.candidates[0].content.parts:
-                for part in response.candidates[0].content.parts:
+                # Extract function calls
+                for part in parts:
                     if hasattr(part, "function_call") and part.function_call:
                         fn = part.function_call
                         processed_response["tool_calls"].append(
@@ -59,8 +62,8 @@ class GeminiLLM(LLMBase):
 
             return processed_response
         else:
-            if response.candidates and response.candidates[0].content.parts:
-                for part in response.candidates[0].content.parts:
+            if parts:
+                for part in parts:
                     if hasattr(part, "text") and part.text:
                         return part.text
             return ""

mem0/llms/minimax.py

@@ -0,0 +1,114 @@
+import json
+import os
+from typing import Dict, List, Optional, Union
+
+from openai import OpenAI
+
+from mem0.configs.llms.base import BaseLlmConfig
+from mem0.configs.llms.minimax import MinimaxConfig
+from mem0.llms.base import LLMBase
+from mem0.memory.utils import extract_json
+
+
+class MiniMaxLLM(LLMBase):
+    def __init__(self, config: Optional[Union[BaseLlmConfig, MinimaxConfig, Dict]] = None):
+        # Convert to MinimaxConfig if needed
+        if config is None:
+            config = MinimaxConfig()
+        elif isinstance(config, dict):
+            config = MinimaxConfig(**config)
+        elif isinstance(config, BaseLlmConfig) and not isinstance(config, MinimaxConfig):
+            # Convert BaseLlmConfig to MinimaxConfig
+            config = MinimaxConfig(
+                model=config.model,
+                temperature=config.temperature,
+                api_key=config.api_key,
+                max_tokens=config.max_tokens,
+                top_p=config.top_p,
+                top_k=config.top_k,
+                enable_vision=config.enable_vision,
+                vision_details=config.vision_details,
+                http_client_proxies=config.http_client,
+            )
+
+        super().__init__(config)
+
+        if not self.config.model:
+            self.config.model = "MiniMax-M2.7"
+
+        api_key = self.config.api_key or os.getenv("MINIMAX_API_KEY")
+        base_url = (
+            self.config.minimax_base_url
+            or os.getenv("MINIMAX_API_BASE")
+            or "https://api.minimax.io/v1"
+        )
+        self.client = OpenAI(api_key=api_key, base_url=base_url)
+
+    def _parse_response(self, response, tools):
+        """
+        Process the response based on whether tools are used or not.
+
+        Args:
+            response: The raw response from API.
+            tools: The list of tools provided in the request.
+
+        Returns:
+            str or dict: The processed response.
+        """
+        if tools:
+            processed_response = {
+                "content": response.choices[0].message.content,
+                "tool_calls": [],
+            }
+
+            if response.choices[0].message.tool_calls:
+                for tool_call in response.choices[0].message.tool_calls:
+                    processed_response["tool_calls"].append(
+                        {
+                            "name": tool_call.function.name,
+                            "arguments": json.loads(extract_json(tool_call.function.arguments)),
+                        }
+                    )
+
+            return processed_response
+        else:
+            return response.choices[0].message.content
+
+    def generate_response(
+        self,
+        messages: List[Dict[str, str]],
+        response_format=None,
+        tools: Optional[List[Dict]] = None,
+        tool_choice: str = "auto",
+        **kwargs,
+    ):
+        """
+        Generate a response based on the given messages using MiniMax.
+
+        Args:
+            messages (list): List of message dicts containing 'role' and 'content'.
+            response_format (str or object, optional): Format of the response. Defaults to None.
+            tools (list, optional): List of tools that the model can call. Defaults to None.
+            tool_choice (str, optional): Tool choice method. Defaults to "auto".
+            **kwargs: Additional MiniMax-specific parameters.
+
+        Returns:
+            str: The generated response.
+        """
+        params = self._get_supported_params(messages=messages, **kwargs)
+        params.update(
+            {
+                "model": self.config.model,
+                "messages": messages,
+            }
+        )
+
+        if response_format:
+            params["response_format"] = response_format
+
+        if tools:
+            params["tools"] = tools
+            params["tool_choice"] = tool_choice
+
+        response = self.client.chat.completions.create(**params)
+        return self._parse_response(response, tools)

mem0/memory/apache_age_memory.py

@@ -0,0 +1,595 @@
+import json
+import logging
+import time
+
+from mem0.memory.utils import format_entities, sanitize_relationship_for_cypher
+
+try:
+    import age
+except ImportError:
+    raise ImportError("apache-age-python is not installed. Please install it using pip install apache-age-python")
+
+try:
+    from rank_bm25 import BM25Okapi
+except ImportError:
+    raise ImportError("rank_bm25 is not installed. Please install it using pip install rank-bm25")
+
+from mem0.graphs.tools import (
+    DELETE_MEMORY_STRUCT_TOOL_GRAPH,
+    DELETE_MEMORY_TOOL_GRAPH,
+    EXTRACT_ENTITIES_STRUCT_TOOL,
+    EXTRACT_ENTITIES_TOOL,
+    RELATIONS_STRUCT_TOOL,
+    RELATIONS_TOOL,
+)
+from mem0.graphs.utils import EXTRACT_RELATIONS_PROMPT, get_delete_messages
+from mem0.utils.factory import EmbedderFactory, LlmFactory
+
+logger = logging.getLogger(__name__)
+
+
+def _cosine_similarity(vec1, vec2):
+    """Compute cosine similarity between two vectors without numpy."""
+    dot = sum(a * b for a, b in zip(vec1, vec2))
+    norm1 = sum(a * a for a in vec1) ** 0.5
+    norm2 = sum(b * b for b in vec2) ** 0.5
+    if norm1 == 0 or norm2 == 0:
+        return 0.0
+    return dot / (norm1 * norm2)
+
+
+def _get_similar_nodes(nodes, query_embedding, filters, threshold):
+    """Find nodes above the similarity threshold from a fetched node list.
+
+    Shared by ``_find_similar_node`` and ``_search_graph_db`` to avoid
+    duplicating the client-side cosine similarity logic.
+    """
+    matches = []
+    for node in nodes:
+        props = node if isinstance(node, dict) else {}
+        stored_emb = props.get("embedding")
+        if not stored_emb:
+            continue
+        if isinstance(stored_emb, str):
+            stored_emb = json.loads(stored_emb)
+
+        if filters.get("agent_id") and props.get("agent_id") != filters["agent_id"]:
+            continue
+        if filters.get("run_id") and props.get("run_id") != filters["run_id"]:
+            continue
+
+        sim = _cosine_similarity(query_embedding, stored_emb)
+        if sim >= threshold:
+            matches.append({"name": props.get("name"), "similarity": sim, "props": props})
+
+    matches.sort(key=lambda x: x["similarity"], reverse=True)
+    return matches
+
+
+class MemoryGraph:
+    def __init__(self, config):
+        self.config = config
+
+        graph_cfg = self.config.graph_store.config
+        self.graph_name = graph_cfg.graph_name
+
+        # Connect using the Apache AGE Python driver (psycopg2-based)
+        self.ag = age.connect(
+            graph=self.graph_name,
+            host=graph_cfg.host,
+            port=graph_cfg.port,
+            dbname=graph_cfg.database,
+            user=graph_cfg.username,
+            password=graph_cfg.password,
+        )
+
+        self.embedding_model = EmbedderFactory.create(
+            self.config.embedder.provider, self.config.embedder.config, self.config.vector_store.config
+        )
+
+        # Default to openai if no specific provider is configured
+        self.llm_provider = "openai"
+        if self.config.llm and self.config.llm.provider:
+            self.llm_provider = self.config.llm.provider
+        if self.config.graph_store and self.config.graph_store.llm and self.config.graph_store.llm.provider:
+            self.llm_provider = self.config.graph_store.llm.provider
+
+        # Get LLM config with proper null checks
+        llm_config = None
+        if self.config.graph_store and self.config.graph_store.llm and hasattr(self.config.graph_store.llm, "config"):
+            llm_config = self.config.graph_store.llm.config
+        elif hasattr(self.config.llm, "config"):
+            llm_config = self.config.llm.config
+        self.llm = LlmFactory.create(self.llm_provider, llm_config)
+        self.user_id = None
+        self.threshold = self.config.graph_store.threshold if hasattr(self.config.graph_store, "threshold") else 0.7
+
+    # -- helpers ---------------------------------------------------------------
+
+    def _exec_cypher(self, cypher_stmt, cols=None, params=None):
+        """Execute a Cypher query via the AGE driver and return results.
+
+        Uses ``ag.execCypher`` which delegates to psycopg2's safe parameter
+        substitution (``%s`` placeholders).  The *cols* argument specifies the
+        column names in the ``AS (…)`` clause — when ``None`` the driver
+        defaults to a single ``v agtype`` column.
+
+        When *cols* are provided, returns a list of dicts keyed by column name.
+        When *cols* is ``None``, returns vertex/edge property dicts or raw values.
+        """
+        cursor = self.ag.execCypher(cypher_stmt, cols=cols, params=params)
+        rows = cursor.fetchall()
+        if not rows:
+            return []
+
+        col_names = [desc[0] for desc in cursor.description] if cursor.description else None
+        results = []
+        for row in rows:
+            if col_names and len(col_names) > 1:
+                record = {}
+                for i, col_name in enumerate(col_names):
+                    val = row[i]
+                    if hasattr(val, "properties"):
+                        record[col_name] = val.properties
+                    else:
+                        record[col_name] = val
+                results.append(record)
+            else:
+                val = row[0] if len(row) == 1 else row
+                if hasattr(val, "properties"):
+                    results.append(val.properties)
+                else:
+                    results.append(val)
+        return results
+
+    def _fetch_user_nodes_with_embeddings(self, user_id):
+        """Fetch all nodes with embeddings for a given user_id."""
+        return self._exec_cypher(
+            "MATCH (n {user_id: %s}) WHERE n.embedding IS NOT NULL RETURN n",
+            params=(user_id,),
+        )
+
+    def _find_similar_node(self, embedding, filters, threshold=0.9):
+        """Find the most similar existing node by cosine similarity.
+
+        Apache AGE does not have a built-in vector index, so we fetch all
+        node embeddings matching the filters and compute cosine similarity
+        on the client side.  This is adequate for moderate graph sizes; for
+        very large graphs consider pairing AGE with pgvector.
+        """
+        nodes = self._fetch_user_nodes_with_embeddings(filters["user_id"])
+        matches = _get_similar_nodes(nodes, embedding, filters, threshold)
+        return matches[0]["props"] if matches else None
+
+    def _merge_node(self, user_id, name, embedding, agent_id=None, run_id=None):
+        """Create a node if it doesn't exist, or update mentions if it does.
+
+        Apache AGE does not support ``ON CREATE SET`` / ``ON MATCH SET``, so
+        we use ``MERGE … SET`` which always applies the SET clause.  Embeddings
+        and optional filter properties are set in a single query.
+        """
+        set_parts = [
+            "n.embedding = %s",
+            "n.mentions = coalesce(n.mentions, 0) + 1",
+            "n.created = coalesce(n.created, %s)",
+        ]
+        params = [user_id, name, json.dumps(embedding), int(time.time() * 1000)]
+
+        if agent_id:
+            set_parts.append("n.agent_id = %s")
+            params.append(agent_id)
+        if run_id:
+            set_parts.append("n.run_id = %s")
+            params.append(run_id)
+
+        set_clause = ", ".join(set_parts)
+        self._exec_cypher(
+            f"MERGE (n {{user_id: %s, name: %s}}) SET {set_clause}",
+            params=tuple(params),
+        )
+
+    def close(self):
+        """Close the underlying database connection."""
+        if self.ag:
+            self.ag.close()
+
+    # -- public API ------------------------------------------------------------
+
+    def add(self, data, filters):
+        """
+        Adds data to the graph.
+
+        Args:
+            data (str): The data to add to the graph.
+            filters (dict): A dictionary containing filters to be applied during the addition.
+        """
+        entity_type_map = self._retrieve_nodes_from_data(data, filters)
+        to_be_added = self._establish_nodes_relations_from_data(data, filters, entity_type_map)
+        search_output = self._search_graph_db(node_list=list(entity_type_map.keys()), filters=filters)
+        to_be_deleted = self._get_delete_entities_from_search_output(search_output, data, filters)
+
+        deleted_entities = self._delete_entities(to_be_deleted, filters)
+        added_entities = self._add_entities(to_be_added, filters, entity_type_map)
+
+        return {"deleted_entities": deleted_entities, "added_entities": added_entities}
+
+    def search(self, query, filters, limit=100):
+        """
+        Search for memories and related graph data.
+
+        Args:
+            query (str): Query to search for.
+            filters (dict): A dictionary containing filters to be applied during the search.
+            limit (int): The maximum number of nodes and relationships to retrieve. Defaults to 100.
+
+        Returns:
+            list: A list of dicts with keys "source", "relationship", "destination".
+        """
+        entity_type_map = self._retrieve_nodes_from_data(query, filters)
+        search_output = self._search_graph_db(node_list=list(entity_type_map.keys()), filters=filters)
+
+        if not search_output:
+            return []
+
+        search_outputs_sequence = [
+            [item["source"], item["relationship"], item["destination"]] for item in search_output
+        ]
+        bm25 = BM25Okapi(search_outputs_sequence)
+
+        tokenized_query = query.split(" ")
+        reranked_results = bm25.get_top_n(tokenized_query, search_outputs_sequence, n=5)
+
+        search_results = []
+        for item in reranked_results:
+            search_results.append({"source": item[0], "relationship": item[1], "destination": item[2]})
+
+        logger.info(f"Returned {len(search_results)} search results")
+        return search_results
+
+    def delete(self, data, filters):
+        """
+        Delete graph entities associated with the given memory text.
+
+        Extracts entities and relationships from the memory text using the same
+        pipeline as add(), then deletes the matching relationships in the graph.
+
+        Args:
+            data (str): The memory text whose graph entities should be removed.
+            filters (dict): Scope filters (user_id, agent_id, run_id).
+        """
+        try:
+            entity_type_map = self._retrieve_nodes_from_data(data, filters)
+            if not entity_type_map:
+                logger.debug("No entities found in memory text, skipping graph cleanup")
+                return
+            to_be_deleted = self._establish_nodes_relations_from_data(data, filters, entity_type_map)
+            if to_be_deleted:
+                self._delete_entities(to_be_deleted, filters)
+        except Exception as e:
+            logger.error(f"Error during graph cleanup for memory delete: {e}")
+
+    def delete_all(self, filters):
+        """Delete all nodes and relationships for a user or specific agent."""
+        where_parts = ["n.user_id = %s"]
+        params = [filters["user_id"]]
+        if filters.get("agent_id"):
+            where_parts.append("n.agent_id = %s")
+            params.append(filters["agent_id"])
+        if filters.get("run_id"):
+            where_parts.append("n.run_id = %s")
+            params.append(filters["run_id"])
+        where_clause = " AND ".join(where_parts)
+
+        self._exec_cypher(
+            f"MATCH (n) WHERE {where_clause} DETACH DELETE n",
+            params=tuple(params),
+        )
+        self.ag.commit()
+
+    def get_all(self, filters, limit=100):
+        """
+        Retrieves all nodes and relationships from the graph database based on optional filtering criteria.
+
+        Args:
+            filters (dict): A dictionary containing filters to be applied during the retrieval.
+            limit (int): The maximum number of nodes and relationships to retrieve. Defaults to 100.
+        Returns:
+            list: A list of dictionaries, each containing:
+                - 'source': The source node name.
+                - 'relationship': The relationship type.
+                - 'target': The target node name.
+        """
+        where_parts = ["n.user_id = %s", "m.user_id = %s"]
+        params = [filters["user_id"], filters["user_id"]]
+        if filters.get("agent_id"):
+            where_parts.extend(["n.agent_id = %s", "m.agent_id = %s"])
+            params.extend([filters["agent_id"], filters["agent_id"]])
+        if filters.get("run_id"):
+            where_parts.extend(["n.run_id = %s", "m.run_id = %s"])
+            params.extend([filters["run_id"], filters["run_id"]])
+        where_clause = " AND ".join(where_parts)
+        params.append(limit)
+
+        results = self._exec_cypher(
+            f"MATCH (n)-[r]->(m) WHERE {where_clause} "
+            f"RETURN n.name, type(r), m.name LIMIT %s",
+            cols=["source", "relationship", "target"],
+            params=tuple(params),
+        )
+
+        final_results = []
+        for result in results:
+            final_results.append(
+                {
+                    "source": result["source"],
+                    "relationship": result["relationship"],
+                    "target": result["target"],
+                }
+            )
+
+        logger.info(f"Retrieved {len(final_results)} relationships")
+        return final_results
+
+    # -- LLM-driven extraction -------------------------------------------------
+
+    def _retrieve_nodes_from_data(self, data, filters):
+        """Extracts all the entities mentioned in the query."""
+        _tools = [EXTRACT_ENTITIES_TOOL]
+        if self.llm_provider in ["azure_openai_structured", "openai_structured"]:
+            _tools = [EXTRACT_ENTITIES_STRUCT_TOOL]
+        search_results = self.llm.generate_response(
+            messages=[
+                {
+                    "role": "system",
+                    "content": f"You are a smart assistant who understands entities and their types in a given text. If user message contains self reference such as 'I', 'me', 'my' etc. then use {filters['user_id']} as the source entity. Extract all the entities from the text. ***DO NOT*** answer the question itself if the given text is a question.",
+                },
+                {"role": "user", "content": data},
+            ],
+            tools=_tools,
+        )
+
+        entity_type_map = {}
+
+        try:
+            for tool_call in search_results["tool_calls"]:
+                if tool_call["name"] != "extract_entities":
+                    continue
+                for item in tool_call.get("arguments", {}).get("entities", []):
+                    if "entity" in item and "entity_type" in item:
+                        entity_type_map[item["entity"]] = item["entity_type"]
+        except Exception as e:
+            logger.exception(
+                f"Error in search tool: {e}, llm_provider={self.llm_provider}, search_results={search_results}"
+            )
+
+        entity_type_map = {k.lower().replace(" ", "_"): v.lower().replace(" ", "_") for k, v in entity_type_map.items()}
+        logger.debug(f"Entity type map: {entity_type_map}\n search_results={search_results}")
+        return entity_type_map
+
+    def _establish_nodes_relations_from_data(self, data, filters, entity_type_map):
+        """Establish relations among the extracted nodes."""
+
+        user_identity = f"user_id: {filters['user_id']}"
+        if filters.get("agent_id"):
+            user_identity += f", agent_id: {filters['agent_id']}"
+        if filters.get("run_id"):
+            user_identity += f", run_id: {filters['run_id']}"
+
+        if self.config.graph_store.custom_prompt:
+            system_content = EXTRACT_RELATIONS_PROMPT.replace("USER_ID", user_identity)
+            system_content = system_content.replace("CUSTOM_PROMPT", f"4. {self.config.graph_store.custom_prompt}")
+            messages = [
+                {"role": "system", "content": system_content},
+                {"role": "user", "content": data},
+            ]
+        else:
+            system_content = EXTRACT_RELATIONS_PROMPT.replace("USER_ID", user_identity)
+            messages = [
+                {"role": "system", "content": system_content},
+                {"role": "user", "content": f"List of entities: {list(entity_type_map.keys())}. \n\nText: {data}"},
+            ]
+
+        _tools = [RELATIONS_TOOL]
+        if self.llm_provider in ["azure_openai_structured", "openai_structured"]:
+            _tools = [RELATIONS_STRUCT_TOOL]
+
+        extracted_entities = self.llm.generate_response(
+            messages=messages,
+            tools=_tools,
+        )
+
+        entities = []
+        if extracted_entities and extracted_entities.get("tool_calls"):
+            entities = extracted_entities["tool_calls"][0].get("arguments", {}).get("entities", [])
+
+        entities = self._remove_spaces_from_entities(entities)
+        logger.debug(f"Extracted entities: {entities}")
+        return entities
+
+    # -- graph DB operations ---------------------------------------------------
+
+    def _search_graph_db(self, node_list, filters, limit=100):
+        """Search similar nodes and their respective incoming and outgoing relations."""
+        result_relations = []
+
+        for node in node_list:
+            n_embedding = self.embedding_model.embed(node)
+
+            nodes = self._fetch_user_nodes_with_embeddings(filters["user_id"])
+            similar_nodes = _get_similar_nodes(nodes, n_embedding, filters, self.threshold)
+
+            # Build WHERE clause for relationship target filtering
+            rel_where_parts = ["m.user_id = %s"]
+            rel_params_suffix = [filters["user_id"]]
+            if filters.get("agent_id"):
+                rel_where_parts.append("m.agent_id = %s")
+                rel_params_suffix.append(filters["agent_id"])
+            if filters.get("run_id"):
+                rel_where_parts.append("m.run_id = %s")
+                rel_params_suffix.append(filters["run_id"])
+            rel_where = " AND ".join(rel_where_parts)
+
+            # For each similar node, fetch its relationships
+            for sn in similar_nodes[:limit]:
+                node_name = sn["name"]
+                similarity = sn["similarity"]
+
+                out_params = (filters["user_id"], node_name) + tuple(rel_params_suffix)
+                out_results = self._exec_cypher(
+                    f"MATCH (n {{user_id: %s, name: %s}})-[r]->(m) "
+                    f"WHERE {rel_where} "
+                    f"RETURN n.name, type(r), m.name",
+                    cols=["source", "relationship", "destination"],
+                    params=out_params,
+                )
+
+                in_params = (filters["user_id"], node_name) + tuple(rel_params_suffix)
+                in_results = self._exec_cypher(
+                    f"MATCH (n {{user_id: %s, name: %s}})<-[r]-(m) "
+                    f"WHERE {rel_where} "
+                    f"RETURN m.name, type(r), n.name",
+                    cols=["source", "relationship", "destination"],
+                    params=in_params,
+                )
+
+                for rel in out_results + in_results:
+                    rel["similarity"] = similarity
+                    result_relations.append(rel)
+
+        return result_relations
+
+    def _get_delete_entities_from_search_output(self, search_output, data, filters):
+        """Get the entities to be deleted from the search output."""
+        search_output_string = format_entities(search_output)
+
+        user_identity = f"user_id: {filters['user_id']}"
+        if filters.get("agent_id"):
+            user_identity += f", agent_id: {filters['agent_id']}"
+        if filters.get("run_id"):
+            user_identity += f", run_id: {filters['run_id']}"
+
+        system_prompt, user_prompt = get_delete_messages(search_output_string, data, user_identity)
+
+        _tools = [DELETE_MEMORY_TOOL_GRAPH]
+        if self.llm_provider in ["azure_openai_structured", "openai_structured"]:
+            _tools = [DELETE_MEMORY_STRUCT_TOOL_GRAPH]
+
+        memory_updates = self.llm.generate_response(
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt},
+            ],
+            tools=_tools,
+        )
+
+        to_be_deleted = []
+        for item in memory_updates.get("tool_calls", []):
+            if item.get("name") == "delete_graph_memory":
+                to_be_deleted.append(item.get("arguments"))
+        to_be_deleted = self._remove_spaces_from_entities(to_be_deleted)
+        logger.debug(f"Deleted relationships: {to_be_deleted}")
+        return to_be_deleted
+
+    def _delete_entities(self, to_be_deleted, filters):
+        """Delete the entities from the graph."""
+        user_id = filters["user_id"]
+        agent_id = filters.get("agent_id")
+        run_id = filters.get("run_id")
+        results = []
+
+        try:
+            for item in to_be_deleted:
+                source = item["source"]
+                destination = item["destination"]
+                relationship = item["relationship"]
+
+                where_parts = [
+                    "n.user_id = %s", "n.name = %s",
+                    "m.user_id = %s", "m.name = %s",
+                ]
+                params = [user_id, source, user_id, destination]
+                if agent_id:
+                    where_parts.extend(["n.agent_id = %s", "m.agent_id = %s"])
+                    params.extend([agent_id, agent_id])
+                if run_id:
+                    where_parts.extend(["n.run_id = %s", "m.run_id = %s"])
+                    params.extend([run_id, run_id])
+                where_clause = " AND ".join(where_parts)
+
+                result = self._exec_cypher(
+                    f"MATCH (n)-[r:{relationship}]->(m) "
+                    f"WHERE {where_clause} "
+                    f"DELETE r "
+                    f"RETURN n.name, type(r), m.name",
+                    cols=["source", "relationship", "target"],
+                    params=tuple(params),
+                )
+                results.append(result)
+
+            self.ag.commit()
+        except Exception:
+            self.ag.rollback()
+            raise
+
+        return results
+
+    def _add_entities(self, to_be_added, filters, entity_type_map):
+        """Add new entities to the graph. Merge nodes if they already exist.
+
+        Apache AGE does not support ``ON CREATE SET`` / ``ON MATCH SET``, so we
+        use ``MERGE … SET`` which always applies.  The ``coalesce`` pattern
+        ensures ``created`` is only set on the first merge.
+        """
+        user_id = filters["user_id"]
+        agent_id = filters.get("agent_id")
+        run_id = filters.get("run_id")
+        results = []
+
+        try:
+            for item in to_be_added:
+                source = item["source"]
+                destination = item["destination"]
+                relationship = item["relationship"]
+
+                source_embedding = self.embedding_model.embed(source)
+                dest_embedding = self.embedding_model.embed(destination)
+
+                source_match = self._find_similar_node(source_embedding, filters, threshold=self.threshold)
+                dest_match = self._find_similar_node(dest_embedding, filters, threshold=self.threshold)
+
+                effective_source = source_match["name"] if source_match else source
+                effective_dest = dest_match["name"] if dest_match else destination
+
+                # Merge source and destination nodes
+                self._merge_node(user_id, effective_source, source_embedding, agent_id, run_id)
+                self._merge_node(user_id, effective_dest, dest_embedding, agent_id, run_id)
+
+                # Merge relationship
+                result = self._exec_cypher(
+                    f"MATCH (s {{user_id: %s, name: %s}}), (d {{user_id: %s, name: %s}}) "
+                    f"MERGE (s)-[r:{relationship}]->(d) "
+                    f"RETURN s.name, type(r), d.name",
+                    cols=["source", "relationship", "target"],
+                    params=(user_id, effective_source, user_id, effective_dest),
+                )
+                results.append(result)
+
+            self.ag.commit()
+        except Exception:
+            self.ag.rollback()
+            raise
+
+        return results
+
+    def _remove_spaces_from_entities(self, entity_list):
+        for item in entity_list:
+            item["source"] = item["source"].lower().replace(" ", "_")
+            item["relationship"] = sanitize_relationship_for_cypher(item["relationship"].lower().replace(" ", "_"))
+            item["destination"] = item["destination"].lower().replace(" ", "_")
+        return entity_list
+
+    def reset(self):
+        """Reset the graph by clearing all nodes and relationships."""
+        logger.warning("Clearing graph...")
+        self._exec_cypher("MATCH (n) DETACH DELETE n")
+        self.ag.commit()

mem0/memory/graph_memory.py

@@ -129,6 +129,28 @@ class MemoryGraph:
 
         return search_results
 
+    def delete(self, data, filters):
+        """
+        Delete graph entities associated with the given memory text.
+
+        Extracts entities and relationships from the memory text using the same
+        pipeline as add(), then soft-deletes the matching relationships in the graph.
+
+        Args:
+            data (str): The memory text whose graph entities should be removed.
+            filters (dict): Scope filters (user_id, agent_id, run_id).
+        """
+        try:
+            entity_type_map = self._retrieve_nodes_from_data(data, filters)
+            if not entity_type_map:
+                logger.debug("No entities found in memory text, skipping graph cleanup")
+                return
+            to_be_deleted = self._establish_nodes_relations_from_data(data, filters, entity_type_map)
+            if to_be_deleted:
+                self._delete_entities(to_be_deleted, filters)
+        except Exception as e:
+            logger.error(f"Error during graph cleanup for memory delete: {e}")
+
     def delete_all(self, filters):
         # Build node properties for filtering
         node_props = ["user_id: $user_id"]
@@ -174,6 +196,7 @@ class MemoryGraph:
 
         query = f"""
         MATCH (n {self.node_label} {{{node_props_str}}})-[r]->(m {self.node_label} {{{node_props_str}}})
+        WHERE r.valid IS NULL OR r.valid = true
         RETURN n.name AS source, type(r) AS relationship, m.name AS target
         LIMIT $limit
         """
@@ -291,10 +314,12 @@ class MemoryGraph:
             CALL {{
                 WITH n
                 MATCH (n)-[r]->(m {self.node_label} {{{node_props_str}}})
+                WHERE r.valid IS NULL OR r.valid = true
                 RETURN n.name AS source, elementId(n) AS source_id, type(r) AS relationship, elementId(r) AS relation_id, m.name AS destination, elementId(m) AS destination_id
                 UNION
                 WITH n  
                 MATCH (n)<-[r]-(m {self.node_label} {{{node_props_str}}})
+                WHERE r.valid IS NULL OR r.valid = true
                 RETURN m.name AS source, elementId(m) AS source_id, type(r) AS relationship, elementId(r) AS relation_id, n.name AS destination, elementId(n) AS destination_id
             }}
             WITH distinct source, source_id, relationship, relation_id, destination, destination_id, similarity
@@ -392,13 +417,15 @@ class MemoryGraph:
             source_props_str = ", ".join(source_props)
             dest_props_str = ", ".join(dest_props)
 
-            # Delete the specific relationship between nodes
+            # Soft-delete: mark relationship as invalid instead of removing it,
+            # enabling temporal reasoning over historical graph state.
+            # See: https://github.com/mem0ai/mem0/issues/4187
             cypher = f"""
             MATCH (n {self.node_label} {{{source_props_str}}})
             -[r:{relationship}]->
             (m {self.node_label} {{{dest_props_str}}})
-            
-            DELETE r
+            WHERE r.valid IS NULL OR r.valid = true
+            SET r.valid = false, r.invalidated_at = datetime()
             RETURN 
                 n.name AS source,
                 m.name AS target,
@@ -464,11 +491,16 @@ class MemoryGraph:
                 CALL db.create.setNodeVectorProperty(destination, 'embedding', $destination_embedding)
                 WITH source, destination
                 MERGE (source)-[r:{relationship}]->(destination)
-                ON CREATE SET 
-                    r.created = timestamp(),
-                    r.mentions = 1
+                ON CREATE SET
+                    r.created_at = timestamp(),
+                    r.updated_at = timestamp(),
+                    r.mentions = 1,
+                    r.valid = true
                 ON MATCH SET
-                    r.mentions = coalesce(r.mentions, 0) + 1
+                    r.mentions = coalesce(r.mentions, 0) + 1,
+                    r.valid = true,
+                    r.updated_at = timestamp(),
+                    r.invalidated_at = null
                 RETURN source.name AS source, type(r) AS relationship, destination.name AS target
                 """
 
@@ -508,11 +540,16 @@ class MemoryGraph:
                 CALL db.create.setNodeVectorProperty(source, 'embedding', $source_embedding)
                 WITH source, destination
                 MERGE (source)-[r:{relationship}]->(destination)
-                ON CREATE SET 
-                    r.created = timestamp(),
-                    r.mentions = 1
+                ON CREATE SET
+                    r.created_at = timestamp(),
+                    r.updated_at = timestamp(),
+                    r.mentions = 1,
+                    r.valid = true
                 ON MATCH SET
-                    r.mentions = coalesce(r.mentions, 0) + 1
+                    r.mentions = coalesce(r.mentions, 0) + 1,
+                    r.valid = true,
+                    r.updated_at = timestamp(),
+                    r.invalidated_at = null
                 RETURN source.name AS source, type(r) AS relationship, destination.name AS target
                 """
 
@@ -537,11 +574,16 @@ class MemoryGraph:
                 WHERE elementId(destination) = $destination_id
                 SET destination.mentions = coalesce(destination.mentions, 0) + 1
                 MERGE (source)-[r:{relationship}]->(destination)
-                ON CREATE SET 
+                ON CREATE SET
                     r.created_at = timestamp(),
                     r.updated_at = timestamp(),
-                    r.mentions = 1
-                ON MATCH SET r.mentions = coalesce(r.mentions, 0) + 1
+                    r.mentions = 1,
+                    r.valid = true
+                ON MATCH SET
+                    r.mentions = coalesce(r.mentions, 0) + 1,
+                    r.valid = true,
+                    r.updated_at = timestamp(),
+                    r.invalidated_at = null
                 RETURN source.name AS source, type(r) AS relationship, destination.name AS target
                 """
 
@@ -585,10 +627,18 @@ class MemoryGraph:
                 WITH source, destination
                 CALL db.create.setNodeVectorProperty(destination, 'embedding', $dest_embedding)
                 WITH source, destination
-                MERGE (source)-[rel:{relationship}]->(destination)
-                ON CREATE SET rel.created = timestamp(), rel.mentions = 1
-                ON MATCH SET rel.mentions = coalesce(rel.mentions, 0) + 1
-                RETURN source.name AS source, type(rel) AS relationship, destination.name AS target
+                MERGE (source)-[r:{relationship}]->(destination)
+                ON CREATE SET
+                    r.created_at = timestamp(),
+                    r.updated_at = timestamp(),
+                    r.mentions = 1,
+                    r.valid = true
+                ON MATCH SET
+                    r.mentions = coalesce(r.mentions, 0) + 1,
+                    r.valid = true,
+                    r.updated_at = timestamp(),
+                    r.invalidated_at = null
+                RETURN source.name AS source, type(r) AS relationship, destination.name AS target
                 """
 
                 params = {

mem0/memory/kuzu_memory.py

@@ -149,6 +149,28 @@ class MemoryGraph:
 
         return search_results
 
+    def delete(self, data, filters):
+        """
+        Delete graph entities associated with the given memory text.
+
+        Extracts entities and relationships from the memory text using the same
+        pipeline as add(), then deletes the matching relationships in the graph.
+
+        Args:
+            data (str): The memory text whose graph entities should be removed.
+            filters (dict): Scope filters (user_id, agent_id, run_id).
+        """
+        try:
+            entity_type_map = self._retrieve_nodes_from_data(data, filters)
+            if not entity_type_map:
+                logger.debug("No entities found in memory text, skipping graph cleanup")
+                return
+            to_be_deleted = self._establish_nodes_relations_from_data(data, filters, entity_type_map)
+            if to_be_deleted:
+                self._delete_entities(to_be_deleted, filters)
+        except Exception as e:
+            logger.error(f"Error during graph cleanup for memory delete: {e}")
+
     def delete_all(self, filters):
         # Build node properties for filtering
         node_props = ["user_id: $user_id"]

mem0/memory/main.py

@@ -644,6 +644,9 @@ class Memory(MemoryBase):
                         if memory_id and (metadata.get("agent_id") or metadata.get("run_id")):
                             # Update only the session identifiers, keep content the same
                             existing_memory = self.vector_store.get(vector_id=memory_id)
+                            if existing_memory is None:
+                                logger.warning(f"Memory {memory_id} not found for session ID update, skipping")
+                                continue
                             updated_metadata = deepcopy(existing_memory.payload)
                             if metadata.get("agent_id"):
                                 updated_metadata["agent_id"] = metadata["agent_id"]
@@ -1098,7 +1101,27 @@ class Memory(MemoryBase):
             memory_id (str): ID of the memory to delete.
         """
         capture_event("mem0.delete", self, {"memory_id": memory_id, "sync_type": "sync"})
-        self._delete_memory(memory_id)
+
+        existing_memory = self.vector_store.get(vector_id=memory_id)
+        if existing_memory is None:
+            raise ValueError(f"Memory with id {memory_id} not found")
+
+        # Clean up graph entities before deleting from vector store
+        if self.enable_graph:
+            try:
+                memory_text = existing_memory.payload.get("data", "")
+                if memory_text:
+                    filters = {}
+                    for key in ("user_id", "agent_id", "run_id"):
+                        val = existing_memory.payload.get(key)
+                        if val:
+                            filters[key] = val
+                    if filters.get("user_id"):
+                        self.graph.delete(memory_text, filters)
+            except Exception as e:
+                logger.error(f"Error cleaning up graph for memory {memory_id}: {e}")
+
+        self._delete_memory(memory_id, existing_memory)
         return {"message": "Memory deleted successfully!"}
 
     def delete_all(self, user_id: Optional[str] = None, agent_id: Optional[str] = None, run_id: Optional[str] = None):
@@ -1157,24 +1180,24 @@ class Memory(MemoryBase):
         else:
             embeddings = self.embedding_model.embed(data, memory_action="add")
         memory_id = str(uuid.uuid4())
-        metadata = metadata or {}
-        metadata["data"] = data
-        metadata["hash"] = hashlib.md5(data.encode()).hexdigest()
-        metadata["created_at"] = datetime.now(timezone.utc).isoformat()
+        new_metadata = deepcopy(metadata) if metadata is not None else {}
+        new_metadata["data"] = data
+        new_metadata["hash"] = hashlib.md5(data.encode()).hexdigest()
+        new_metadata["created_at"] = datetime.now(timezone.utc).isoformat()
 
         self.vector_store.insert(
             vectors=[embeddings],
             ids=[memory_id],
-            payloads=[metadata],
+            payloads=[new_metadata],
         )
         self.db.add_history(
             memory_id,
             None,
             data,
             "ADD",
-            created_at=metadata.get("created_at"),
-            actor_id=metadata.get("actor_id"),
-            role=metadata.get("role"),
+            created_at=new_metadata.get("created_at"),
+            actor_id=new_metadata.get("actor_id"),
+            role=new_metadata.get("role"),
         )
         return memory_id
 
@@ -1208,9 +1231,10 @@ class Memory(MemoryBase):
         if metadata is None:
             raise ValueError("Metadata cannot be done for procedural memory.")
 
-        metadata["memory_type"] = MemoryType.PROCEDURAL.value
+        new_metadata = deepcopy(metadata)
+        new_metadata["memory_type"] = MemoryType.PROCEDURAL.value
         embeddings = self.embedding_model.embed(procedural_memory, memory_action="add")
-        memory_id = self._create_memory(procedural_memory, {procedural_memory: embeddings}, metadata=metadata)
+        memory_id = self._create_memory(procedural_memory, {procedural_memory: embeddings}, metadata=new_metadata)
         capture_event("mem0._create_procedural_memory", self, {"memory_id": memory_id, "sync_type": "sync"})
 
         result = {"results": [{"id": memory_id, "memory": procedural_memory, "event": "ADD"}]}
@@ -1226,6 +1250,9 @@ class Memory(MemoryBase):
             logger.error(f"Error getting memory with ID {memory_id} during update.")
             raise ValueError(f"Error getting memory with ID {memory_id}. Please provide a valid 'memory_id'")
 
+        if existing_memory is None:
+            raise ValueError(f"Memory with id {memory_id} not found. Please provide a valid 'memory_id'")
+
         prev_value = existing_memory.payload.get("data")
 
         new_metadata = deepcopy(metadata) if metadata is not None else {}
@@ -1271,9 +1298,12 @@ class Memory(MemoryBase):
         )
         return memory_id
 
-    def _delete_memory(self, memory_id):
+    def _delete_memory(self, memory_id, existing_memory=None):
         logger.info(f"Deleting memory with {memory_id=}")
-        existing_memory = self.vector_store.get(vector_id=memory_id)
+        if existing_memory is None:
+            existing_memory = self.vector_store.get(vector_id=memory_id)
+            if existing_memory is None:
+                raise ValueError(f"Memory with id {memory_id} not found")
         prev_value = existing_memory.payload.get("data", "")
         self.vector_store.delete(vector_id=memory_id)
         self.db.add_history(
@@ -1671,6 +1701,9 @@ class AsyncMemory(MemoryBase):
                             # Create async task to update only the session identifiers
                             async def update_session_ids(mem_id, meta):
                                 existing_memory = await asyncio.to_thread(self.vector_store.get, vector_id=mem_id)
+                                if existing_memory is None:
+                                    logger.warning(f"Memory {mem_id} not found for session ID update, skipping")
+                                    return
                                 updated_metadata = deepcopy(existing_memory.payload)
                                 if meta.get("agent_id"):
                                     updated_metadata["agent_id"] = meta["agent_id"]
@@ -2163,7 +2196,27 @@ class AsyncMemory(MemoryBase):
             memory_id (str): ID of the memory to delete.
         """
         capture_event("mem0.delete", self, {"memory_id": memory_id, "sync_type": "async"})
-        await self._delete_memory(memory_id)
+
+        existing_memory = await asyncio.to_thread(self.vector_store.get, vector_id=memory_id)
+        if existing_memory is None:
+            raise ValueError(f"Memory with id {memory_id} not found")
+
+        # Clean up graph entities before deleting from vector store
+        if self.enable_graph:
+            try:
+                memory_text = existing_memory.payload.get("data", "")
+                if memory_text:
+                    filters = {}
+                    for key in ("user_id", "agent_id", "run_id"):
+                        val = existing_memory.payload.get(key)
+                        if val:
+                            filters[key] = val
+                    if filters.get("user_id"):
+                        await asyncio.to_thread(self.graph.delete, memory_text, filters)
+            except Exception as e:
+                logger.error(f"Error cleaning up graph for memory {memory_id}: {e}")
+
+        await self._delete_memory(memory_id, existing_memory)
         return {"message": "Memory deleted successfully!"}
 
     async def delete_all(self, user_id=None, agent_id=None, run_id=None):
@@ -2226,16 +2279,16 @@ class AsyncMemory(MemoryBase):
             embeddings = await asyncio.to_thread(self.embedding_model.embed, data, memory_action="add")
 
         memory_id = str(uuid.uuid4())
-        metadata = metadata or {}
-        metadata["data"] = data
-        metadata["hash"] = hashlib.md5(data.encode()).hexdigest()
-        metadata["created_at"] = datetime.now(timezone.utc).isoformat()
+        new_metadata = deepcopy(metadata) if metadata is not None else {}
+        new_metadata["data"] = data
+        new_metadata["hash"] = hashlib.md5(data.encode()).hexdigest()
+        new_metadata["created_at"] = datetime.now(timezone.utc).isoformat()
 
         await asyncio.to_thread(
             self.vector_store.insert,
             vectors=[embeddings],
             ids=[memory_id],
-            payloads=[metadata],
+            payloads=[new_metadata],
         )
 
         await asyncio.to_thread(
@@ -2244,9 +2297,9 @@ class AsyncMemory(MemoryBase):
             None,
             data,
             "ADD",
-            created_at=metadata.get("created_at"),
-            actor_id=metadata.get("actor_id"),
-            role=metadata.get("role"),
+            created_at=new_metadata.get("created_at"),
+            actor_id=new_metadata.get("actor_id"),
+            role=new_metadata.get("role"),
         )
 
         return memory_id
@@ -2295,9 +2348,10 @@ class AsyncMemory(MemoryBase):
         if metadata is None:
             raise ValueError("Metadata cannot be done for procedural memory.")
 
-        metadata["memory_type"] = MemoryType.PROCEDURAL.value
+        new_metadata = deepcopy(metadata)
+        new_metadata["memory_type"] = MemoryType.PROCEDURAL.value
         embeddings = await asyncio.to_thread(self.embedding_model.embed, procedural_memory, memory_action="add")
-        memory_id = await self._create_memory(procedural_memory, {procedural_memory: embeddings}, metadata=metadata)
+        memory_id = await self._create_memory(procedural_memory, {procedural_memory: embeddings}, metadata=new_metadata)
         capture_event("mem0._create_procedural_memory", self, {"memory_id": memory_id, "sync_type": "async"})
 
         result = {"results": [{"id": memory_id, "memory": procedural_memory, "event": "ADD"}]}
@@ -2313,6 +2367,9 @@ class AsyncMemory(MemoryBase):
             logger.error(f"Error getting memory with ID {memory_id} during update.")
             raise ValueError(f"Error getting memory with ID {memory_id}. Please provide a valid 'memory_id'")
 
+        if existing_memory is None:
+            raise ValueError(f"Memory with id {memory_id} not found. Please provide a valid 'memory_id'")
+
         prev_value = existing_memory.payload.get("data")
 
         new_metadata = deepcopy(metadata) if metadata is not None else {}
@@ -2361,9 +2418,12 @@ class AsyncMemory(MemoryBase):
         )
         return memory_id
 
-    async def _delete_memory(self, memory_id):
+    async def _delete_memory(self, memory_id, existing_memory=None):
         logger.info(f"Deleting memory with {memory_id=}")
-        existing_memory = await asyncio.to_thread(self.vector_store.get, vector_id=memory_id)
+        if existing_memory is None:
+            existing_memory = await asyncio.to_thread(self.vector_store.get, vector_id=memory_id)
+            if existing_memory is None:
+                raise ValueError(f"Memory with id {memory_id} not found")
         prev_value = existing_memory.payload.get("data", "")
 
         await asyncio.to_thread(self.vector_store.delete, vector_id=memory_id)

mem0/memory/memgraph_memory.py

@@ -134,6 +134,28 @@ class MemoryGraph:
 
         return search_results
 
+    def delete(self, data, filters):
+        """
+        Delete graph entities associated with the given memory text.
+
+        Extracts entities and relationships from the memory text using the same
+        pipeline as add(), then deletes the matching relationships in the graph.
+
+        Args:
+            data (str): The memory text whose graph entities should be removed.
+            filters (dict): Scope filters (user_id, agent_id).
+        """
+        try:
+            entity_type_map = self._retrieve_nodes_from_data(data, filters)
+            if not entity_type_map:
+                logger.debug("No entities found in memory text, skipping graph cleanup")
+                return
+            to_be_deleted = self._establish_nodes_relations_from_data(data, filters, entity_type_map)
+            if to_be_deleted:
+                self._delete_entities(to_be_deleted, filters)
+        except Exception as e:
+            logger.error(f"Error during graph cleanup for memory delete: {e}")
+
     def delete_all(self, filters):
         """Delete all nodes and relationships for a user or specific agent."""
         if filters.get("agent_id"):

mem0/utils/factory.py

@@ -6,6 +6,7 @@ from mem0.configs.llms.anthropic import AnthropicConfig
 from mem0.configs.llms.azure import AzureOpenAIConfig
 from mem0.configs.llms.base import BaseLlmConfig
 from mem0.configs.llms.deepseek import DeepSeekConfig
+from mem0.configs.llms.minimax import MinimaxConfig
 from mem0.configs.llms.lmstudio import LMStudioConfig
 from mem0.configs.llms.ollama import OllamaConfig
 from mem0.configs.llms.openai import OpenAIConfig
@@ -45,6 +46,7 @@ class LlmFactory:
         "azure_openai_structured": ("mem0.llms.azure_openai_structured.AzureOpenAIStructuredLLM", AzureOpenAIConfig),
         "gemini": ("mem0.llms.gemini.GeminiLLM", BaseLlmConfig),
         "deepseek": ("mem0.llms.deepseek.DeepSeekLLM", DeepSeekConfig),
+        "minimax": ("mem0.llms.minimax.MiniMaxLLM", MinimaxConfig),
         "xai": ("mem0.llms.xai.XAILLM", BaseLlmConfig),
         "sarvam": ("mem0.llms.sarvam.SarvamLLM", BaseLlmConfig),
         "lmstudio": ("mem0.llms.lmstudio.LMStudioLLM", LMStudioConfig),
@@ -186,6 +188,7 @@ class VectorStoreFactory:
         "baidu": "mem0.vector_stores.baidu.BaiduDB",
         "cassandra": "mem0.vector_stores.cassandra.CassandraDB",
         "neptune": "mem0.vector_stores.neptune_analytics.NeptuneAnalyticsVector",
+        "turbopuffer": "mem0.vector_stores.turbopuffer.TurbopufferDB",
     }
 
     @classmethod
@@ -216,6 +219,7 @@ class GraphStoreFactory:
         "neptune": "mem0.graphs.neptune.neptunegraph.MemoryGraph",
         "neptunedb": "mem0.graphs.neptune.neptunedb.MemoryGraph",
         "kuzu": "mem0.memory.kuzu_memory.MemoryGraph",
+        "apache_age": "mem0.memory.apache_age_memory.MemoryGraph",
         "default": "mem0.memory.graph_memory.MemoryGraph",
     }
 

mem0/vector_stores/configs.py

@@ -34,6 +34,7 @@ class VectorStoreConfig(BaseModel):
         "faiss": "FAISSConfig",
         "langchain": "LangchainConfig",
         "s3_vectors": "S3VectorsConfig",
+        "turbopuffer": "TurbopufferConfig",
     }
 
     @model_validator(mode="after")

mem0/vector_stores/databricks.py

@@ -65,7 +65,7 @@ class Databricks(VectorStoreBase):
             catalog (str): Unity Catalog catalog name.
             schema (str): Unity Catalog schema name.
             table_name (str): Source Delta table name.
-            index_name (str, optional): Vector search index name (default: "mem0").
+            collection_name (str, optional): Vector search index name (default: "mem0").
             index_type (str, optional): Index type, either "DELTA_SYNC" or "DIRECT_ACCESS" (default: "DELTA_SYNC").
             embedding_model_endpoint_name (str, optional): Embedding model endpoint for Databricks-computed embeddings.
             embedding_dimension (int, optional): Vector embedding dimensions (default: 1536).
@@ -85,7 +85,7 @@ class Databricks(VectorStoreBase):
         self.fully_qualified_index_name = f"{self.catalog}.{self.schema}.{self.index_name}"
 
         # Configuration
-        self.index_type = index_type
+        self.index_type = VectorIndexType(index_type) if isinstance(index_type, str) else index_type
         self.embedding_model_endpoint_name = embedding_model_endpoint_name
         self.embedding_dimension = embedding_dimension
         self.endpoint_type = endpoint_type
@@ -261,11 +261,11 @@ class Databricks(VectorStoreBase):
             )
             logger.info(f"Successfully created source table '{self.fully_qualified_table_name}'")
             self.client.table_constraints.create(
-                full_name_arg="logistics_dev.ai.dev_memory",
+                full_name_arg=self.fully_qualified_table_name,
                 constraint=TableConstraint(
                     primary_key_constraint=PrimaryKeyConstraint(
-                        name="pk_dev_memory",  # Name of the primary key constraint
-                        child_columns=["memory_id"],  # Columns that make up the primary key
+                        name=f"pk_{self.table_name}",
+                        child_columns=["memory_id"],
                     )
                 ),
             )
@@ -439,29 +439,29 @@ class Databricks(VectorStoreBase):
         try:
             filters_json = json.dumps(filters) if filters else None
 
-            # Choose query type
-            if self.index_type == VectorIndexType.DELTA_SYNC and query:
-                # Text-based search
-                sdk_results = self.client.vector_search_indexes.query_index(
-                    index_name=self.fully_qualified_index_name,
-                    columns=self.column_names,
-                    query_text=query,
-                    num_results=limit,
-                    query_type=self.query_type,
-                    filters_json=filters_json,
-                )
-            elif self.index_type == VectorIndexType.DIRECT_ACCESS and vectors:
-                # Vector-based search
-                sdk_results = self.client.vector_search_indexes.query_index(
-                    index_name=self.fully_qualified_index_name,
-                    columns=self.column_names,
-                    query_vector=vectors,
-                    num_results=limit,
-                    query_type=self.query_type,
-                    filters_json=filters_json,
-                )
+            # Choose query mode per Databricks SDK contract:
+            # - query_text: for Delta Sync Index with model endpoint
+            # - query_vector: for Direct Access Index and Delta Sync Index with self-managed vectors
+            query_kwargs = {
+                "index_name": self.fully_qualified_index_name,
+                "columns": self.column_names,
+                "num_results": limit,
+                "query_type": self.query_type,
+                "filters_json": filters_json,
+            }
+            uses_model_endpoint = (
+                self.index_type == VectorIndexType.DELTA_SYNC and self.embedding_model_endpoint_name
+            )
+            if uses_model_endpoint:
+                if not query:
+                    raise ValueError("Query text is required for Delta Sync Index with model endpoint.")
+                query_kwargs["query_text"] = query
+            elif vectors:
+                query_kwargs["query_vector"] = vectors
             else:
-                raise ValueError("Must provide query text for DELTA_SYNC or vectors for DIRECT_ACCESS.")
+                raise ValueError("Must provide vectors for search.")
+
+            sdk_results = self.client.vector_search_indexes.query_index(**query_kwargs)
 
             # Parse results
             result_data = sdk_results.result if hasattr(sdk_results, "result") else sdk_results
@@ -572,14 +572,23 @@ class Databricks(VectorStoreBase):
             filters = {"memory_id": vector_id}
             filters_json = json.dumps(filters)
 
-            results = self.client.vector_search_indexes.query_index(
-                index_name=self.fully_qualified_index_name,
-                columns=self.column_names,
-                query_text=" ",  # Empty query, rely on filters
-                num_results=1,
-                query_type=self.query_type,
-                filters_json=filters_json,
+            # Use query_text for Delta Sync with model endpoint, query_vector otherwise
+            query_kwargs = {
+                "index_name": self.fully_qualified_index_name,
+                "columns": self.column_names,
+                "num_results": 1,
+                "query_type": self.query_type,
+                "filters_json": filters_json,
+            }
+            uses_model_endpoint = (
+                self.index_type == VectorIndexType.DELTA_SYNC and self.embedding_model_endpoint_name
             )
+            if uses_model_endpoint:
+                query_kwargs["query_text"] = " "
+            else:
+                query_kwargs["query_vector"] = [0.0] * self.embedding_dimension
+
+            results = self.client.vector_search_indexes.query_index(**query_kwargs)
 
             # Process results
             result_data = results.result if hasattr(results, "result") else results
@@ -589,7 +598,7 @@ class Databricks(VectorStoreBase):
                 raise KeyError(f"Vector with ID {vector_id} not found")
 
             result = data_array[0]
-            columns = columns = [col.name for col in results.manifest.columns] if results.manifest and results.manifest.columns else []
+            columns = [col.name for col in results.manifest.columns] if results.manifest and results.manifest.columns else []
             row_data = dict(zip(columns, result))
 
             # Build payload following the standard schema
@@ -686,14 +695,23 @@ class Databricks(VectorStoreBase):
             filters_json = json.dumps(filters) if filters else None
             num_results = limit or 100
             columns = self.column_names
-            sdk_results = self.client.vector_search_indexes.query_index(
-                index_name=self.fully_qualified_index_name,
-                columns=columns,
-                query_text=" ",
-                num_results=num_results,
-                query_type=self.query_type,
-                filters_json=filters_json,
+            # Use query_text for Delta Sync with model endpoint, query_vector otherwise
+            query_kwargs = {
+                "index_name": self.fully_qualified_index_name,
+                "columns": columns,
+                "num_results": num_results,
+                "query_type": self.query_type,
+                "filters_json": filters_json,
+            }
+            uses_model_endpoint = (
+                self.index_type == VectorIndexType.DELTA_SYNC and self.embedding_model_endpoint_name
             )
+            if uses_model_endpoint:
+                query_kwargs["query_text"] = " "
+            else:
+                query_kwargs["query_vector"] = [0.0] * self.embedding_dimension
+
+            sdk_results = self.client.vector_search_indexes.query_index(**query_kwargs)
             result_data = sdk_results.result if hasattr(sdk_results, "result") else sdk_results
             data_array = result_data.data_array if hasattr(result_data, "data_array") else []
 

mem0/vector_stores/mongodb.py

@@ -26,7 +26,7 @@ class OutputData(BaseModel):
 
 
 class MongoDB(VectorStoreBase):
-    VECTOR_TYPE = "knnVector"
+    VECTOR_TYPE = "vector"
     SIMILARITY_METRIC = "cosine"
 
     def __init__(self, db_name: str, collection_name: str, embedding_model_dims: int, mongo_uri: str):
@@ -69,17 +69,16 @@ class MongoDB(VectorStoreBase):
             else:
                 search_index_model = SearchIndexModel(
                     name=self.index_name,
+                    type="vectorSearch",
                     definition={
-                        "mappings": {
-                            "dynamic": False,
-                            "fields": {
-                                "embedding": {
-                                    "type": self.VECTOR_TYPE,
-                                    "dimensions": self.embedding_model_dims,
-                                    "similarity": self.SIMILARITY_METRIC,
-                                }
-                            },
-                        }
+                        "fields": [
+                            {
+                                "type": self.VECTOR_TYPE,
+                                "path": "embedding",
+                                "numDimensions": self.embedding_model_dims,
+                                "similarity": self.SIMILARITY_METRIC,
+                            }
+                        ]
                     },
                 )
                 collection.create_search_index(search_index_model)
@@ -141,7 +140,7 @@ class MongoDB(VectorStoreBase):
                     "$vectorSearch": {
                         "index": self.index_name,
                         "limit": limit,
-                        "numCandidates": limit,
+                        "numCandidates": min(limit * 20, 10000),
                         "queryVector": vectors,
                         "path": "embedding",
                     }
@@ -198,7 +197,8 @@ class MongoDB(VectorStoreBase):
         if vector is not None:
             update_fields["embedding"] = vector
         if payload is not None:
-            update_fields["payload"] = payload
+            for key, value in payload.items():
+                update_fields[f"payload.{key}"] = value
 
         if update_fields:
             try:

mem0/vector_stores/opensearch.py

@@ -113,6 +113,25 @@ class OpenSearchDB(VectorStoreBase):
         if payloads is None:
             payloads = [{} for _ in range(len(vectors))]
 
+        for idx, vec in enumerate(vectors):
+            if vec is None:
+                raise ValueError(
+                    f"Vector at index {idx} is null. "
+                    f"This usually means the embedding model failed to generate an embedding. "
+                    f"Check that your embedding model is configured correctly and returning valid vectors."
+                )
+            if len(vec) == 0:
+                raise ValueError(
+                    f"Vector at index {idx} is empty. "
+                    f"Expected a vector of dimension {self.embedding_model_dims}, got an empty vector."
+                )
+            if len(vec) != self.embedding_model_dims:
+                raise ValueError(
+                    f"Vector at index {idx} has dimension {len(vec)}, "
+                    f"but the index '{self.collection_name}' expects dimension {self.embedding_model_dims}. "
+                    f"Ensure your embedding model's output dimensions match the vector store configuration."
+                )
+
         results = []
         for i, (vec, id_) in enumerate(zip(vectors, ids)):
             body = {
@@ -124,14 +143,16 @@ class OpenSearchDB(VectorStoreBase):
                 self.client.index(index=self.collection_name, body=body)
                 # Force refresh to make documents immediately searchable for tests
                 self.client.indices.refresh(index=self.collection_name)
-                
-                results.append(OutputData(
-                    id=id_,
-                    score=1.0,  # No score for inserts
-                    payload=payloads[i]
-                ))
+
+                results.append(
+                    OutputData(
+                        id=id_,
+                        score=1.0,  # No score for inserts
+                        payload=payloads[i],
+                    )
+                )
             except Exception as e:
-                logger.error(f"Error inserting vector {id_}: {e}")
+                logger.error(f"Error inserting vector {id_}: {e}", exc_info=True)
                 raise
 
         return results
@@ -179,7 +200,7 @@ class OpenSearchDB(VectorStoreBase):
             ]
             return results
         except Exception as e:
-            logger.error(f"Error during search: {e}")
+            logger.error(f"Error during search: {e}", exc_info=True)
             return []
 
     def delete(self, vector_id: str) -> None:
@@ -200,6 +221,15 @@ class OpenSearchDB(VectorStoreBase):
 
     def update(self, vector_id: str, vector: Optional[List[float]] = None, payload: Optional[Dict] = None) -> None:
         """Update a vector and its payload using the custom 'id' field."""
+        if vector is not None:
+            if len(vector) == 0:
+                raise ValueError("Cannot update with an empty vector.")
+            if len(vector) != self.embedding_model_dims:
+                raise ValueError(
+                    f"Update vector has dimension {len(vector)}, "
+                    f"but the index '{self.collection_name}' expects dimension {self.embedding_model_dims}. "
+                    f"Ensure your embedding model's output dimensions match the vector store configuration."
+                )
 
         # First, find the document by custom ID
         search_query = {"query": {"term": {"id": vector_id}}}
@@ -222,8 +252,9 @@ class OpenSearchDB(VectorStoreBase):
         if doc:
             try:
                 response = self.client.update(index=self.collection_name, id=opensearch_id, body={"doc": doc})
-            except Exception:
-                pass
+            except Exception as e:
+                logger.error(f"Error updating vector {vector_id}: {e}", exc_info=True)
+                raise
 
     def get(self, vector_id: str) -> Optional[OutputData]:
         """Retrieve a vector by ID."""
@@ -238,7 +269,7 @@ class OpenSearchDB(VectorStoreBase):
 
             return OutputData(id=hits[0]["_source"].get("id"), score=1.0, payload=hits[0]["_source"].get("payload", {}))
         except Exception as e:
-            logger.error(f"Error retrieving vector {vector_id}: {str(e)}")
+            logger.error(f"Error retrieving vector {vector_id}: {str(e)}", exc_info=True)
             return None
 
     def list_cols(self) -> List[str]:
@@ -281,9 +312,8 @@ class OpenSearchDB(VectorStoreBase):
             ]
             return [results]  # VectorStore expects tuple/list format
         except Exception as e:
-            logger.error(f"Error listing vectors: {e}")
+            logger.error(f"Error listing vectors: {e}", exc_info=True)
             return []
-        
 
     def reset(self):
         """Reset the index by deleting and recreating it."""

mem0/vector_stores/qdrant.py

@@ -1,12 +1,14 @@
 import logging
-import os
-import shutil
+from typing import Optional
 
 from qdrant_client import QdrantClient
 from qdrant_client.models import (
     Distance,
     FieldCondition,
     Filter,
+    MatchAny,
+    MatchExcept,
+    MatchText,
     MatchValue,
     PointIdsList,
     PointStruct,
@@ -44,7 +46,8 @@ class Qdrant(VectorStoreBase):
             path (str, optional): Path for local Qdrant database. Defaults to None.
             url (str, optional): Full URL for Qdrant server. Defaults to None.
             api_key (str, optional): API key for Qdrant server. Defaults to None.
-            on_disk (bool, optional): Enables persistent storage. Defaults to False.
+            on_disk (bool, optional): Enables persistent storage. Vectors are stored on disk (True) or in memory (False).
+                Does not delete the local database path. Defaults to False.
         """
         if client:
             self.client = client
@@ -62,9 +65,6 @@ class Qdrant(VectorStoreBase):
             if not params:
                 params["path"] = path
                 self.is_local = True
-                if not on_disk:
-                    if os.path.exists(path) and os.path.isdir(path):
-                        shutil.rmtree(path)
             else:
                 self.is_local = False
 
@@ -138,26 +138,148 @@ class Qdrant(VectorStoreBase):
         ]
         self.client.upsert(collection_name=self.collection_name, points=points)
 
-    def _create_filter(self, filters: dict) -> Filter:
+    def _build_field_condition(self, key: str, value) -> Optional[FieldCondition]:
+        """
+        Build a single FieldCondition from a key-value filter pair.
+
+        Supports the enhanced filter syntax documented at
+        https://docs.mem0.ai/open-source/features/metadata-filtering
+
+        Args:
+            key (str): The payload field name.
+            value: A scalar for simple equality, or a dict with one operator key.
+
+        Returns:
+            Optional[FieldCondition]: The Qdrant field condition, or None if the
+            value is the wildcard '*' (match any / field exists — skip filter).
+        """
+        if not isinstance(value, dict):
+            if value == "*":
+                # Wildcard: match any value. Qdrant has no direct "field exists"
+                # condition via FieldCondition, so we skip this filter (match all).
+                return None
+            if isinstance(value, list):
+                # List shorthand: {"field": ["a", "b"]} treated as in-operator.
+                return FieldCondition(key=key, match=MatchAny(any=value))
+            # Simple equality: {"field": "value"}
+            return FieldCondition(key=key, match=MatchValue(value=value))
+
+        ops = set(value.keys())
+        range_ops = {"gt", "gte", "lt", "lte"}
+        non_range_ops = ops - range_ops
+
+        if ops & range_ops:
+            if non_range_ops:
+                raise ValueError(
+                    f"Cannot mix range operators ({ops & range_ops}) with "
+                    f"non-range operators ({non_range_ops}) for field '{key}'. "
+                    f"Use AND to combine them as separate conditions."
+                )
+            range_kwargs = {op: value[op] for op in range_ops if op in value}
+            return FieldCondition(key=key, range=Range(**range_kwargs))
+        elif "eq" in value:
+            return FieldCondition(key=key, match=MatchValue(value=value["eq"]))
+        elif "ne" in value:
+            return FieldCondition(key=key, match=MatchExcept(**{"except": [value["ne"]]}))
+        elif "in" in value:
+            return FieldCondition(key=key, match=MatchAny(any=value["in"]))
+        elif "nin" in value:
+            return FieldCondition(key=key, match=MatchExcept(**{"except": value["nin"]}))
+        elif "contains" in value or "icontains" in value:
+            # MatchText: with a full-text index, tokenized matching (all words must appear).
+            # Without a full-text index, exact substring match.
+            op = "icontains" if "icontains" in value else "contains"
+            text = value[op]
+            if op == "icontains":
+                logger.debug(
+                    "icontains on field '%s': Qdrant MatchText case sensitivity depends on "
+                    "full-text index configuration. Without a full-text index this behaves "
+                    "as a case-sensitive substring match (same as 'contains').",
+                    key,
+                )
+            return FieldCondition(key=key, match=MatchText(text=text))
+        else:
+            supported = {"eq", "ne", "gt", "gte", "lt", "lte", "in", "nin", "contains", "icontains"}
+            raise ValueError(
+                f"Unsupported filter operator(s) for field '{key}': {ops}. "
+                f"Supported operators: {supported}"
+            )
+
+    def _create_filter(self, filters: dict) -> Optional[Filter]:
         """
         Create a Filter object from the provided filters.
 
+        Supports the enhanced filter syntax with comparison operators (eq, ne,
+        gt, gte, lt, lte), list operators (in, nin), string operators (contains,
+        icontains), and logical operators (AND, OR, NOT).
+
         Args:
             filters (dict): Filters to apply.
 
         Returns:
-            Filter: The created Filter object.
+            Filter: The created Filter object, or None if filters is empty.
         """
         if not filters:
             return None
-            
-        conditions = []
+
+        # Normalize $or/$not/$and → OR/NOT/AND and deduplicate.
+        # Memory._process_metadata_filters() renames OR→$or and NOT→$not,
+        # but effective_filters retains the original OR/NOT keys from
+        # deepcopy(input_filters).  Without dedup the same sub-conditions
+        # would be evaluated twice.
+        key_map = {"$or": "OR", "$not": "NOT", "$and": "AND"}
+        normalized = {}
         for key, value in filters.items():
-            if isinstance(value, dict) and "gte" in value and "lte" in value:
-                conditions.append(FieldCondition(key=key, range=Range(gte=value["gte"], lte=value["lte"])))
+            norm_key = key_map.get(key, key)
+            if norm_key not in normalized:
+                normalized[norm_key] = value
+
+        must = []
+        should = []
+        must_not = []
+
+        for key, value in normalized.items():
+            if key in ("AND", "OR", "NOT"):
+                if not isinstance(value, list):
+                    raise ValueError(
+                        f"{key} filter value must be a list of filter dicts, "
+                        f"got {type(value).__name__}"
+                    )
+                for i, item in enumerate(value):
+                    if not isinstance(item, dict):
+                        raise ValueError(
+                            f"{key} filter list item at index {i} must be a dict, "
+                            f"got {type(item).__name__}: {item!r}"
+                        )
+
+            if key == "AND":
+                for sub in value:
+                    built = self._create_filter(sub)
+                    if built:
+                        must.append(built)
+            elif key == "OR":
+                for sub in value:
+                    built = self._create_filter(sub)
+                    if built:
+                        should.append(built)
+            elif key == "NOT":
+                for sub in value:
+                    built = self._create_filter(sub)
+                    if built:
+                        must_not.append(built)
             else:
-                conditions.append(FieldCondition(key=key, match=MatchValue(value=value)))
-        return Filter(must=conditions) if conditions else None
+                condition = self._build_field_condition(key, value)
+                if condition is not None:
+                    must.append(condition)
+
+        if not any([must, should, must_not]):
+            return None
+
+        return Filter(
+            must=must or None,
+            should=should or None,
+            must_not=must_not or None,
+        )
 
     def search(self, query: str, vectors: list, limit: int = 5, filters: dict = None) -> list:
         """

mem0/vector_stores/turbopuffer.py

@@ -0,0 +1,337 @@
+import logging
+import os
+from typing import Any, Dict, List, Optional, Union
+
+try:
+    from turbopuffer import Turbopuffer as TurbopufferClient
+except ImportError:
+    raise ImportError(
+        "Turbopuffer requires extra dependencies. Install with `pip install turbopuffer`"
+    ) from None
+
+from pydantic import BaseModel
+
+from mem0.vector_stores.base import VectorStoreBase
+
+logger = logging.getLogger(__name__)
+
+
+class OutputData(BaseModel):
+    id: Optional[str]
+    score: Optional[float]
+    payload: Optional[Dict]
+
+
+class TurbopufferDB(VectorStoreBase):
+    def __init__(
+        self,
+        collection_name: str,
+        embedding_model_dims: int,
+        api_key: Optional[str] = None,
+        region: str = "gcp-us-central1",
+        distance_metric: str = "cosine_distance",
+        batch_size: int = 100,
+        extra_params: Optional[Dict[str, Any]] = None,
+    ):
+        """
+        Initialize the Turbopuffer vector store.
+
+        Args:
+            collection_name (str): Name of the namespace/collection.
+            embedding_model_dims (int): Dimensions of the embedding model.
+            api_key (str, optional): API key for Turbopuffer. Defaults to None.
+            region (str, optional): Turbopuffer region. Defaults to "gcp-us-central1".
+            distance_metric (str, optional): Distance metric for vector similarity.
+                Options: "cosine_distance" or "euclidean_squared". Defaults to "cosine_distance".
+            batch_size (int, optional): Batch size for operations. Defaults to 100.
+            extra_params (Dict, optional): Additional parameters for Turbopuffer client. Defaults to None.
+        """
+        api_key = api_key or os.environ.get("TURBOPUFFER_API_KEY")
+        if not api_key:
+            raise ValueError(
+                "Turbopuffer API key must be provided either as a parameter or via TURBOPUFFER_API_KEY environment variable"
+            )
+
+        params = extra_params or {}
+        params["region"] = region
+
+        self.client = TurbopufferClient(api_key=api_key, **params)
+        self.collection_name = collection_name
+        self.embedding_model_dims = embedding_model_dims
+        self.distance_metric = distance_metric
+        self.batch_size = batch_size
+
+        self.namespace = self.client.namespace(self.collection_name)
+
+    def create_col(self, name=None, vector_size=None, distance=None):
+        """
+        Create a new namespace in Turbopuffer.
+        Namespaces are created implicitly on first upsert, so this is a no-op.
+        """
+        pass
+
+    def insert(
+        self,
+        vectors: List[List[float]],
+        payloads: Optional[List[Dict]] = None,
+        ids: Optional[List[Union[str, int]]] = None,
+    ):
+        """
+        Insert vectors into the namespace.
+
+        Args:
+            vectors (list): List of vectors to insert.
+            payloads (list, optional): List of payloads corresponding to vectors. Defaults to None.
+            ids (list, optional): List of IDs corresponding to vectors. Defaults to None.
+        """
+        logger.info(f"Inserting {len(vectors)} vectors into namespace {self.collection_name}")
+
+        if ids is None:
+            ids = [str(i) for i in range(len(vectors))]
+
+        for i in range(0, len(vectors), self.batch_size):
+            batch_end = i + self.batch_size
+            rows = []
+            for j in range(i, min(batch_end, len(vectors))):
+                row = {}
+                if payloads and payloads[j]:
+                    row.update(payloads[j])
+                row["id"] = str(ids[j])
+                row["vector"] = vectors[j]
+                rows.append(row)
+
+            self.namespace.write(
+                upsert_rows=rows,
+                distance_metric=self.distance_metric,
+            )
+
+    def _parse_output(self, rows) -> List[OutputData]:
+        """
+        Parse the output data from Turbopuffer query results.
+
+        Args:
+            rows: List of Row objects from Turbopuffer query.
+
+        Returns:
+            List[OutputData]: Parsed output data.
+        """
+        results = []
+        for row in rows:
+            row_dict = row.model_dump()
+            row_id = str(row_dict.pop("id"))
+            dist = row_dict.pop("$dist", None)
+            row_dict.pop("vector", None)
+
+            score = 1 - dist if dist is not None else None
+
+            results.append(OutputData(
+                id=row_id,
+                score=score,
+                payload=row_dict,
+            ))
+        return results
+
+    def _convert_filters(self, filters: Optional[Dict]):
+        """
+        Convert mem0 filters to Turbopuffer filter format.
+
+        Turbopuffer filters use tuple format: ("And", (("field", "Op", value), ...))
+        """
+        if not filters:
+            return None
+
+        conditions = []
+        for key, value in filters.items():
+            if isinstance(value, dict):
+                if "gte" in value:
+                    conditions.append((key, "Gte", value["gte"]))
+                if "lte" in value:
+                    conditions.append((key, "Lte", value["lte"]))
+            else:
+                conditions.append((key, "Eq", value))
+
+        if not conditions:
+            return None
+        if len(conditions) == 1:
+            return conditions[0]
+        return ("And", tuple(conditions))
+
+    def search(
+        self, query: str, vectors: List[float], limit: int = 5, filters: Optional[Dict] = None
+    ) -> List[OutputData]:
+        """
+        Search for similar vectors.
+
+        Args:
+            query (str): Query text (unused in vector search, kept for interface consistency).
+            vectors (list): Query vector to search with.
+            limit (int, optional): Number of results to return. Defaults to 5.
+            filters (dict, optional): Filters to apply to the search. Defaults to None.
+
+        Returns:
+            list: Search results.
+        """
+        query_params = {
+            "rank_by": ("vector", "ANN", vectors),
+            "top_k": limit,
+            "include_attributes": True,
+        }
+
+        tpuf_filters = self._convert_filters(filters)
+        if tpuf_filters is not None:
+            query_params["filters"] = tpuf_filters
+
+        response = self.namespace.query(**query_params)
+        return self._parse_output(response.rows or [])
+
+    def delete(self, vector_id: Union[str, int]):
+        """
+        Delete a vector by ID.
+
+        Args:
+            vector_id (Union[str, int]): ID of the vector to delete.
+        """
+        self.namespace.write(deletes=[str(vector_id)])
+
+    def update(
+        self,
+        vector_id: Union[str, int],
+        vector: Optional[List[float]] = None,
+        payload: Optional[Dict] = None,
+    ):
+        """
+        Update a vector and its payload.
+
+        Args:
+            vector_id (Union[str, int]): ID of the vector to update.
+            vector (list, optional): Updated vector. Defaults to None.
+            payload (dict, optional): Updated payload. Defaults to None.
+        """
+        if vector is not None:
+            row = {}
+            if payload:
+                row.update(payload)
+            row["id"] = str(vector_id)
+            row["vector"] = vector
+            self.namespace.write(
+                upsert_rows=[row],
+                distance_metric=self.distance_metric,
+            )
+        elif payload is not None:
+            row = dict(payload)
+            row["id"] = str(vector_id)
+            self.namespace.write(patch_rows=[row])
+
+    def get(self, vector_id: Union[str, int]) -> Optional[OutputData]:
+        """
+        Retrieve a vector by ID.
+
+        Args:
+            vector_id (Union[str, int]): ID of the vector to retrieve.
+
+        Returns:
+            OutputData: Retrieved vector data, or None if not found.
+        """
+        try:
+            response = self.namespace.query(
+                top_k=1,
+                rank_by=("vector", "ANN", [0.0] * self.embedding_model_dims),
+                filters=("id", "Eq", str(vector_id)),
+                include_attributes=True,
+            )
+            rows = response.rows or []
+            if rows:
+                return self._parse_output(rows)[0]
+            return None
+        except Exception as e:
+            logger.error(f"Error retrieving vector {vector_id}: {e}")
+            return None
+
+    def list_cols(self) -> list:
+        """
+        List all namespaces.
+
+        Returns:
+            list: List of namespace summaries.
+        """
+        try:
+            result = []
+            for ns in self.client.namespaces():
+                result.append(ns)
+            return result
+        except Exception as e:
+            logger.error(f"Error listing namespaces: {e}")
+            return []
+
+    def delete_col(self):
+        """Delete the entire namespace."""
+        try:
+            self.namespace.delete_all()
+            logger.info(f"Namespace {self.collection_name} deleted successfully")
+        except Exception as e:
+            logger.error(f"Error deleting namespace {self.collection_name}: {e}")
+
+    def col_info(self) -> Dict:
+        """
+        Get information about the namespace.
+
+        Returns:
+            dict: Namespace metadata.
+        """
+        try:
+            metadata = self.namespace.metadata()
+            return {
+                "name": self.collection_name,
+                "approx_row_count": metadata.approx_row_count,
+                "approx_logical_bytes": metadata.approx_logical_bytes,
+                "created_at": str(metadata.created_at),
+                "updated_at": str(metadata.updated_at),
+            }
+        except Exception:
+            return {"name": self.collection_name}
+
+    def list(self, filters: Optional[Dict] = None, limit: int = 100) -> list:
+        """
+        List vectors in the namespace with optional filtering.
+
+        Args:
+            filters (dict, optional): Filters to apply. Defaults to None.
+            limit (int, optional): Number of vectors to return. Defaults to 100.
+
+        Returns:
+            list: Wrapped list of OutputData objects ([[results]]).
+        """
+        query_params = {
+            "rank_by": ("vector", "ANN", [0.0] * self.embedding_model_dims),
+            "top_k": limit,
+            "include_attributes": True,
+        }
+
+        tpuf_filters = self._convert_filters(filters)
+        if tpuf_filters is not None:
+            query_params["filters"] = tpuf_filters
+
+        try:
+            response = self.namespace.query(**query_params)
+            results = self._parse_output(response.rows or [])
+        except Exception as e:
+            logger.error(f"Error listing vectors: {e}")
+            results = []
+        return [results]
+
+    def count(self) -> int:
+        """
+        Get approximate count of vectors in the namespace.
+
+        Returns:
+            int: Approximate number of vectors.
+        """
+        try:
+            metadata = self.namespace.metadata()
+            return metadata.approx_row_count
+        except Exception:
+            return 0
+
+    def reset(self):
+        """Reset the namespace by deleting all vectors."""
+        self.delete_col()

openmemory/api/app/mcp_server.py

@@ -184,7 +184,7 @@ async def search_memory(query: str) -> str:
             for h in hits:
                 # All vector db search functions return OutputData class
                 id, score, payload = h.id, h.score, h.payload
-                if allowed and h.id is None or h.id not in allowed: 
+                if allowed and (h.id is None or h.id not in allowed):
                     continue
                 
                 results.append({

pyproject.toml

@@ -31,6 +31,7 @@ graph = [
     "neo4j>=5.23.1",
     "rank-bm25>=0.2.2",
     "kuzu>=0.11.0",
+    "apache-age-python>=0.0.6",
 ]
 vector_stores = [
     "vecs>=0.4.0",

server/.env.example

@@ -10,3 +10,6 @@ POSTGRES_DB=
 POSTGRES_USER=
 POSTGRES_PASSWORD=
 POSTGRES_COLLECTION_NAME=
+
+# Optional: set to enable API key authentication on all endpoints
+ADMIN_API_KEY=

server/main.py

@@ -1,10 +1,12 @@
 import logging
 import os
+import secrets
 from typing import Any, Dict, List, Optional
 
 from dotenv import load_dotenv
-from fastapi import FastAPI, HTTPException
+from fastapi import Depends, FastAPI, HTTPException
 from fastapi.responses import JSONResponse, RedirectResponse
+from fastapi.security import APIKeyHeader
 from pydantic import BaseModel, Field
 
 from mem0 import Memory
@@ -14,6 +16,22 @@ logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(
 # Load environment variables
 load_dotenv()
 
+ADMIN_API_KEY = os.environ.get("ADMIN_API_KEY", "")
+
+MIN_KEY_LENGTH = 16
+
+if not ADMIN_API_KEY:
+    logging.warning(
+        "ADMIN_API_KEY not set - API endpoints are UNSECURED! "
+        "Set ADMIN_API_KEY environment variable for production use."
+    )
+else:
+    if len(ADMIN_API_KEY) < MIN_KEY_LENGTH:
+        logging.warning(
+            "ADMIN_API_KEY is shorter than %d characters - consider using a longer key for production.",
+            MIN_KEY_LENGTH,
+        )
+    logging.info("API key authentication enabled")
 
 POSTGRES_HOST = os.environ.get("POSTGRES_HOST", "postgres")
 POSTGRES_PORT = os.environ.get("POSTGRES_PORT", "5432")
@@ -60,10 +78,35 @@ MEMORY_INSTANCE = Memory.from_config(DEFAULT_CONFIG)
 
 app = FastAPI(
     title="Mem0 REST APIs",
-    description="A REST API for managing and searching memories for your AI Agents and Apps.",
+    description=(
+        "A REST API for managing and searching memories for your AI Agents and Apps.\n\n"
+        "## Authentication\n"
+        "When the ADMIN_API_KEY environment variable is set, all endpoints require "
+        "the `X-API-Key` header for authentication."
+    ),
     version="1.0.0",
 )
 
+api_key_header = APIKeyHeader(name="X-API-Key", auto_error=False)
+
+
+async def verify_api_key(api_key: Optional[str] = Depends(api_key_header)):
+    """Validate the API key when ADMIN_API_KEY is configured. No-op otherwise."""
+    if ADMIN_API_KEY:
+        if api_key is None:
+            raise HTTPException(
+                status_code=401,
+                detail="X-API-Key header is required.",
+                headers={"WWW-Authenticate": "ApiKey"},
+            )
+        if not secrets.compare_digest(api_key, ADMIN_API_KEY):
+            raise HTTPException(
+                status_code=401,
+                detail="Invalid API key.",
+                headers={"WWW-Authenticate": "ApiKey"},
+            )
+    return api_key
+
 
 class Message(BaseModel):
     role: str = Field(..., description="Role of the message (user or assistant).")
@@ -76,6 +119,9 @@ class MemoryCreate(BaseModel):
     agent_id: Optional[str] = None
     run_id: Optional[str] = None
     metadata: Optional[Dict[str, Any]] = None
+    infer: Optional[bool] = Field(None, description="Whether to extract facts from messages. Defaults to True.")
+    memory_type: Optional[str] = Field(None, description="Type of memory to store (e.g. 'core').")
+    prompt: Optional[str] = Field(None, description="Custom prompt to use for fact extraction.")
 
 
 class SearchRequest(BaseModel):
@@ -84,10 +130,12 @@ class SearchRequest(BaseModel):
     run_id: Optional[str] = None
     agent_id: Optional[str] = None
     filters: Optional[Dict[str, Any]] = None
+    limit: Optional[int] = Field(None, description="Maximum number of results to return.")
+    threshold: Optional[float] = Field(None, description="Minimum similarity score for results.")
 
 
 @app.post("/configure", summary="Configure Mem0")
-def set_config(config: Dict[str, Any]):
+def set_config(config: Dict[str, Any], _api_key: Optional[str] = Depends(verify_api_key)):
     """Set memory configuration."""
     global MEMORY_INSTANCE
     MEMORY_INSTANCE = Memory.from_config(config)
@@ -95,7 +143,7 @@ def set_config(config: Dict[str, Any]):
 
 
 @app.post("/memories", summary="Create memories")
-def add_memory(memory_create: MemoryCreate):
+def add_memory(memory_create: MemoryCreate, _api_key: Optional[str] = Depends(verify_api_key)):
     """Store new memories."""
     if not any([memory_create.user_id, memory_create.agent_id, memory_create.run_id]):
         raise HTTPException(status_code=400, detail="At least one identifier (user_id, agent_id, run_id) is required.")
@@ -114,6 +162,7 @@ def get_all_memories(
     user_id: Optional[str] = None,
     run_id: Optional[str] = None,
     agent_id: Optional[str] = None,
+    _api_key: Optional[str] = Depends(verify_api_key),
 ):
     """Retrieve stored memories."""
     if not any([user_id, run_id, agent_id]):
@@ -129,7 +178,7 @@ def get_all_memories(
 
 
 @app.get("/memories/{memory_id}", summary="Get a memory")
-def get_memory(memory_id: str):
+def get_memory(memory_id: str, _api_key: Optional[str] = Depends(verify_api_key)):
     """Retrieve a specific memory by ID."""
     try:
         return MEMORY_INSTANCE.get(memory_id)
@@ -139,7 +188,7 @@ def get_memory(memory_id: str):
 
 
 @app.post("/search", summary="Search memories")
-def search_memories(search_req: SearchRequest):
+def search_memories(search_req: SearchRequest, _api_key: Optional[str] = Depends(verify_api_key)):
     """Search for memories based on a query."""
     try:
         params = {k: v for k, v in search_req.model_dump().items() if v is not None and k != "query"}
@@ -150,7 +199,7 @@ def search_memories(search_req: SearchRequest):
 
 
 @app.put("/memories/{memory_id}", summary="Update a memory")
-def update_memory(memory_id: str, updated_memory: Dict[str, Any]):
+def update_memory(memory_id: str, updated_memory: Dict[str, Any], _api_key: Optional[str] = Depends(verify_api_key)):
     """Update an existing memory with new content.
     
     Args:
@@ -168,7 +217,7 @@ def update_memory(memory_id: str, updated_memory: Dict[str, Any]):
 
 
 @app.get("/memories/{memory_id}/history", summary="Get memory history")
-def memory_history(memory_id: str):
+def memory_history(memory_id: str, _api_key: Optional[str] = Depends(verify_api_key)):
     """Retrieve memory history."""
     try:
         return MEMORY_INSTANCE.history(memory_id=memory_id)
@@ -178,7 +227,7 @@ def memory_history(memory_id: str):
 
 
 @app.delete("/memories/{memory_id}", summary="Delete a memory")
-def delete_memory(memory_id: str):
+def delete_memory(memory_id: str, _api_key: Optional[str] = Depends(verify_api_key)):
     """Delete a specific memory by ID."""
     try:
         MEMORY_INSTANCE.delete(memory_id=memory_id)
@@ -193,6 +242,7 @@ def delete_all_memories(
     user_id: Optional[str] = None,
     run_id: Optional[str] = None,
     agent_id: Optional[str] = None,
+    _api_key: Optional[str] = Depends(verify_api_key),
 ):
     """Delete all memories for a given identifier."""
     if not any([user_id, run_id, agent_id]):
@@ -209,7 +259,7 @@ def delete_all_memories(
 
 
 @app.post("/reset", summary="Reset all memories")
-def reset_memory():
+def reset_memory(_api_key: Optional[str] = Depends(verify_api_key)):
     """Completely reset stored memories."""
     try:
         MEMORY_INSTANCE.reset()

tests/llms/test_anthropic.py

@@ -0,0 +1,100 @@
+from unittest.mock import Mock, patch
+
+import pytest
+
+pytest.importorskip("anthropic", reason="anthropic package not installed")
+
+from mem0.configs.llms.anthropic import AnthropicConfig
+from mem0.configs.llms.base import BaseLlmConfig
+from mem0.llms.anthropic import AnthropicLLM
+
+
+@pytest.fixture
+def mock_anthropic_client():
+    with patch("mem0.llms.anthropic.anthropic") as mock_anthropic:
+        mock_client = Mock()
+        mock_anthropic.Anthropic.return_value = mock_client
+        yield mock_client
+
+
+def test_default_config_omits_top_p(mock_anthropic_client):
+    """Default AnthropicConfig should not set top_p to avoid conflict with temperature."""
+    config = AnthropicConfig(model="claude-3-5-sonnet-20240620", api_key="test-key")
+    assert config.top_p is None
+    assert config.temperature == 0.1
+
+
+def test_generate_response_does_not_send_top_p_by_default(mock_anthropic_client):
+    """Anthropic API rejects temperature and top_p together; top_p must be omitted by default."""
+    config = AnthropicConfig(model="claude-3-5-sonnet-20240620", api_key="test-key")
+    llm = AnthropicLLM(config)
+
+    mock_response = Mock()
+    mock_response.content = [Mock(text="Hello!")]
+    mock_anthropic_client.messages.create.return_value = mock_response
+
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Hi"},
+    ]
+
+    llm.generate_response(messages)
+
+    call_kwargs = mock_anthropic_client.messages.create.call_args[1]
+    assert "top_p" not in call_kwargs
+    assert call_kwargs["temperature"] == 0.1
+
+
+def test_generate_response_sends_top_p_alone_when_no_temperature(mock_anthropic_client):
+    """When user sets only top_p (no temperature), top_p should be sent."""
+    config = AnthropicConfig(model="claude-3-5-sonnet-20240620", api_key="test-key", top_p=0.9, temperature=None)
+    llm = AnthropicLLM(config)
+
+    mock_response = Mock()
+    mock_response.content = [Mock(text="Hello!")]
+    mock_anthropic_client.messages.create.return_value = mock_response
+
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Hi"},
+    ]
+
+    llm.generate_response(messages)
+
+    call_kwargs = mock_anthropic_client.messages.create.call_args[1]
+    assert call_kwargs["top_p"] == 0.9
+    assert "temperature" not in call_kwargs
+
+
+def test_both_set_prefers_temperature_over_top_p(mock_anthropic_client):
+    """When both temperature and top_p are set, temperature wins and top_p is dropped."""
+    config = AnthropicConfig(model="claude-3-5-sonnet-20240620", api_key="test-key", top_p=0.9, temperature=0.5)
+    llm = AnthropicLLM(config)
+
+    mock_response = Mock()
+    mock_response.content = [Mock(text="Hello!")]
+    mock_anthropic_client.messages.create.return_value = mock_response
+
+    messages = [{"role": "user", "content": "Hi"}]
+    llm.generate_response(messages)
+
+    call_kwargs = mock_anthropic_client.messages.create.call_args[1]
+    assert call_kwargs["temperature"] == 0.5
+    assert "top_p" not in call_kwargs
+
+
+def test_base_config_conversion_does_not_send_both(mock_anthropic_client):
+    """BaseLlmConfig defaults both temperature=0.1 and top_p=0.1; Anthropic must not send both."""
+    base_config = BaseLlmConfig(model="claude-3-5-sonnet-20240620", api_key="test-key")
+    llm = AnthropicLLM(base_config)
+
+    mock_response = Mock()
+    mock_response.content = [Mock(text="Hello!")]
+    mock_anthropic_client.messages.create.return_value = mock_response
+
+    messages = [{"role": "user", "content": "Hi"}]
+    llm.generate_response(messages)
+
+    call_kwargs = mock_anthropic_client.messages.create.call_args[1]
+    assert "temperature" in call_kwargs
+    assert "top_p" not in call_kwargs

tests/llms/test_azure_openai_structured.py

@@ -1,4 +1,5 @@
 from unittest import mock
+from unittest.mock import Mock
 
 from mem0.llms.azure_openai_structured import SCOPE, AzureOpenAIStructuredLLM
 
@@ -98,3 +99,75 @@ def test_init_with_placeholder_api_key_uses_default_credential(
         args, kwargs = mock_azure_openai.call_args
         assert kwargs["api_key"] is None
         assert kwargs["azure_ad_token_provider"] == "token-provider"
+
+
+@mock.patch("mem0.llms.azure_openai_structured.AzureOpenAI")
+def test_generate_response_without_tools(mock_azure_openai):
+    mock_client = Mock()
+    mock_azure_openai.return_value = mock_client
+
+    config = DummyConfig(model="test-model", azure_kwargs=DummyAzureKwargs(api_key="real-key"))
+    llm = AzureOpenAIStructuredLLM(config)
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="Hello there!"))]
+    mock_client.chat.completions.create.return_value = mock_response
+
+    messages = [{"role": "user", "content": "Hi"}]
+    response = llm.generate_response(messages)
+
+    assert response == "Hello there!"
+
+
+@mock.patch("mem0.llms.azure_openai_structured.AzureOpenAI")
+def test_generate_response_with_tools(mock_azure_openai):
+    mock_client = Mock()
+    mock_azure_openai.return_value = mock_client
+
+    config = DummyConfig(model="test-model", azure_kwargs=DummyAzureKwargs(api_key="real-key"))
+    llm = AzureOpenAIStructuredLLM(config)
+
+    mock_tool_call = Mock()
+    mock_tool_call.function.name = "add_memory"
+    mock_tool_call.function.arguments = '{"data": "sunny day"}'
+
+    mock_message = Mock()
+    mock_message.content = "I've added the memory."
+    mock_message.tool_calls = [mock_tool_call]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=mock_message)]
+    mock_client.chat.completions.create.return_value = mock_response
+
+    tools = [{"type": "function", "function": {"name": "add_memory"}}]
+    messages = [{"role": "user", "content": "Remember sunny day"}]
+    response = llm.generate_response(messages, tools=tools)
+
+    assert response["content"] == "I've added the memory."
+    assert len(response["tool_calls"]) == 1
+    assert response["tool_calls"][0]["name"] == "add_memory"
+    assert response["tool_calls"][0]["arguments"] == {"data": "sunny day"}
+
+
+@mock.patch("mem0.llms.azure_openai_structured.AzureOpenAI")
+def test_generate_response_with_tools_no_tool_calls(mock_azure_openai):
+    mock_client = Mock()
+    mock_azure_openai.return_value = mock_client
+
+    config = DummyConfig(model="test-model", azure_kwargs=DummyAzureKwargs(api_key="real-key"))
+    llm = AzureOpenAIStructuredLLM(config)
+
+    mock_message = Mock()
+    mock_message.content = "No tools needed."
+    mock_message.tool_calls = None
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=mock_message)]
+    mock_client.chat.completions.create.return_value = mock_response
+
+    tools = [{"type": "function", "function": {"name": "add_memory"}}]
+    messages = [{"role": "user", "content": "Hello"}]
+    response = llm.generate_response(messages, tools=tools)
+
+    assert response["content"] == "No tools needed."
+    assert response["tool_calls"] == []

tests/llms/test_gemini.py

@@ -117,3 +117,73 @@ def test_generate_response_with_tools(mock_gemini_client: Mock):
     assert len(response["tool_calls"]) == 1
     assert response["tool_calls"][0]["name"] == "add_memory"
     assert response["tool_calls"][0]["arguments"] == {"data": "Today is a sunny day."}
+
+
+def test_parse_response_none_content_no_tools(mock_gemini_client: Mock):
+    """Gemini can return content=None when response is blocked by safety filters."""
+    config = BaseLlmConfig(model="gemini-2.0-flash", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GeminiLLM(config)
+
+    mock_candidate = Mock(content=None)
+    mock_response = Mock(candidates=[mock_candidate])
+
+    result = llm._parse_response(mock_response, tools=None)
+    assert result == ""
+
+
+def test_parse_response_none_content_with_tools(mock_gemini_client: Mock):
+    """Gemini can return content=None when response is blocked by safety filters (tools path)."""
+    config = BaseLlmConfig(model="gemini-2.0-flash", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GeminiLLM(config)
+
+    mock_candidate = Mock(content=None)
+    mock_response = Mock(candidates=[mock_candidate])
+
+    result = llm._parse_response(mock_response, tools=[{"function": {"name": "test"}}])
+    assert result == {"content": None, "tool_calls": []}
+
+
+def test_parse_response_empty_candidates(mock_gemini_client: Mock):
+    """Gemini can return an empty candidates list."""
+    config = BaseLlmConfig(model="gemini-2.0-flash", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GeminiLLM(config)
+
+    mock_response = Mock(candidates=[])
+    result = llm._parse_response(mock_response, tools=None)
+    assert result == ""
+
+
+def test_parse_response_none_candidates(mock_gemini_client: Mock):
+    """Gemini can return candidates=None."""
+    config = BaseLlmConfig(model="gemini-2.0-flash", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GeminiLLM(config)
+
+    mock_response = Mock(candidates=None)
+    result = llm._parse_response(mock_response, tools=None)
+    assert result == ""
+
+
+def test_parse_response_empty_parts_no_tools(mock_gemini_client: Mock):
+    """Gemini can return content with an empty parts list."""
+    config = BaseLlmConfig(model="gemini-2.0-flash", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GeminiLLM(config)
+
+    mock_content = Mock(parts=[])
+    mock_candidate = Mock(content=mock_content)
+    mock_response = Mock(candidates=[mock_candidate])
+
+    result = llm._parse_response(mock_response, tools=None)
+    assert result == ""
+
+
+def test_parse_response_empty_parts_with_tools(mock_gemini_client: Mock):
+    """Gemini can return content with an empty parts list (tools path)."""
+    config = BaseLlmConfig(model="gemini-2.0-flash", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GeminiLLM(config)
+
+    mock_content = Mock(parts=[])
+    mock_candidate = Mock(content=mock_content)
+    mock_response = Mock(candidates=[mock_candidate])
+
+    result = llm._parse_response(mock_response, tools=[{"function": {"name": "test"}}])
+    assert result == {"content": None, "tool_calls": []}

tests/llms/test_minimax.py

@@ -0,0 +1,194 @@
+import os
+from unittest.mock import Mock, patch
+
+import pytest
+
+from mem0.configs.llms.base import BaseLlmConfig
+from mem0.configs.llms.minimax import MinimaxConfig
+from mem0.llms.minimax import MiniMaxLLM
+from mem0.utils.factory import LlmFactory
+
+
+@pytest.fixture
+def mock_minimax_client():
+    with patch("mem0.llms.minimax.OpenAI") as mock_openai:
+        mock_client = Mock()
+        mock_openai.return_value = mock_client
+        yield mock_client
+
+
+def test_minimax_llm_default_base_url():
+    """Default config uses MiniMax official base URL."""
+    config = BaseLlmConfig(
+        model="MiniMax-M2.7", temperature=0.7, max_tokens=100, top_p=1.0, api_key="api_key"
+    )
+    llm = MiniMaxLLM(config)
+    # OpenAI client may normalize URL with trailing slash
+    assert str(llm.client.base_url).rstrip("/") == "https://api.minimax.io/v1"
+
+
+def test_minimax_llm_env_base_url():
+    """Config uses MINIMAX_API_BASE env variable when set."""
+    provider_base_url = "https://api.provider.com/v1/"
+    os.environ["MINIMAX_API_BASE"] = provider_base_url
+    try:
+        config = MinimaxConfig(
+            model="MiniMax-M2.7",
+            temperature=0.7,
+            max_tokens=100,
+            top_p=1.0,
+            api_key="api_key",
+        )
+        llm = MiniMaxLLM(config)
+        assert str(llm.client.base_url).rstrip("/") == provider_base_url.rstrip("/")
+    finally:
+        os.environ.pop("MINIMAX_API_BASE", None)
+
+
+def test_minimax_llm_config_base_url():
+    """Config uses minimax_base_url when provided."""
+    config_base_url = "https://api.config.com/v1/"
+    config = MinimaxConfig(
+        model="MiniMax-M2.7",
+        temperature=0.7,
+        max_tokens=100,
+        top_p=1.0,
+        api_key="api_key",
+        minimax_base_url=config_base_url,
+    )
+    llm = MiniMaxLLM(config)
+    assert str(llm.client.base_url).rstrip("/") == config_base_url.rstrip("/")
+
+
+def test_minimax_llm_default_model(mock_minimax_client):
+    """Default model is MiniMax-M2.7 when not specified."""
+    config = MinimaxConfig(temperature=0.7, max_tokens=100, api_key="api_key")
+    llm = MiniMaxLLM(config)
+    assert llm.config.model == "MiniMax-M2.7"
+
+
+def test_minimax_llm_env_api_key():
+    """Uses MINIMAX_API_KEY env when api_key not in config."""
+    os.environ["MINIMAX_API_KEY"] = "env-api-key"
+    try:
+        with patch("mem0.llms.minimax.OpenAI") as mock_openai:
+            mock_client = Mock()
+            mock_openai.return_value = mock_client
+            config = MinimaxConfig(model="MiniMax-M2.7", api_key=None)
+            MiniMaxLLM(config)
+            mock_openai.assert_called_once_with(
+                api_key="env-api-key",
+                base_url="https://api.minimax.io/v1",
+            )
+    finally:
+        os.environ.pop("MINIMAX_API_KEY", None)
+
+
+def test_generate_response_without_tools(mock_minimax_client):
+    """generate_response returns text when no tools provided."""
+    config = BaseLlmConfig(
+        model="MiniMax-M2.7", temperature=0.7, max_tokens=100, top_p=1.0, api_key="api_key"
+    )
+    llm = MiniMaxLLM(config)
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Hello, how are you?"},
+    ]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="I'm doing well, thank you for asking!"))]
+    mock_minimax_client.chat.completions.create.return_value = mock_response
+
+    response = llm.generate_response(messages)
+
+    mock_minimax_client.chat.completions.create.assert_called_once_with(
+        model="MiniMax-M2.7", messages=messages, temperature=0.7, max_tokens=100, top_p=1.0
+    )
+    assert response == "I'm doing well, thank you for asking!"
+
+
+def test_generate_response_with_tools(mock_minimax_client):
+    """generate_response returns tool_calls when tools provided."""
+    config = BaseLlmConfig(
+        model="MiniMax-M2.7", temperature=0.7, max_tokens=100, top_p=1.0, api_key="api_key"
+    )
+    llm = MiniMaxLLM(config)
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Add a new memory: Today is a sunny day."},
+    ]
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "add_memory",
+                "description": "Add a memory",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"data": {"type": "string", "description": "Data to add to memory"}},
+                    "required": ["data"],
+                },
+            },
+        }
+    ]
+
+    mock_response = Mock()
+    mock_message = Mock()
+    mock_message.content = "I've added the memory for you."
+
+    mock_tool_call = Mock()
+    mock_tool_call.function.name = "add_memory"
+    mock_tool_call.function.arguments = '{"data": "Today is a sunny day."}'
+
+    mock_message.tool_calls = [mock_tool_call]
+    mock_response.choices = [Mock(message=mock_message)]
+    mock_minimax_client.chat.completions.create.return_value = mock_response
+
+    response = llm.generate_response(messages, tools=tools)
+
+    mock_minimax_client.chat.completions.create.assert_called_once_with(
+        model="MiniMax-M2.7",
+        messages=messages,
+        temperature=0.7,
+        max_tokens=100,
+        top_p=1.0,
+        tools=tools,
+        tool_choice="auto",
+    )
+
+    assert response["content"] == "I've added the memory for you."
+    assert len(response["tool_calls"]) == 1
+    assert response["tool_calls"][0]["name"] == "add_memory"
+    assert response["tool_calls"][0]["arguments"] == {"data": "Today is a sunny day."}
+
+
+def test_generate_response_with_response_format(mock_minimax_client):
+    """generate_response passes response_format to the API."""
+    config = BaseLlmConfig(
+        model="MiniMax-M2.7", temperature=0.7, max_tokens=100, top_p=1.0, api_key="api_key"
+    )
+    llm = MiniMaxLLM(config)
+    messages = [{"role": "user", "content": "Return JSON."}]
+    response_format = {"type": "json_object"}
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content='{"key": "value"}'))]
+    mock_minimax_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages, response_format=response_format)
+
+    mock_minimax_client.chat.completions.create.assert_called_once_with(
+        model="MiniMax-M2.7",
+        messages=messages,
+        temperature=0.7,
+        max_tokens=100,
+        top_p=1.0,
+        response_format={"type": "json_object"},
+    )
+
+
+def test_factory_creates_minimax_llm(mock_minimax_client):
+    """LlmFactory.create returns MiniMaxLLM for provider 'minimax'."""
+    llm = LlmFactory.create("minimax", {"model": "MiniMax-M2.7", "api_key": "test-key"})
+    assert isinstance(llm, MiniMaxLLM)
+    assert llm.config.model == "MiniMax-M2.7"

tests/memory/test_apache_age_memory.py

@@ -0,0 +1,226 @@
+from unittest.mock import MagicMock, Mock, patch
+
+# age and rank_bm25 are optional deps — mock them so tests run without install
+_age_mock = Mock()
+patch.dict("sys.modules", {
+    "age": _age_mock,
+    "age.models": Mock(),
+    "rank_bm25": Mock(),
+}).start()
+
+from mem0.memory.apache_age_memory import MemoryGraph, _cosine_similarity  # noqa: E402
+
+
+def _make_instance():
+    with patch.object(MemoryGraph, "__init__", return_value=None):
+        instance = MemoryGraph.__new__(MemoryGraph)
+        instance.llm_provider = "openai"
+        instance.llm = MagicMock()
+        instance.embedding_model = MagicMock()
+        instance.config = MagicMock()
+        instance.config.graph_store.custom_prompt = None
+        instance.ag = MagicMock()
+        instance.graph_name = "test_graph"
+        instance.threshold = 0.7
+        return instance
+
+
+class TestCosineSimilarity:
+    """Tests for the _cosine_similarity helper."""
+
+    def test_identical_vectors(self):
+        assert abs(_cosine_similarity([1, 0, 0], [1, 0, 0]) - 1.0) < 1e-6
+
+    def test_orthogonal_vectors(self):
+        assert abs(_cosine_similarity([1, 0, 0], [0, 1, 0])) < 1e-6
+
+    def test_zero_vector(self):
+        assert _cosine_similarity([0, 0, 0], [1, 2, 3]) == 0.0
+
+
+class TestRetrieveNodesFromData:
+    """Tests for _retrieve_nodes_from_data in Apache AGE MemoryGraph."""
+
+    def test_normal_entities_extracted(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = {
+            "tool_calls": [{"name": "extract_entities", "arguments": {"entities": [
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "hiking", "entity_type": "activity"},
+            ]}}]
+        }
+        result = instance._retrieve_nodes_from_data("Alice loves hiking", {"user_id": "u1"})
+        assert result == {"alice": "person", "hiking": "activity"}
+
+    def test_malformed_entity_missing_entity_type_is_skipped(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = {
+            "tool_calls": [{"name": "extract_entities", "arguments": {"entities": [
+                {"entity": "matrix multiplication", "entity_type": "task"},
+                {"entity": "task"},
+                {"entity": "ReLU", "entity_type": "task"},
+            ]}}]
+        }
+        result = instance._retrieve_nodes_from_data("some text", {"user_id": "u1"})
+        assert "matrix_multiplication" in result
+        assert "relu" in result
+        assert "task" not in result
+
+    def test_missing_entities_key_returns_empty(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = {
+            "tool_calls": [{"name": "extract_entities", "arguments": {"text": "Hello."}}]
+        }
+        result = instance._retrieve_nodes_from_data("Hello.", {"user_id": "u1"})
+        assert result == {}
+
+    def test_none_tool_calls_returns_empty(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = {"tool_calls": None}
+        result = instance._retrieve_nodes_from_data("hello world", {"user_id": "u1"})
+        assert result == {}
+
+
+class TestEstablishNodesRelationsFromData:
+    """Tests for _establish_nodes_relations_from_data in Apache AGE MemoryGraph."""
+
+    def test_none_response_does_not_crash(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = None
+        result = instance._establish_nodes_relations_from_data(
+            "Hello world", {"user_id": "u1"}, {}
+        )
+        assert result == []
+
+    def test_empty_tool_calls_returns_empty(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = {"tool_calls": []}
+        result = instance._establish_nodes_relations_from_data(
+            "Hello world", {"user_id": "u1"}, {}
+        )
+        assert result == []
+
+    def test_valid_entities_returned(self):
+        instance = _make_instance()
+        instance.llm.generate_response.return_value = {
+            "tool_calls": [{"name": "add_entities", "arguments": {"entities": [
+                {"source": "alice", "relationship": "loves", "destination": "hiking"}
+            ]}}]
+        }
+        result = instance._establish_nodes_relations_from_data(
+            "Alice loves hiking", {"user_id": "u1"}, {"alice": "person"}
+        )
+        assert len(result) == 1
+        assert result[0]["source"] == "alice"
+
+
+class TestRemoveSpacesFromEntities:
+    """Tests for _remove_spaces_from_entities."""
+
+    def test_spaces_and_case(self):
+        instance = _make_instance()
+        entities = [{"source": "Alice Smith", "relationship": "Works At", "destination": "Big Corp"}]
+        result = instance._remove_spaces_from_entities(entities)
+        assert result[0]["source"] == "alice_smith"
+        assert result[0]["relationship"] == "works_at"
+        assert result[0]["destination"] == "big_corp"
+
+
+class TestFindSimilarNode:
+    """Tests for _find_similar_node."""
+
+    def test_returns_none_when_no_nodes(self):
+        instance = _make_instance()
+        instance._exec_cypher = MagicMock(return_value=[])
+        result = instance._find_similar_node([1.0, 0.0], {"user_id": "u1"}, threshold=0.9)
+        assert result is None
+
+    def test_returns_best_match_above_threshold(self):
+        instance = _make_instance()
+        instance._exec_cypher = MagicMock(return_value=[
+            {"name": "alice", "embedding": [1.0, 0.0], "user_id": "u1"},
+            {"name": "bob", "embedding": [0.0, 1.0], "user_id": "u1"},
+        ])
+        result = instance._find_similar_node([1.0, 0.0], {"user_id": "u1"}, threshold=0.9)
+        assert result["name"] == "alice"
+
+    def test_filters_by_agent_id(self):
+        instance = _make_instance()
+        instance._exec_cypher = MagicMock(return_value=[
+            {"name": "alice", "embedding": [1.0, 0.0], "user_id": "u1", "agent_id": "a2"},
+        ])
+        result = instance._find_similar_node(
+            [1.0, 0.0], {"user_id": "u1", "agent_id": "a1"}, threshold=0.9
+        )
+        assert result is None
+
+
+class TestDeleteAll:
+    """Tests for delete_all."""
+
+    def test_calls_exec_cypher_and_commits(self):
+        instance = _make_instance()
+        instance._exec_cypher = MagicMock(return_value=[])
+        instance.delete_all({"user_id": "u1"})
+        instance._exec_cypher.assert_called_once()
+        instance.ag.commit.assert_called_once()
+
+
+class TestGetAll:
+    """Tests for get_all."""
+
+    def test_returns_formatted_results(self):
+        instance = _make_instance()
+        instance._exec_cypher = MagicMock(return_value=[
+            {"source": "alice", "relationship": "KNOWS", "target": "bob"},
+            {"source": "alice", "relationship": "LIKES", "target": "hiking"},
+        ])
+        results = instance.get_all({"user_id": "u1"}, limit=10)
+        assert len(results) == 2
+        assert results[0]["source"] == "alice"
+        assert results[0]["relationship"] == "KNOWS"
+        assert results[0]["target"] == "bob"
+
+    def test_passes_limit_to_cypher(self):
+        """Limit is enforced via LIMIT in the Cypher query, not Python slicing."""
+        instance = _make_instance()
+        instance._exec_cypher = MagicMock(return_value=[
+            {"source": "n0", "relationship": "R", "target": "m0"},
+        ])
+        instance.get_all({"user_id": "u1"}, limit=3)
+        # Verify limit was passed as a parameter to the query
+        cypher_stmt = instance._exec_cypher.call_args[0][0]
+        assert "LIMIT %s" in cypher_stmt
+        params = instance._exec_cypher.call_args[1].get("params") or instance._exec_cypher.call_args[0][2]
+        assert 3 in params
+
+
+class TestAdd:
+    """Tests for the add orchestration method."""
+
+    def test_add_returns_added_and_deleted(self):
+        instance = _make_instance()
+        instance._retrieve_nodes_from_data = MagicMock(return_value={"alice": "person"})
+        instance._establish_nodes_relations_from_data = MagicMock(return_value=[
+            {"source": "alice", "relationship": "knows", "destination": "bob"}
+        ])
+        instance._search_graph_db = MagicMock(return_value=[])
+        instance._get_delete_entities_from_search_output = MagicMock(return_value=[])
+        instance._delete_entities = MagicMock(return_value=[])
+        instance._add_entities = MagicMock(return_value=["added"])
+
+        result = instance.add("Alice knows Bob", {"user_id": "u1"})
+        assert "deleted_entities" in result
+        assert "added_entities" in result
+        assert result["added_entities"] == ["added"]
+
+
+class TestSearch:
+    """Tests for the search method."""
+
+    def test_returns_empty_when_no_search_output(self):
+        instance = _make_instance()
+        instance._retrieve_nodes_from_data = MagicMock(return_value={"alice": "person"})
+        instance._search_graph_db = MagicMock(return_value=[])
+        result = instance.search("Who is Alice?", {"user_id": "u1"})
+        assert result == []

tests/memory/test_graph_memory_soft_delete.py

@@ -0,0 +1,315 @@
+"""Tests for graph memory soft-delete behavior.
+
+Verifies that _delete_entities marks relationships as invalid (soft-delete)
+rather than permanently removing them, and that search/retrieval queries
+exclude soft-deleted relationships by default.
+
+See: https://github.com/mem0ai/mem0/issues/4187
+"""
+
+from unittest.mock import Mock, patch
+
+# Mock optional deps at module level so the import works across all Python
+# versions without triggering transitive C-extension reloads (numpy via
+# qdrant_client). This matches the pattern in test_memgraph_memory.py.
+_neo4j_mock = Mock()
+patch.dict("sys.modules", {
+    "langchain_neo4j": _neo4j_mock,
+    "rank_bm25": Mock(),
+}).start()
+
+from mem0.memory.graph_memory import MemoryGraph  # noqa: E402
+
+
+def _create_graph_memory():
+    """Create a MemoryGraph instance with mocked dependencies."""
+    with patch.object(MemoryGraph, "__init__", lambda self, *a, **kw: None):
+        mg = MemoryGraph.__new__(MemoryGraph)
+        mg.graph = Mock()
+        mg.graph.query = Mock(return_value=[])
+        mg.embedding_model = Mock()
+        mg.embedding_model.embed = Mock(return_value=[0.1] * 128)
+        mg.llm = Mock()
+        mg.node_label = ":Entity"
+        mg.threshold = 0.7
+        mg.llm_provider = "openai"
+    return mg
+
+
+class TestSoftDelete:
+    """Verify _delete_entities uses SET r.valid = false, not DELETE r."""
+
+    def test_delete_entities_sends_soft_delete_cypher(self):
+        mg = _create_graph_memory()
+        mg.graph.query.return_value = [
+            {"source": "Alice", "target": "Bob", "relationship": "KNOWS"}
+        ]
+
+        mg._delete_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1"},
+        )
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "SET r.valid = false" in cypher
+        assert "r.invalidated_at = datetime()" in cypher
+        assert "DELETE r" not in cypher
+
+    def test_delete_entities_only_targets_valid_edges(self):
+        mg = _create_graph_memory()
+        mg._delete_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1"},
+        )
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "r.valid IS NULL OR r.valid = true" in cypher
+
+    def test_delete_entities_is_idempotent(self):
+        mg = _create_graph_memory()
+        item = [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}]
+        filters = {"user_id": "user1"}
+
+        mg.graph.query.return_value = [
+            {"source": "Alice", "target": "Bob", "relationship": "KNOWS"}
+        ]
+        mg._delete_entities(item, filters)
+
+        mg.graph.query.return_value = []
+        mg._delete_entities(item, filters)
+
+        # Both calls should have the same WHERE filter
+        for c in mg.graph.query.call_args_list:
+            assert "r.valid IS NULL OR r.valid = true" in c[0][0]
+
+
+class TestSearchExcludesSoftDeleted:
+    """Verify search and get_all filter out soft-deleted relationships."""
+
+    def test_get_all_filters_soft_deleted(self):
+        mg = _create_graph_memory()
+        mg.get_all(filters={"user_id": "user1"}, limit=10)
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "r.valid IS NULL OR r.valid = true" in cypher
+
+    def test_search_graph_db_filters_both_directions(self):
+        """_search_graph_db must filter soft-deleted edges in both outgoing and incoming queries."""
+        mg = _create_graph_memory()
+        mg.graph.query.return_value = []
+
+        mg._search_graph_db(node_list=["alice"], filters={"user_id": "user1"})
+
+        cypher = mg.graph.query.call_args[0][0]
+        # The UNION query has two MATCH branches — both must filter
+        occurrences = cypher.count("r.valid IS NULL OR r.valid = true")
+        assert occurrences >= 2, (
+            f"_search_graph_db has {occurrences} valid-filter(s) but needs >= 2 "
+            "(one for outgoing, one for incoming relationships)"
+        )
+
+    def test_delete_all_still_hard_deletes(self):
+        mg = _create_graph_memory()
+        mg.delete_all(filters={"user_id": "user1"})
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "DETACH DELETE" in cypher
+
+
+class TestMergeResetsValidFlag:
+    """Verify MERGE in _add_entities sets r.valid = true.
+
+    Critical: after soft-delete, a MERGE that matches the existing
+    (invalidated) edge must reset valid=true, or the edge becomes
+    a zombie -- exists but invisible to queries.
+    """
+
+    def _run_add_entities(self, source_found, dest_found):
+        """Helper: call _add_entities with configurable node search results."""
+        mg = _create_graph_memory()
+
+        source_result = (
+            [{"elementId(source_candidate)": "src_id_1"}] if source_found else []
+        )
+        dest_result = (
+            [{"elementId(destination_candidate)": "dst_id_1"}] if dest_found else []
+        )
+
+        mg._search_source_node = Mock(return_value=source_result)
+        mg._search_destination_node = Mock(return_value=dest_result)
+
+        mg._add_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1"},
+            entity_type_map={},
+        )
+
+        cypher = mg.graph.query.call_args[0][0]
+        return cypher
+
+    def test_merge_sets_valid_true_when_source_found(self):
+        cypher = self._run_add_entities(source_found=True, dest_found=False)
+        assert "r.valid = true" in cypher
+
+    def test_merge_sets_valid_true_when_dest_found(self):
+        cypher = self._run_add_entities(source_found=False, dest_found=True)
+        assert "r.valid = true" in cypher
+
+    def test_merge_sets_valid_true_when_both_found(self):
+        cypher = self._run_add_entities(source_found=True, dest_found=True)
+        assert "r.valid = true" in cypher
+
+    def test_merge_sets_valid_true_when_neither_found(self):
+        cypher = self._run_add_entities(source_found=False, dest_found=False)
+        assert "r.valid = true" in cypher
+
+    def test_merge_clears_invalidated_at_on_resurrection(self):
+        """When a soft-deleted edge is resurrected via MERGE, invalidated_at must be cleared.
+
+        Without this, a resurrected edge (valid=true) still carries stale
+        invalidated_at metadata, which corrupts temporal reasoning queries.
+        """
+        for label, src, dst in [
+            ("source found", True, False),
+            ("dest found", False, True),
+            ("both found", True, True),
+            ("neither found", False, False),
+        ]:
+            cypher = self._run_add_entities(source_found=src, dest_found=dst)
+            assert "r.invalidated_at = null" in cypher, (
+                f"MERGE path '{label}': ON MATCH SET does not clear r.invalidated_at. "
+                "Resurrected edges will have stale invalidation timestamps."
+            )
+
+
+class TestCypherConsistency:
+    """Verify all MERGE blocks use consistent property names and variable aliases."""
+
+    def _get_merge_cypher(self, source_found, dest_found):
+        mg = _create_graph_memory()
+        mg._search_source_node = Mock(
+            return_value=[{"elementId(source_candidate)": "id1"}] if source_found else []
+        )
+        mg._search_destination_node = Mock(
+            return_value=[{"elementId(destination_candidate)": "id2"}] if dest_found else []
+        )
+        mg._add_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1"},
+            entity_type_map={},
+        )
+        return mg.graph.query.call_args[0][0]
+
+    def test_all_blocks_use_created_at_not_created(self):
+        """All MERGE blocks must use r.created_at, not r.created."""
+        for label, src, dst in [
+            ("source found", True, False),
+            ("dest found", False, True),
+            ("both found", True, True),
+            ("neither found", False, False),
+        ]:
+            cypher = self._get_merge_cypher(src, dst)
+            assert "r.created_at" in cypher, (
+                f"MERGE path '{label}': uses r.created instead of r.created_at"
+            )
+
+    def test_all_blocks_use_r_not_rel(self):
+        """All MERGE blocks must use 'r' as the relationship variable, not 'rel'."""
+        for label, src, dst in [
+            ("source found", True, False),
+            ("dest found", False, True),
+            ("both found", True, True),
+            ("neither found", False, False),
+        ]:
+            cypher = self._get_merge_cypher(src, dst)
+            assert "rel." not in cypher, (
+                f"MERGE path '{label}': uses 'rel' variable instead of 'r'"
+            )
+
+    def test_all_blocks_set_updated_at_on_create(self):
+        """All MERGE blocks must set r.updated_at on CREATE for consistent timestamps."""
+        for label, src, dst in [
+            ("source found", True, False),
+            ("dest found", False, True),
+            ("both found", True, True),
+            ("neither found", False, False),
+        ]:
+            cypher = self._get_merge_cypher(src, dst)
+            assert "r.updated_at = timestamp()" in cypher, (
+                f"MERGE path '{label}': missing r.updated_at on CREATE SET"
+            )
+
+
+class TestSoftDeleteWithFilters:
+    """Verify soft-delete works correctly with agent_id and run_id filters."""
+
+    def test_delete_entities_with_agent_id(self):
+        mg = _create_graph_memory()
+        mg._delete_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1", "agent_id": "agent1"},
+        )
+
+        cypher = mg.graph.query.call_args[0][0]
+        params = mg.graph.query.call_args[1]["params"]
+        assert "SET r.valid = false" in cypher
+        assert "agent_id: $agent_id" in cypher
+        assert params["agent_id"] == "agent1"
+
+    def test_delete_entities_with_run_id(self):
+        mg = _create_graph_memory()
+        mg._delete_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1", "run_id": "run1"},
+        )
+
+        cypher = mg.graph.query.call_args[0][0]
+        params = mg.graph.query.call_args[1]["params"]
+        assert "SET r.valid = false" in cypher
+        assert "run_id: $run_id" in cypher
+        assert params["run_id"] == "run1"
+
+    def test_get_all_with_agent_id_filters_soft_deleted(self):
+        mg = _create_graph_memory()
+        mg.get_all(filters={"user_id": "user1", "agent_id": "agent1"}, limit=10)
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "r.valid IS NULL OR r.valid = true" in cypher
+        assert "agent_id: $agent_id" in cypher
+
+    def test_merge_with_agent_id_sets_valid_true(self):
+        mg = _create_graph_memory()
+        mg._search_source_node = Mock(
+            return_value=[{"elementId(source_candidate)": "id1"}]
+        )
+        mg._search_destination_node = Mock(return_value=[])
+
+        mg._add_entities(
+            [{"source": "alice", "destination": "bob", "relationship": "KNOWS"}],
+            {"user_id": "user1", "agent_id": "agent1"},
+            entity_type_map={},
+        )
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "r.valid = true" in cypher
+        assert "agent_id: $agent_id" in cypher
+
+
+class TestResetAndCleanup:
+    """Verify reset and delete_all use hard-delete (DETACH DELETE)."""
+
+    def test_reset_uses_detach_delete(self):
+        mg = _create_graph_memory()
+        mg.reset()
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "DETACH DELETE" in cypher
+        assert "valid" not in cypher.lower()
+
+    def test_delete_all_does_not_soft_delete(self):
+        mg = _create_graph_memory()
+        mg.delete_all(filters={"user_id": "user1"})
+
+        cypher = mg.graph.query.call_args[0][0]
+        assert "DETACH DELETE" in cypher
+        assert "r.valid = false" not in cypher

tests/memory/test_main.py

@@ -188,6 +188,201 @@ async def test_async_update_memory_uses_utc_timestamps(mocker):
     _assert_utc_timestamp(payload["updated_at"])
 
 
+class TestMetadataNotMutated:
+    """Tests that metadata dicts passed to memory methods are not mutated in-place (issue #2648)."""
+
+    def test_create_memory_does_not_mutate_metadata(self, mocker):
+        memory = _build_memory_instance(mocker, Memory)
+        original_metadata = {"user_id": "test_user", "category": "sports"}
+        metadata_copy = original_metadata.copy()
+
+        memory._create_memory("test data", {"test data": [0.1, 0.2, 0.3]}, metadata=original_metadata)
+
+        assert original_metadata == metadata_copy, (
+            f"_create_memory mutated the caller's metadata dict: {original_metadata} != {metadata_copy}"
+        )
+
+    def test_create_memory_stores_correct_payload(self, mocker):
+        memory = _build_memory_instance(mocker, Memory)
+        metadata = {"user_id": "test_user", "category": "sports"}
+
+        memory._create_memory("test data", {"test data": [0.1, 0.2, 0.3]}, metadata=metadata)
+
+        payload = memory.vector_store.insert.call_args.kwargs["payloads"][0]
+        assert payload["data"] == "test data"
+        assert payload["user_id"] == "test_user"
+        assert payload["category"] == "sports"
+        assert "hash" in payload
+        assert "created_at" in payload
+
+    def test_create_memory_with_none_metadata(self, mocker):
+        memory = _build_memory_instance(mocker, Memory)
+
+        memory._create_memory("test data", {"test data": [0.1, 0.2, 0.3]}, metadata=None)
+
+        payload = memory.vector_store.insert.call_args.kwargs["payloads"][0]
+        assert payload["data"] == "test data"
+        assert "hash" in payload
+
+    def test_create_memory_shared_metadata_across_calls(self, mocker):
+        """Verify that sharing a metadata dict between multiple _create_memory calls is safe."""
+        memory = _build_memory_instance(mocker, Memory)
+        shared_metadata = {"user_id": "test_user"}
+
+        memory._create_memory("first memory", {"first memory": [0.1, 0.2, 0.3]}, metadata=shared_metadata)
+        memory._create_memory("second memory", {"second memory": [0.4, 0.5, 0.6]}, metadata=shared_metadata)
+
+        assert shared_metadata == {"user_id": "test_user"}, "shared metadata was mutated across calls"
+
+        # Verify each call got the correct data
+        first_payload = memory.vector_store.insert.call_args_list[0].kwargs["payloads"][0]
+        second_payload = memory.vector_store.insert.call_args_list[1].kwargs["payloads"][0]
+        assert first_payload["data"] == "first memory"
+        assert second_payload["data"] == "second memory"
+
+    def test_create_memory_preserves_role_and_actor_id_in_history(self, mocker):
+        """Verify that role and actor_id from metadata flow through to add_history after deepcopy."""
+        memory = _build_memory_instance(mocker, Memory)
+        metadata = {"user_id": "test_user", "role": "assistant", "actor_id": "bot-1"}
+
+        memory._create_memory("test data", {"test data": [0.1, 0.2, 0.3]}, metadata=metadata)
+
+        # Verify the payload stored in vector store has all fields
+        payload = memory.vector_store.insert.call_args.kwargs["payloads"][0]
+        assert payload["role"] == "assistant"
+        assert payload["actor_id"] == "bot-1"
+        assert payload["user_id"] == "test_user"
+        assert payload["data"] == "test data"
+
+        # Verify add_history received the correct role and actor_id
+        history_call = memory.db.add_history.call_args
+        assert history_call.kwargs["role"] == "assistant"
+        assert history_call.kwargs["actor_id"] == "bot-1"
+
+        # And the original metadata is still untouched
+        assert metadata == {"user_id": "test_user", "role": "assistant", "actor_id": "bot-1"}
+
+    def test_create_memory_with_nested_metadata_not_mutated(self, mocker):
+        """Verify deepcopy protects nested structures in metadata."""
+        memory = _build_memory_instance(mocker, Memory)
+        metadata = {"user_id": "test_user", "tags": ["important", "urgent"], "config": {"key": "val"}}
+        import copy
+        metadata_snapshot = copy.deepcopy(metadata)
+
+        memory._create_memory("test data", {"test data": [0.1, 0.2, 0.3]}, metadata=metadata)
+
+        assert metadata == metadata_snapshot, "Nested metadata structures were mutated"
+
+    def test_update_memory_does_not_mutate_metadata(self, mocker):
+        memory = _build_memory_instance(mocker, Memory)
+        memory.vector_store.get.return_value = MagicMock(
+            payload={"data": "old data", "user_id": "test_user", "created_at": "2026-01-01T00:00:00+00:00"}
+        )
+        original_metadata = {"category": "updated"}
+        metadata_copy = original_metadata.copy()
+
+        memory._update_memory("mem-id", "new data", {"new data": [0.1, 0.2, 0.3]}, metadata=original_metadata)
+
+        assert original_metadata == metadata_copy, (
+            f"_update_memory mutated the caller's metadata dict: {original_metadata} != {metadata_copy}"
+        )
+
+    def test_add_to_vector_store_no_infer_does_not_mutate_metadata(self, mocker):
+        """Verify _add_to_vector_store with infer=False doesn't leak metadata between messages."""
+        memory = _build_memory_instance(mocker, Memory)
+        memory.embedding_model.embed.return_value = [0.1, 0.2, 0.3]
+
+        original_metadata = {"user_id": "test_user"}
+        metadata_copy = original_metadata.copy()
+
+        messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there", "name": "bot-1"},
+        ]
+
+        result = memory._add_to_vector_store(messages, original_metadata, filters={}, infer=False)
+
+        # Metadata should not be mutated
+        assert original_metadata == metadata_copy, (
+            f"_add_to_vector_store mutated the caller's metadata: {original_metadata}"
+        )
+
+        # Should have created 2 memories
+        assert len(result) == 2
+        assert result[0]["role"] == "user"
+        assert result[1]["role"] == "assistant"
+        assert result[1]["actor_id"] == "bot-1"
+
+        # Verify each insert got distinct payloads with correct roles
+        insert_calls = memory.vector_store.insert.call_args_list
+        first_payload = insert_calls[0].kwargs["payloads"][0]
+        second_payload = insert_calls[1].kwargs["payloads"][0]
+        assert first_payload["role"] == "user"
+        assert "actor_id" not in first_payload  # user message has no name
+        assert second_payload["role"] == "assistant"
+        assert second_payload["actor_id"] == "bot-1"
+
+    @pytest.mark.asyncio
+    async def test_async_create_memory_does_not_mutate_metadata(self, mocker):
+        memory = _build_memory_instance(mocker, AsyncMemory)
+        original_metadata = {"user_id": "test_user", "category": "sports"}
+        metadata_copy = original_metadata.copy()
+
+        await memory._create_memory("test data", {"test data": [0.1, 0.2, 0.3]}, metadata=original_metadata)
+
+        assert original_metadata == metadata_copy, (
+            f"async _create_memory mutated the caller's metadata dict: {original_metadata} != {metadata_copy}"
+        )
+
+    @pytest.mark.asyncio
+    async def test_async_create_memory_shared_metadata_across_calls(self, mocker):
+        memory = _build_memory_instance(mocker, AsyncMemory)
+        shared_metadata = {"user_id": "test_user"}
+
+        await memory._create_memory("first memory", {"first memory": [0.1, 0.2, 0.3]}, metadata=shared_metadata)
+        await memory._create_memory("second memory", {"second memory": [0.4, 0.5, 0.6]}, metadata=shared_metadata)
+
+        assert shared_metadata == {"user_id": "test_user"}, "shared metadata was mutated across async calls"
+
+    @pytest.mark.asyncio
+    async def test_async_add_to_vector_store_no_infer_does_not_mutate_metadata(self, mocker):
+        """Verify async _add_to_vector_store with infer=False doesn't leak metadata between messages."""
+        memory = _build_memory_instance(mocker, AsyncMemory)
+        memory.embedding_model.embed.return_value = [0.1, 0.2, 0.3]
+
+        original_metadata = {"user_id": "test_user"}
+        metadata_copy = original_metadata.copy()
+
+        messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there", "name": "bot-1"},
+        ]
+
+        result = await memory._add_to_vector_store(messages, original_metadata, effective_filters={}, infer=False)
+
+        assert original_metadata == metadata_copy, (
+            f"async _add_to_vector_store mutated the caller's metadata: {original_metadata}"
+        )
+        assert len(result) == 2
+        assert result[0]["role"] == "user"
+        assert result[1]["role"] == "assistant"
+
+    @pytest.mark.asyncio
+    async def test_async_update_memory_does_not_mutate_metadata(self, mocker):
+        memory = _build_memory_instance(mocker, AsyncMemory)
+        memory.vector_store.get.return_value = MagicMock(
+            payload={"data": "old data", "user_id": "test_user", "created_at": "2026-01-01T00:00:00+00:00"}
+        )
+        original_metadata = {"category": "updated"}
+        metadata_copy = original_metadata.copy()
+
+        await memory._update_memory("mem-id", "new data", {"new data": [0.1, 0.2, 0.3]}, metadata=original_metadata)
+
+        assert original_metadata == metadata_copy, (
+            f"async _update_memory mutated the caller's metadata dict: {original_metadata} != {metadata_copy}"
+        )
+
+
 def test_normalize_iso_timestamp_to_utc_preserves_naive_values():
     assert _normalize_iso_timestamp_to_utc("2026-03-18T00:00:00") == "2026-03-18T00:00:00"
 

tests/test_graph_delete.py

@@ -0,0 +1,517 @@
+"""Tests for graph cleanup on memory deletion (issue #3245)."""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from mem0.configs.base import MemoryConfig
+
+
+class MockVectorMemory:
+    def __init__(self, memory_id, payload, score=0.8):
+        self.id = memory_id
+        self.payload = payload
+        self.score = score
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_calls_graph_cleanup_when_graph_enabled(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """When graph is enabled, delete() should call graph.delete() with memory text and filters."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    # Enable graph with a mock
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    # Set up vector store to return a memory with graph-relevant data
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1",
+        {
+            "data": "Alice likes Bob",
+            "user_id": "user-1",
+            "agent_id": "agent-1",
+            "hash": "abc",
+        },
+    )
+
+    memory.delete("mem-1")
+
+    # graph.delete should have been called with the memory text and filters
+    memory.graph.delete.assert_called_once_with(
+        "Alice likes Bob", {"user_id": "user-1", "agent_id": "agent-1"}
+    )
+
+    # _delete_memory should still have been called (vector store + history cleanup)
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_skips_graph_when_not_enabled(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """When graph is not enabled, delete() should not attempt graph cleanup."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    assert memory.enable_graph is False
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"}
+    )
+
+    result = memory.delete("mem-1")
+
+    assert result == {"message": "Memory deleted successfully!"}
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_continues_if_graph_cleanup_fails(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """If graph cleanup raises an exception, delete() should still succeed."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+    memory.graph.delete.side_effect = RuntimeError("Neo4j connection lost")
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"}
+    )
+
+    # Should not raise
+    result = memory.delete("mem-1")
+    assert result == {"message": "Memory deleted successfully!"}
+
+    # Vector store deletion should still proceed
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_skips_graph_when_no_user_id(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """Graph cleanup should be skipped if the memory has no user_id."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    # Memory with no user_id
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"data": "Some data", "hash": "abc"}
+    )
+
+    memory.delete("mem-1")
+
+    # graph.delete should NOT have been called since there's no user_id
+    memory.graph.delete.assert_not_called()
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_skips_graph_when_no_memory_text(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """Graph cleanup should be skipped if the memory has no text data."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"user_id": "user-1", "hash": "abc"}
+    )
+
+    memory.delete("mem-1")
+
+    memory.graph.delete.assert_not_called()
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_passes_all_filters_to_graph(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """Graph cleanup should include all available filters (user_id, agent_id, run_id)."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1",
+        {
+            "data": "Alice likes Bob",
+            "user_id": "user-1",
+            "agent_id": "agent-1",
+            "run_id": "run-1",
+            "hash": "abc",
+        },
+    )
+
+    memory.delete("mem-1")
+
+    memory.graph.delete.assert_called_once_with(
+        "Alice likes Bob",
+        {"user_id": "user-1", "agent_id": "agent-1", "run_id": "run-1"},
+    )
+
+
+@pytest.mark.asyncio
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+async def test_async_delete_calls_graph_cleanup(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """Async delete() should also perform graph cleanup."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import AsyncMemory
+
+    config = MemoryConfig()
+    memory = AsyncMemory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1",
+        {
+            "data": "Alice likes Bob",
+            "user_id": "user-1",
+            "hash": "abc",
+        },
+    )
+
+    result = await memory.delete("mem-1")
+
+    assert result == {"message": "Memory deleted successfully!"}
+    memory.graph.delete.assert_called_once_with("Alice likes Bob", {"user_id": "user-1"})
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@pytest.mark.asyncio
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+async def test_async_delete_continues_if_graph_cleanup_fails(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """Async delete() should continue even if graph cleanup fails."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import AsyncMemory
+
+    config = MemoryConfig()
+    memory = AsyncMemory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+    memory.graph.delete.side_effect = RuntimeError("Graph error")
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"}
+    )
+
+    result = await memory.delete("mem-1")
+    assert result == {"message": "Memory deleted successfully!"}
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_raises_for_nonexistent_memory_with_graph_enabled(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """delete() should raise ValueError for non-existent memory even with graph enabled."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    mock_vector_store.get.return_value = None
+
+    with pytest.raises(ValueError, match="Memory with id non-existent not found"):
+        memory.delete("non-existent")
+
+    memory.graph.delete.assert_not_called()
+    mock_vector_store.delete.assert_not_called()
+
+
+@pytest.mark.asyncio
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+async def test_async_delete_raises_for_nonexistent_memory_with_graph_enabled(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """Async delete() should raise ValueError for non-existent memory even with graph enabled."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_store.get.return_value = None
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import AsyncMemory
+
+    config = MemoryConfig()
+    memory = AsyncMemory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    with pytest.raises(ValueError, match="Memory with id non-existent not found"):
+        await memory.delete("non-existent")
+
+    memory.graph.delete.assert_not_called()
+    mock_vector_store.delete.assert_not_called()
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_delete_all_does_not_trigger_per_memory_graph_cleanup(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """delete_all() should use graph.delete_all(), not per-memory graph.delete()."""
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    mem1 = MockVectorMemory("mem-1", {"data": "Alice likes Bob", "user_id": "user-1"})
+    mem2 = MockVectorMemory("mem-2", {"data": "Bob likes Charlie", "user_id": "user-1"})
+    mock_vector_store.list.return_value = ([mem1, mem2], 2)
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"data": "Alice likes Bob", "user_id": "user-1"}
+    )
+
+    memory.delete_all(user_id="user-1")
+
+    # graph.delete (per-memory) should NOT be called
+    memory.graph.delete.assert_not_called()
+    # graph.delete_all (bulk) SHOULD be called
+    memory.graph.delete_all.assert_called_once_with({"user_id": "user-1"})
+
+
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_internal_delete_memory_does_not_trigger_graph_cleanup(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
+    """_delete_memory() should NOT call graph.delete() — only the public delete() does.
+
+    This ensures that the DELETE branch inside _add_to_vector_store() (which calls
+    _delete_memory directly) does not interfere with the parallel graph pipeline
+    running in _add_to_graph().
+    """
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory
+
+    config = MemoryConfig()
+    memory = Memory(config)
+
+    memory.enable_graph = True
+    memory.graph = MagicMock()
+
+    mock_vector_store.get.return_value = MockVectorMemory(
+        "mem-1", {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"}
+    )
+
+    # Call _delete_memory directly (as _add_to_vector_store does for DELETE events)
+    memory._delete_memory("mem-1")
+
+    # graph.delete should NOT have been called — graph cleanup is only in delete()
+    memory.graph.delete.assert_not_called()
+    # But vector store deletion should proceed
+    mock_vector_store.delete.assert_called_once_with(vector_id="mem-1")
+
+
+def test_graph_memory_delete_calls_internal_methods():
+    """Test that MemoryGraph.delete() calls the expected internal pipeline methods."""
+    from unittest.mock import patch as _patch
+
+    # We need to mock the Neo4j import
+    with _patch.dict("sys.modules", {"langchain_neo4j": MagicMock(), "rank_bm25": MagicMock()}):
+        from mem0.memory.graph_memory import MemoryGraph
+
+        with _patch.object(MemoryGraph, "__init__", return_value=None):
+            graph = MemoryGraph.__new__(MemoryGraph)
+
+            # Mock the internal methods
+            graph._retrieve_nodes_from_data = MagicMock(
+                return_value={"alice": "person", "bob": "person"}
+            )
+            graph._establish_nodes_relations_from_data = MagicMock(
+                return_value=[
+                    {"source": "alice", "destination": "bob", "relationship": "likes"}
+                ]
+            )
+            graph._delete_entities = MagicMock(return_value=[])
+
+            filters = {"user_id": "user-1"}
+            graph.delete("Alice likes Bob", filters)
+
+            graph._retrieve_nodes_from_data.assert_called_once_with("Alice likes Bob", filters)
+            graph._establish_nodes_relations_from_data.assert_called_once_with(
+                "Alice likes Bob", filters, {"alice": "person", "bob": "person"}
+            )
+            graph._delete_entities.assert_called_once_with(
+                [{"source": "alice", "destination": "bob", "relationship": "likes"}],
+                filters,
+            )
+
+
+def test_graph_memory_delete_skips_when_no_entities():
+    """Test that MemoryGraph.delete() does nothing when no entities are extracted."""
+    from unittest.mock import patch as _patch
+
+    with _patch.dict("sys.modules", {"langchain_neo4j": MagicMock(), "rank_bm25": MagicMock()}):
+        from mem0.memory.graph_memory import MemoryGraph
+
+        with _patch.object(MemoryGraph, "__init__", return_value=None):
+            graph = MemoryGraph.__new__(MemoryGraph)
+
+            graph._retrieve_nodes_from_data = MagicMock(return_value={})
+            graph._establish_nodes_relations_from_data = MagicMock()
+            graph._delete_entities = MagicMock()
+
+            graph.delete("Some text", {"user_id": "user-1"})
+
+            graph._retrieve_nodes_from_data.assert_called_once()
+            graph._establish_nodes_relations_from_data.assert_not_called()
+            graph._delete_entities.assert_not_called()
+
+
+def test_graph_memory_delete_handles_exception():
+    """Test that MemoryGraph.delete() catches exceptions without raising."""
+    from unittest.mock import patch as _patch
+
+    with _patch.dict("sys.modules", {"langchain_neo4j": MagicMock(), "rank_bm25": MagicMock()}):
+        from mem0.memory.graph_memory import MemoryGraph
+
+        with _patch.object(MemoryGraph, "__init__", return_value=None):
+            graph = MemoryGraph.__new__(MemoryGraph)
+
+            graph._retrieve_nodes_from_data = MagicMock(
+                side_effect=RuntimeError("LLM error")
+            )
+
+            # Should not raise
+            graph.delete("Some text", {"user_id": "user-1"})

tests/test_graph_delete_docker.py

@@ -0,0 +1,821 @@
+"""
+End-to-end tests for graph cleanup on memory deletion against real
+Neo4j, Memgraph, and Apache AGE instances running in Docker.
+
+Requires:
+    docker run -d --name mem0-neo4j-test -p 7687:7687 -e NEO4J_AUTH=neo4j/testpassword neo4j:5.23
+    docker run -d --name mem0-memgraph-test -p 7688:7687 memgraph/memgraph:latest
+    docker run -d --name mem0-age-test -p 5432:5432 -e POSTGRES_USER=postgres -e POSTGRES_PASSWORD=testpassword -e POSTGRES_DB=testdb apache/age:latest
+
+Tests are skipped automatically if the databases or required Python
+packages are not available.
+"""
+
+import hashlib
+import sys
+import warnings
+from unittest.mock import MagicMock
+
+import pytest
+
+warnings.filterwarnings("ignore")
+
+EMBEDDING_DIMS = 64
+
+# ---------------------------------------------------------------------------
+# Deterministic embedding helper (shared across backends)
+# ---------------------------------------------------------------------------
+
+
+def _make_deterministic_embedder():
+    cache = {}
+    counter = [0]
+
+    def embed(text, *args, **kwargs):
+        t = text.lower().strip()
+        if t not in cache:
+            vec = [0.0] * EMBEDDING_DIMS
+            idx = counter[0] % EMBEDDING_DIMS
+            vec[idx] = 1.0
+            h = hashlib.sha256(t.encode()).digest()
+            for i in range(EMBEDDING_DIMS):
+                vec[i] += float(h[i % len(h)]) / 25500.0
+            norm = sum(v * v for v in vec) ** 0.5
+            cache[t] = [v / norm for v in vec]
+            counter[0] += 1
+        return cache[t]
+
+    mock = MagicMock()
+    mock.embed.side_effect = embed
+    mock.config.embedding_dims = EMBEDDING_DIMS
+    return mock
+
+
+def _make_mock_llm(entities, relations):
+    """Create an LLM mock that returns specific entities and relations."""
+    mock = MagicMock()
+
+    def generate_response(messages, tools):
+        tool_names = []
+        for t in tools:
+            if isinstance(t, dict):
+                fn = t.get("function", t)
+                tool_names.append(fn.get("name", ""))
+            else:
+                tool_names.append(getattr(t, "name", str(t)))
+
+        if any("extract_entities" in n for n in tool_names):
+            return {
+                "tool_calls": [
+                    {"name": "extract_entities", "arguments": {"entities": entities}}
+                ]
+            }
+        elif any("establish" in n or "relation" in n for n in tool_names):
+            return {
+                "tool_calls": [
+                    {"name": "establish_nodes_relations", "arguments": {"entities": relations}}
+                ]
+            }
+        elif any("delete" in n for n in tool_names):
+            return {"tool_calls": []}
+        return {"tool_calls": []}
+
+    mock.generate_response.side_effect = generate_response
+    return mock
+
+
+# ===========================================================================
+# NEO4J
+# ===========================================================================
+
+
+def _port_open(host, port, timeout=1):
+    """Quick TCP check — avoids slow driver-level timeouts."""
+    import socket
+
+    try:
+        with socket.create_connection((host, port), timeout=timeout):
+            return True
+    except OSError:
+        return False
+
+
+def _neo4j_available():
+    if not _port_open("localhost", 7687):
+        return False
+    try:
+        from langchain_neo4j import Neo4jGraph
+
+        g = Neo4jGraph(
+            url="bolt://localhost:7687",
+            username="neo4j",
+            password="testpassword",
+            refresh_schema=False,
+            driver_config={"notifications_min_severity": "OFF"},
+        )
+        g.query("RETURN 1")
+        return True
+    except Exception:
+        return False
+
+
+requires_neo4j = pytest.mark.skipif(not _neo4j_available(), reason="Neo4j not available")
+
+
+@pytest.fixture
+def neo4j_graph():
+    """Create a Neo4j-backed MemoryGraph with mocked LLM/embedder."""
+    from langchain_neo4j import Neo4jGraph
+    from mem0.memory.graph_memory import MemoryGraph
+
+    mg = MemoryGraph.__new__(MemoryGraph)
+    mg.graph = Neo4jGraph(
+        url="bolt://localhost:7687",
+        username="neo4j",
+        password="testpassword",
+        refresh_schema=False,
+        driver_config={"notifications_min_severity": "OFF"},
+    )
+    mg.graph.query("MATCH (n) DETACH DELETE n")
+
+    mg.node_label = ":`__Entity__`"
+    mg.llm_provider = "openai"
+    mg.user_id = None
+    mg.threshold = 0.99
+    mg.embedding_model = _make_deterministic_embedder()
+    mg.llm = MagicMock()
+    mg.config = MagicMock()
+    mg.config.graph_store.custom_prompt = None
+    mg.config.graph_store.config.base_label = True
+
+    yield mg
+
+    mg.graph.query("MATCH (n) DETACH DELETE n")
+
+
+@requires_neo4j
+class TestNeo4jDeleteE2E:
+    def _node_count(self, mg):
+        return mg.graph.query("MATCH (n) RETURN count(n) AS cnt")[0]["cnt"]
+
+    def _valid_edge_count(self, mg):
+        return mg.graph.query(
+            "MATCH ()-[r]->() WHERE r.valid IS NULL OR r.valid = true RETURN count(r) AS cnt"
+        )[0]["cnt"]
+
+    def _invalid_edge_count(self, mg):
+        return mg.graph.query(
+            "MATCH ()-[r]->() WHERE r.valid = false RETURN count(r) AS cnt"
+        )[0]["cnt"]
+
+    def test_add_creates_graph_data(self, neo4j_graph):
+        mg = neo4j_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._node_count(mg) == 2
+        assert self._valid_edge_count(mg) == 1
+
+    def test_delete_soft_deletes_relationships(self, neo4j_graph):
+        """Neo4j delete() should set r.valid=false (soft-delete), not hard-delete."""
+        mg = neo4j_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._valid_edge_count(mg) == 1
+        assert self._invalid_edge_count(mg) == 0
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert self._valid_edge_count(mg) == 0
+        assert self._invalid_edge_count(mg) == 1  # soft-deleted, not removed
+        assert self._node_count(mg) == 2  # nodes preserved
+
+    def test_delete_preserves_other_relationships(self, neo4j_graph):
+        mg = neo4j_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Charlie", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Charlie", "relationship": "knows"}],
+        )
+        mg.add("Alice knows Charlie", {"user_id": "u1"})
+        assert self._valid_edge_count(mg) == 2
+
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert self._valid_edge_count(mg) == 1
+        assert self._invalid_edge_count(mg) == 1
+
+    def test_delete_user_isolation(self, neo4j_graph):
+        mg = neo4j_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        mg.add("Alice likes Bob", {"user_id": "u2"})
+        assert self._valid_edge_count(mg) == 2
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert self._valid_edge_count(mg) == 1
+
+    def test_delete_all_hard_deletes(self, neo4j_graph):
+        mg = neo4j_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._node_count(mg) == 2
+
+        mg.delete_all({"user_id": "u1"})
+        assert self._node_count(mg) == 0
+
+    def test_add_delete_add_cycle(self, neo4j_graph):
+        mg = neo4j_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._valid_edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert self._valid_edge_count(mg) == 0
+
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._valid_edge_count(mg) == 1
+
+
+# ===========================================================================
+# MEMGRAPH
+# ===========================================================================
+
+
+def _memgraph_available():
+    if not _port_open("localhost", 7688):
+        return False
+    try:
+        from langchain_memgraph.graphs.memgraph import Memgraph
+
+        g = Memgraph("bolt://localhost:7688", "memgraph", "memgraph")
+        g.query("RETURN 1")
+        return True
+    except Exception:
+        return False
+
+
+requires_memgraph = pytest.mark.skipif(
+    not _memgraph_available(), reason="Memgraph not available"
+)
+
+
+@pytest.fixture
+def memgraph_graph():
+    """Create a Memgraph-backed MemoryGraph with mocked LLM/embedder."""
+    from langchain_memgraph.graphs.memgraph import Memgraph
+    from mem0.memory.memgraph_memory import MemoryGraph
+
+    mg = MemoryGraph.__new__(MemoryGraph)
+    mg.graph = Memgraph("bolt://localhost:7688", "memgraph", "memgraph")
+    mg.graph.query("MATCH (n) DETACH DELETE n")
+
+    try:
+        mg.graph.query("DROP VECTOR INDEX memzero;")
+    except Exception:
+        pass
+    mg.graph.query(
+        f"CREATE VECTOR INDEX memzero ON :Entity(embedding) "
+        f"WITH CONFIG {{'dimension': {EMBEDDING_DIMS}, 'capacity': 1000, 'metric': 'cos'}};"
+    )
+    try:
+        mg.graph.query("CREATE INDEX ON :Entity(user_id);")
+    except Exception:
+        pass
+    try:
+        mg.graph.query("CREATE INDEX ON :Entity;")
+    except Exception:
+        pass
+
+    mg.llm_provider = "openai"
+    mg.user_id = None
+    mg.threshold = 0.99
+    mg.embedding_model = _make_deterministic_embedder()
+    mg.llm = MagicMock()
+    mg.config = MagicMock()
+    mg.config.graph_store.custom_prompt = None
+    mg.config.embedder.config = {"embedding_dims": EMBEDDING_DIMS}
+
+    yield mg
+
+    mg.graph.query("MATCH (n) DETACH DELETE n")
+
+
+@requires_memgraph
+class TestMemgraphDeleteE2E:
+    def _node_count(self, mg):
+        return mg.graph.query("MATCH (n:Entity) RETURN count(n) AS cnt")[0]["cnt"]
+
+    def _edge_count(self, mg):
+        return mg.graph.query("MATCH ()-[r]->() RETURN count(r) AS cnt")[0]["cnt"]
+
+    def test_add_creates_graph_data(self, memgraph_graph):
+        mg = memgraph_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._node_count(mg) == 2
+        assert self._edge_count(mg) == 1
+
+    def test_delete_hard_deletes_relationships(self, memgraph_graph):
+        """Memgraph delete() should hard-delete the relationship (DELETE r)."""
+        mg = memgraph_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert self._edge_count(mg) == 0
+        assert self._node_count(mg) == 2
+
+    def test_delete_preserves_other_relationships(self, memgraph_graph):
+        mg = memgraph_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Charlie", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Charlie", "relationship": "knows"}],
+        )
+        mg.add("Alice knows Charlie", {"user_id": "u1"})
+        assert self._edge_count(mg) == 2
+
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert self._edge_count(mg) == 1
+
+    def test_delete_user_isolation(self, memgraph_graph):
+        mg = memgraph_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        mg.add("Alice likes Bob", {"user_id": "u2"})
+        assert self._edge_count(mg) == 2
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert self._edge_count(mg) == 1
+
+    def test_delete_all_hard_deletes(self, memgraph_graph):
+        mg = memgraph_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._node_count(mg) == 2
+
+        mg.delete_all({"user_id": "u1"})
+        assert self._node_count(mg) == 0
+        assert self._edge_count(mg) == 0
+
+    def test_add_delete_add_cycle(self, memgraph_graph):
+        mg = memgraph_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert self._edge_count(mg) == 0
+
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert self._edge_count(mg) == 1
+
+
+# ===========================================================================
+# APACHE AGE
+# ===========================================================================
+
+
+def _age_available():
+    if not _port_open("localhost", 5432):
+        return False
+    try:
+        import age
+
+        ag = age.connect(
+            host="localhost",
+            port=5432,
+            dbname="testdb",
+            user="postgres",
+            password="testpassword",
+        )
+        with ag.connection.cursor() as cur:
+            cur.execute("CREATE EXTENSION IF NOT EXISTS age;")
+            cur.execute("SET search_path = ag_catalog, '$user', public;")
+        ag.connection.commit()
+        ag.close()
+        return True
+    except Exception:
+        return False
+
+
+requires_age = pytest.mark.skipif(not _age_available(), reason="Apache AGE not available")
+
+
+@pytest.fixture
+def age_graph():
+    """Create an Apache AGE-backed MemoryGraph with mocked LLM/embedder."""
+    import age
+
+    from mem0.memory.apache_age_memory import MemoryGraph
+
+    graph_name = "mem0_test_delete"
+
+    ag = age.connect(
+        graph=graph_name,
+        host="localhost",
+        port=5432,
+        dbname="testdb",
+        user="postgres",
+        password="testpassword",
+    )
+    with ag.connection.cursor() as cur:
+        cur.execute("CREATE EXTENSION IF NOT EXISTS age;")
+        cur.execute("SET search_path = ag_catalog, '$user', public;")
+    ag.connection.commit()
+    age.setUpAge(ag.connection, graph_name)
+    ag.connection.commit()
+
+    mg = MemoryGraph.__new__(MemoryGraph)
+    mg.ag = ag
+    mg.graph_name = graph_name
+    mg.llm_provider = "openai"
+    mg.user_id = None
+    mg.threshold = 0.99
+    mg.embedding_model = _make_deterministic_embedder()
+    mg.llm = MagicMock()
+    mg.config = MagicMock()
+    mg.config.graph_store.custom_prompt = None
+
+    try:
+        ag.execCypher("MATCH (n) DETACH DELETE n")
+        ag.commit()
+    except Exception:
+        ag.rollback()
+
+    yield mg
+
+    try:
+        ag.execCypher("MATCH (n) DETACH DELETE n")
+        ag.commit()
+    except Exception:
+        ag.rollback()
+    ag.close()
+
+
+def _age_node_count(mg):
+    cursor = mg.ag.execCypher("MATCH (n) RETURN count(n)", cols=["cnt"])
+    rows = cursor.fetchall()
+    return rows[0][0] if rows else 0
+
+
+def _age_edge_count(mg):
+    cursor = mg.ag.execCypher("MATCH ()-[r]->() RETURN count(r)", cols=["cnt"])
+    rows = cursor.fetchall()
+    return rows[0][0] if rows else 0
+
+
+@requires_age
+class TestApacheAgeDeleteE2E:
+
+    def test_add_creates_graph_data(self, age_graph):
+        mg = age_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert _age_node_count(mg) == 2
+        assert _age_edge_count(mg) == 1
+
+    def test_delete_hard_deletes_relationships(self, age_graph):
+        """Apache AGE delete() should hard-delete the relationship."""
+        mg = age_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert _age_edge_count(mg) == 0
+        assert _age_node_count(mg) == 2
+
+    def test_delete_preserves_other_relationships(self, age_graph):
+        mg = age_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Charlie", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Charlie", "relationship": "knows"}],
+        )
+        mg.add("Alice knows Charlie", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 2
+
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 1
+
+    def test_delete_user_isolation(self, age_graph):
+        mg = age_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        mg.add("Alice likes Bob", {"user_id": "u2"})
+        assert _age_edge_count(mg) == 2
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 1
+
+    def test_delete_all_hard_deletes(self, age_graph):
+        mg = age_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert _age_node_count(mg) == 2
+
+        mg.delete_all({"user_id": "u1"})
+        assert _age_node_count(mg) == 0
+        assert _age_edge_count(mg) == 0
+
+    def test_add_delete_add_cycle(self, age_graph):
+        mg = age_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 0
+
+        mg.add("Alice likes Bob", {"user_id": "u1"})
+        assert _age_edge_count(mg) == 1
+
+
+# ===========================================================================
+# NEPTUNE (tested via Neo4j OpenCypher — same query language)
+# ===========================================================================
+
+
+def _neptune_test_available():
+    """Neptune uses OpenCypher — we test NeptuneBase.delete() against Neo4j."""
+    if not _port_open("localhost", 7687):
+        return False
+    try:
+        # Mock langchain_aws so NeptuneBase can be imported without AWS deps
+        sys.modules.setdefault("langchain_aws", MagicMock())
+        sys.modules.setdefault("botocore", MagicMock())
+        sys.modules.setdefault("botocore.config", MagicMock())
+
+        from mem0.graphs.neptune.base import NeptuneBase  # noqa: F401
+        from langchain_neo4j import Neo4jGraph
+
+        g = Neo4jGraph(
+            url="bolt://localhost:7687",
+            username="neo4j",
+            password="testpassword",
+            refresh_schema=False,
+            driver_config={"notifications_min_severity": "OFF"},
+        )
+        g.query("RETURN 1")
+        return True
+    except Exception:
+        return False
+
+
+requires_neptune_test = pytest.mark.skipif(
+    not _neptune_test_available(),
+    reason="Neo4j not available (used as OpenCypher backend for Neptune tests)",
+)
+
+
+def _make_concrete_neptune_subclass():
+    """Create a concrete NeptuneBase subclass for testing, backed by Neo4j."""
+    # Ensure mocks are in place for import
+    sys.modules.setdefault("langchain_aws", MagicMock())
+    sys.modules.setdefault("botocore", MagicMock())
+    sys.modules.setdefault("botocore.config", MagicMock())
+
+    from mem0.graphs.neptune.base import NeptuneBase
+
+    class TestableNeptune(NeptuneBase):
+        def __init__(self):
+            pass
+
+        def _delete_entities_cypher(self, source, destination, relationship, user_id):
+            cypher = f"""
+                MATCH (n:`__Entity__` {{name: $source_name, user_id: $user_id}})
+                -[r:{relationship}]->
+                (m:`__Entity__` {{name: $dest_name, user_id: $user_id}})
+                DELETE r
+                RETURN n.name AS source, m.name AS target, type(r) AS relationship
+            """
+            return cypher, {"source_name": source, "dest_name": destination, "user_id": user_id}
+
+        def _delete_all_cypher(self, filters):
+            return (
+                "MATCH (n:`__Entity__` {user_id: $user_id}) DETACH DELETE n",
+                {"user_id": filters["user_id"]},
+            )
+
+        # Stubs for abstract methods not used in delete path
+        def _add_entities_by_source_cypher(self, *a, **kw): pass
+        def _add_entities_by_destination_cypher(self, *a, **kw): pass
+        def _add_relationship_entities_cypher(self, *a, **kw): pass
+        def _add_new_entities_cypher(self, *a, **kw): pass
+        def _search_source_node_cypher(self, *a, **kw): pass
+        def _search_destination_node_cypher(self, *a, **kw): pass
+        def _get_all_cypher(self, *a, **kw): pass
+        def _search_graph_db_cypher(self, *a, **kw): pass
+
+    return TestableNeptune
+
+
+@pytest.fixture
+def neptune_graph():
+    """NeptuneBase subclass backed by a real Neo4j container."""
+    from langchain_neo4j import Neo4jGraph
+
+    cls = _make_concrete_neptune_subclass()
+    mg = cls()
+    mg.graph = Neo4jGraph(
+        url="bolt://localhost:7687",
+        username="neo4j",
+        password="testpassword",
+        refresh_schema=False,
+        driver_config={"notifications_min_severity": "OFF"},
+    )
+    mg.graph.query("MATCH (n) DETACH DELETE n")
+
+    mg.node_label = ":`__Entity__`"
+    mg.llm_provider = "openai"
+    mg.user_id = None
+    mg.threshold = 0.99
+    mg.embedding_model = _make_deterministic_embedder()
+    mg.llm = MagicMock()
+    mg.config = MagicMock()
+    mg.config.graph_store.custom_prompt = None
+
+    yield mg
+
+    mg.graph.query("MATCH (n) DETACH DELETE n")
+
+
+def _neptune_node_count(mg):
+    return mg.graph.query("MATCH (n) RETURN count(n) AS cnt")[0]["cnt"]
+
+
+def _neptune_edge_count(mg):
+    return mg.graph.query("MATCH ()-[r]->() RETURN count(r) AS cnt")[0]["cnt"]
+
+
+def _neptune_create_entities(mg, user_id):
+    """Create test entities directly via Cypher."""
+    mg.graph.query(f"""
+        CREATE (a:`__Entity__` {{name: 'alice', user_id: '{user_id}'}})
+        CREATE (b:`__Entity__` {{name: 'bob', user_id: '{user_id}'}})
+        CREATE (a)-[:likes]->(b)
+    """)
+
+
+@requires_neptune_test
+class TestNeptuneDeleteE2E:
+    """Test NeptuneBase.delete() using Neo4j as the OpenCypher backend.
+
+    Neptune uses standard OpenCypher, the same query language as Neo4j.
+    This validates that:
+    - NeptuneBase.delete() correctly calls _delete_entities(to_be_deleted, user_id) with a string
+    - The generated Cypher from _delete_entities_cypher runs correctly
+    - User isolation works
+    """
+
+    def test_delete_removes_relationship(self, neptune_graph):
+        mg = neptune_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        _neptune_create_entities(mg, "u1")
+        assert _neptune_edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert _neptune_edge_count(mg) == 0
+        assert _neptune_node_count(mg) == 2
+
+    def test_delete_user_isolation(self, neptune_graph):
+        mg = neptune_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        _neptune_create_entities(mg, "u1")
+        _neptune_create_entities(mg, "u2")
+        assert _neptune_edge_count(mg) == 2
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert _neptune_edge_count(mg) == 1
+
+    def test_delete_passes_user_id_string_not_dict(self, neptune_graph):
+        """Verify NeptuneBase.delete() passes filters['user_id'] (string) to _delete_entities."""
+        mg = neptune_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+
+        original = mg._delete_entities
+        call_args = []
+
+        def spy(to_be_deleted, user_id):
+            call_args.append(("to_be_deleted", to_be_deleted, "user_id", user_id))
+            return original(to_be_deleted, user_id)
+
+        mg._delete_entities = spy
+
+        _neptune_create_entities(mg, "u1")
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+
+        assert len(call_args) == 1
+        assert call_args[0][3] == "u1"
+        assert isinstance(call_args[0][3], str)
+
+    def test_delete_all(self, neptune_graph):
+        mg = neptune_graph
+        _neptune_create_entities(mg, "u1")
+        assert _neptune_node_count(mg) == 2
+
+        mg.delete_all({"user_id": "u1"})
+        assert _neptune_node_count(mg) == 0
+
+    def test_add_delete_add_cycle(self, neptune_graph):
+        mg = neptune_graph
+        mg.llm = _make_mock_llm(
+            [{"entity": "Alice", "entity_type": "person"}, {"entity": "Bob", "entity_type": "person"}],
+            [{"source": "Alice", "destination": "Bob", "relationship": "likes"}],
+        )
+        _neptune_create_entities(mg, "u1")
+        assert _neptune_edge_count(mg) == 1
+
+        mg.delete("Alice likes Bob", {"user_id": "u1"})
+        assert _neptune_edge_count(mg) == 0
+
+        _neptune_create_entities(mg, "u1")
+        assert _neptune_edge_count(mg) == 1

tests/test_graph_delete_e2e.py

@@ -0,0 +1,736 @@
+"""
+End-to-end tests for graph cleanup on memory deletion (issue #3245).
+
+Uses a real Kuzu embedded database to verify that graph entities are
+correctly cleaned up when memories are deleted. LLM and embedding calls
+are mocked to provide deterministic entity extraction.
+
+Tests are skipped automatically if kuzu is not installed.
+"""
+
+import shutil
+import tempfile
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from mem0.configs.base import MemoryConfig
+
+try:
+    import kuzu  # noqa: F401
+    _kuzu_available = True
+except ImportError:
+    _kuzu_available = False
+
+requires_kuzu = pytest.mark.skipif(not _kuzu_available, reason="kuzu is not installed")
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _node_count(kuzu_graph):
+    """Return total node count in the Kuzu graph."""
+    result = kuzu_graph.execute("MATCH (n:Entity) RETURN count(n) AS cnt")
+    rows = list(result.rows_as_dict())
+    return int(rows[0]["cnt"])
+
+
+def _edge_count(kuzu_graph):
+    """Return total edge count in the Kuzu graph."""
+    result = kuzu_graph.execute("MATCH ()-[r:CONNECTED_TO]->() RETURN count(r) AS cnt")
+    rows = list(result.rows_as_dict())
+    return int(rows[0]["cnt"])
+
+
+def _get_edges(kuzu_graph):
+    """Return all edges as list of (source, relationship, destination) tuples."""
+    result = kuzu_graph.execute(
+        "MATCH (s:Entity)-[r:CONNECTED_TO]->(d:Entity) "
+        "RETURN s.name AS src, r.name AS rel, d.name AS dst"
+    )
+    return [(row["src"], row["rel"], row["dst"]) for row in result.rows_as_dict()]
+
+
+def _get_nodes(kuzu_graph):
+    """Return all node names."""
+    result = kuzu_graph.execute("MATCH (n:Entity) RETURN n.name AS name, n.user_id AS uid")
+    return [(row["name"], row["uid"]) for row in result.rows_as_dict()]
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+class MockVectorMemory:
+    """Mimics the object returned by vector_store.get()."""
+
+    def __init__(self, memory_id, payload, score=0.8):
+        self.id = memory_id
+        self.payload = payload
+        self.score = score
+
+
+@pytest.fixture
+def kuzu_graph_memory():
+    """
+    Create a real Kuzu-backed MemoryGraph with mocked LLM and embedder.
+    Yields (graph_memory_instance, kuzu_connection) then cleans up.
+    """
+    import os
+
+    import kuzu
+
+    tmpdir = tempfile.mkdtemp()
+    db_path = os.path.join(tmpdir, "test.kuzu")
+    db = kuzu.Database(db_path)
+    conn = kuzu.Connection(db)
+
+    # We'll construct the MemoryGraph by bypassing __init__ and setting up manually
+    from mem0.memory.kuzu_memory import MemoryGraph
+
+    mg = MemoryGraph.__new__(MemoryGraph)
+
+    # Real Kuzu connection
+    mg.db = db
+    mg.graph = conn
+    mg.node_label = ":Entity"
+    mg.rel_label = ":CONNECTED_TO"
+    mg.kuzu_create_schema()
+
+    # Deterministic embedding: use one-hot-style vectors per entity name
+    # to avoid accidental cosine similarity matches between different entities
+    embedding_dims = 64
+    mg.embedding_dims = embedding_dims
+
+    _embed_cache = {}
+    _embed_counter = [0]
+
+    def deterministic_embed(text):
+        """Generate a deterministic, near-orthogonal embedding for each unique text."""
+        text_lower = text.lower().strip()
+        if text_lower not in _embed_cache:
+            # Create a sparse vector — set a unique dimension to 1.0
+            vec = [0.0] * embedding_dims
+            idx = _embed_counter[0] % embedding_dims
+            vec[idx] = 1.0
+            # Add small noise to other dims so it's not exactly zero
+            import hashlib
+
+            h = hashlib.sha256(text_lower.encode()).digest()
+            for i in range(embedding_dims):
+                vec[i] += float(h[i % len(h)]) / 25500.0  # tiny noise
+            norm = sum(v * v for v in vec) ** 0.5
+            _embed_cache[text_lower] = [v / norm for v in vec]
+            _embed_counter[0] += 1
+        return _embed_cache[text_lower]
+
+    mock_embedder = MagicMock()
+    mock_embedder.embed.side_effect = deterministic_embed
+    mock_embedder.config.embedding_dims = embedding_dims
+    mg.embedding_model = mock_embedder
+
+    # Mock LLM — configured per-test via mock_embedder
+    mg.llm = MagicMock()
+    mg.llm_provider = "openai"
+    mg.user_id = None
+    # High threshold so only identical entity names merge, not similar ones
+    mg.threshold = 0.99
+    mg.config = MagicMock()
+    mg.config.graph_store.custom_prompt = None
+
+    yield mg, conn
+
+    # Cleanup
+    conn.close()
+    shutil.rmtree(tmpdir, ignore_errors=True)
+
+
+def _setup_llm_for_entities(mg, entities, relations):
+    """
+    Configure the mock LLM to return specific entities and relations.
+
+    entities: list of {"entity": str, "entity_type": str}
+    relations: list of {"source": str, "destination": str, "relationship": str}
+    """
+
+    def generate_response(messages, tools):
+        # Detect which tool is being called based on tool definition names
+        tool_names = []
+        for t in tools:
+            if isinstance(t, dict):
+                fn = t.get("function", t)
+                tool_names.append(fn.get("name", ""))
+            else:
+                tool_names.append(getattr(t, "name", str(t)))
+
+        if any("extract_entities" in n for n in tool_names):
+            return {
+                "tool_calls": [
+                    {
+                        "name": "extract_entities",
+                        "arguments": {"entities": entities},
+                    }
+                ]
+            }
+        elif any("establish" in n or "relation" in n for n in tool_names):
+            return {
+                "tool_calls": [
+                    {
+                        "name": "establish_nodes_relations",
+                        "arguments": {"entities": relations},
+                    }
+                ]
+            }
+        elif any("delete" in n for n in tool_names):
+            # For _get_delete_entities_from_search_output during add() — return nothing to delete
+            return {"tool_calls": []}
+        return {"tool_calls": []}
+
+    mg.llm.generate_response.side_effect = generate_response
+
+
+# ---------------------------------------------------------------------------
+# End-to-end tests
+# ---------------------------------------------------------------------------
+
+
+@requires_kuzu
+class TestKuzuGraphDeleteE2E:
+    """End-to-end tests using a real Kuzu database."""
+
+    def test_add_creates_nodes_and_edges(self, kuzu_graph_memory):
+        """Baseline: verify add() actually creates graph data."""
+        mg, conn = kuzu_graph_memory
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+
+        filters = {"user_id": "test_user"}
+        mg.add("Alice likes Bob", filters)
+
+        assert _node_count(conn) == 2
+        assert _edge_count(conn) == 1
+        edges = _get_edges(conn)
+        assert ("alice", "likes", "bob") in edges
+
+    def test_delete_removes_edges_created_by_add(self, kuzu_graph_memory):
+        """Core test: delete() should remove the relationships that add() created."""
+        mg, conn = kuzu_graph_memory
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+
+        filters = {"user_id": "test_user"}
+        mg.add("Alice likes Bob", filters)
+
+        assert _edge_count(conn) == 1
+
+        # Now delete using the same text — should remove the relationship
+        mg.delete("Alice likes Bob", filters)
+
+        assert _edge_count(conn) == 0
+        # Nodes remain (we don't delete nodes on single memory delete)
+        assert _node_count(conn) == 2
+
+    def test_delete_only_removes_matching_edges(self, kuzu_graph_memory):
+        """delete() should only remove edges matching the extracted relationships."""
+        mg, conn = kuzu_graph_memory
+
+        # First add: Alice likes Bob
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        filters = {"user_id": "test_user"}
+        mg.add("Alice likes Bob", filters)
+
+        # Second add: Alice knows Charlie
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Charlie", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Charlie", "relationship": "knows"},
+            ],
+        )
+        mg.add("Alice knows Charlie", filters)
+
+        assert _edge_count(conn) == 2
+
+        # Delete only the "Alice likes Bob" memory
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        mg.delete("Alice likes Bob", filters)
+
+        assert _edge_count(conn) == 1
+        edges = _get_edges(conn)
+        assert ("alice", "knows", "charlie") in edges
+        assert ("alice", "likes", "bob") not in edges
+
+    def test_delete_with_different_user_id_does_not_affect_other_users(self, kuzu_graph_memory):
+        """delete() scoped to user_id should not touch another user's graph data."""
+        mg, conn = kuzu_graph_memory
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+
+        # Add for user1
+        mg.add("Alice likes Bob", {"user_id": "user1"})
+        # Add same data for user2
+        mg.add("Alice likes Bob", {"user_id": "user2"})
+
+        assert _edge_count(conn) == 2
+
+        # Delete only user1's data
+        mg.delete("Alice likes Bob", {"user_id": "user1"})
+
+        assert _edge_count(conn) == 1
+        # Remaining edge belongs to user2
+        nodes = _get_nodes(conn)
+        user2_nodes = [n for n in nodes if n[1] == "user2"]
+        assert len(user2_nodes) == 2
+
+    def test_delete_nonexistent_relationship_is_safe(self, kuzu_graph_memory):
+        """delete() on data that doesn't exist in the graph should be a no-op."""
+        mg, conn = kuzu_graph_memory
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "hates"},
+            ],
+        )
+
+        filters = {"user_id": "test_user"}
+
+        # Nothing in the graph yet
+        assert _edge_count(conn) == 0
+        assert _node_count(conn) == 0
+
+        # Should not raise
+        mg.delete("Alice hates Bob", filters)
+
+        assert _edge_count(conn) == 0
+        assert _node_count(conn) == 0
+
+    def test_delete_with_llm_failure_does_not_raise(self, kuzu_graph_memory):
+        """If LLM fails during entity extraction, delete() should not raise."""
+        mg, conn = kuzu_graph_memory
+
+        # Make LLM raise
+        mg.llm.generate_response.side_effect = RuntimeError("LLM service down")
+
+        filters = {"user_id": "test_user"}
+
+        # Should not raise
+        mg.delete("Alice likes Bob", filters)
+
+    def test_delete_with_empty_entity_extraction(self, kuzu_graph_memory):
+        """If LLM returns no entities, delete() should be a no-op."""
+        mg, conn = kuzu_graph_memory
+
+        # Add real data
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        filters = {"user_id": "test_user"}
+        mg.add("Alice likes Bob", filters)
+        assert _edge_count(conn) == 1
+
+        # Now delete but LLM returns no entities
+        _setup_llm_for_entities(mg, entities=[], relations=[])
+        mg.delete("some text", filters)
+
+        # Data should still be there
+        assert _edge_count(conn) == 1
+
+    def test_delete_all_removes_everything_for_user(self, kuzu_graph_memory):
+        """delete_all() should remove all nodes/edges for a user (baseline behavior)."""
+        mg, conn = kuzu_graph_memory
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        filters = {"user_id": "test_user"}
+        mg.add("Alice likes Bob", filters)
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Bob", "entity_type": "person"},
+                {"entity": "Charlie", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Bob", "destination": "Charlie", "relationship": "knows"},
+            ],
+        )
+        mg.add("Bob knows Charlie", filters)
+
+        assert _node_count(conn) >= 3
+        assert _edge_count(conn) == 2
+
+        mg.delete_all(filters)
+
+        assert _node_count(conn) == 0
+        assert _edge_count(conn) == 0
+
+    def test_add_delete_add_cycle(self, kuzu_graph_memory):
+        """Verify that add → delete → re-add works correctly."""
+        mg, conn = kuzu_graph_memory
+
+        _setup_llm_for_entities(
+            mg,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        filters = {"user_id": "test_user"}
+
+        # Add
+        mg.add("Alice likes Bob", filters)
+        assert _edge_count(conn) == 1
+
+        # Delete
+        mg.delete("Alice likes Bob", filters)
+        assert _edge_count(conn) == 0
+
+        # Re-add
+        mg.add("Alice likes Bob", filters)
+        assert _edge_count(conn) == 1
+        edges = _get_edges(conn)
+        assert ("alice", "likes", "bob") in edges
+
+
+@requires_kuzu
+class TestMemoryDeleteWithGraphE2E:
+    """
+    End-to-end tests for Memory.delete() with graph enabled.
+
+    Uses a real Kuzu database for the graph store and mocks for
+    the vector store, LLM, and embedder.
+    """
+
+    @pytest.fixture
+    def memory_with_graph(self):
+        """Create a Memory instance with a real Kuzu graph backend."""
+        import os
+
+        import kuzu
+
+        tmpdir = tempfile.mkdtemp()
+
+        with (
+            patch("mem0.utils.factory.EmbedderFactory.create") as mock_embedder_factory,
+            patch("mem0.utils.factory.VectorStoreFactory.create") as mock_vector_factory,
+            patch("mem0.utils.factory.LlmFactory.create") as mock_llm_factory,
+            patch("mem0.memory.storage.SQLiteManager") as mock_sqlite,
+        ):
+            _mem_embed_cache = {}
+            _mem_embed_counter = [0]
+
+            def _mem_deterministic_embed(text, *args, **kwargs):
+                text_lower = text.lower().strip()
+                if text_lower not in _mem_embed_cache:
+                    import hashlib
+
+                    vec = [0.0] * 64
+                    idx = _mem_embed_counter[0] % 64
+                    vec[idx] = 1.0
+                    h = hashlib.sha256(text_lower.encode()).digest()
+                    for i in range(64):
+                        vec[i] += float(h[i % len(h)]) / 25500.0
+                    norm = sum(v * v for v in vec) ** 0.5
+                    _mem_embed_cache[text_lower] = [v / norm for v in vec]
+                    _mem_embed_counter[0] += 1
+                return _mem_embed_cache[text_lower]
+
+            mock_embedder = MagicMock()
+            mock_embedder.embed.side_effect = _mem_deterministic_embed
+            mock_embedder.config.embedding_dims = 64
+            mock_embedder_factory.return_value = mock_embedder
+
+            mock_vector_store = MagicMock()
+            mock_vector_factory.return_value = mock_vector_store
+
+            mock_llm = MagicMock()
+            mock_llm_factory.return_value = mock_llm
+
+            mock_sqlite.return_value = MagicMock()
+
+            from mem0.memory.main import Memory
+
+            config = MemoryConfig()
+            memory = Memory(config)
+
+            # Now wire up a real Kuzu graph
+            db_path = os.path.join(tmpdir, "test.kuzu")
+            db = kuzu.Database(db_path)
+            conn = kuzu.Connection(db)
+
+            from mem0.memory.kuzu_memory import MemoryGraph as KuzuMemoryGraph
+
+            graph = KuzuMemoryGraph.__new__(KuzuMemoryGraph)
+            graph.db = db
+            graph.graph = conn
+            graph.node_label = ":Entity"
+            graph.rel_label = ":CONNECTED_TO"
+            graph.kuzu_create_schema()
+            graph.embedding_dims = 64
+            graph.embedding_model = mock_embedder
+            graph.llm = mock_llm
+            graph.llm_provider = "openai"
+            graph.user_id = None
+            graph.threshold = 0.99
+            graph.config = MagicMock()
+            graph.config.graph_store.custom_prompt = None
+
+            memory.graph = graph
+            memory.enable_graph = True
+
+            yield memory, mock_vector_store, mock_llm, conn
+
+            conn.close()
+            shutil.rmtree(tmpdir, ignore_errors=True)
+
+    def test_memory_delete_triggers_graph_cleanup(self, memory_with_graph):
+        """
+        Full integration: Memory.delete() should clean up both vector store and graph.
+        """
+        memory, mock_vs, mock_llm, conn = memory_with_graph
+
+        # 1. Manually add entities to the graph (simulating what add() would do)
+        _setup_llm_for_memory_graph(
+            mock_llm,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        memory.graph.add("Alice likes Bob", {"user_id": "user-1"})
+        assert _edge_count(conn) == 1
+
+        # 2. Set up mock vector store to return this memory
+        mock_vs.get.return_value = MockVectorMemory(
+            "mem-1",
+            {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"},
+        )
+
+        # 3. Delete the memory
+        result = memory.delete("mem-1")
+
+        assert result == {"message": "Memory deleted successfully!"}
+
+        # 4. Verify graph was cleaned up
+        assert _edge_count(conn) == 0
+
+        # 5. Verify vector store was also cleaned up
+        mock_vs.delete.assert_called_once_with(vector_id="mem-1")
+
+    def test_memory_delete_with_graph_preserves_other_users_data(self, memory_with_graph):
+        """Deleting user1's memory should not affect user2's graph data."""
+        memory, mock_vs, mock_llm, conn = memory_with_graph
+
+        _setup_llm_for_memory_graph(
+            mock_llm,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+
+        # Add data for two users
+        memory.graph.add("Alice likes Bob", {"user_id": "user-1"})
+        memory.graph.add("Alice likes Bob", {"user_id": "user-2"})
+        assert _edge_count(conn) == 2
+
+        # Delete only user-1's memory
+        mock_vs.get.return_value = MockVectorMemory(
+            "mem-1",
+            {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"},
+        )
+        memory.delete("mem-1")
+
+        # user-2's data should be intact
+        assert _edge_count(conn) == 1
+        nodes = _get_nodes(conn)
+        remaining_user_ids = set(uid for _, uid in nodes)
+        assert "user-2" in remaining_user_ids
+
+    def test_memory_delete_graph_failure_still_deletes_vector(self, memory_with_graph):
+        """If graph cleanup fails, vector store deletion should still proceed."""
+        memory, mock_vs, mock_llm, conn = memory_with_graph
+
+        # Make LLM raise during entity extraction (graph cleanup will fail)
+        mock_llm.generate_response.side_effect = RuntimeError("LLM exploded")
+
+        mock_vs.get.return_value = MockVectorMemory(
+            "mem-1",
+            {"data": "Alice likes Bob", "user_id": "user-1", "hash": "abc"},
+        )
+
+        result = memory.delete("mem-1")
+
+        assert result == {"message": "Memory deleted successfully!"}
+        mock_vs.delete.assert_called_once_with(vector_id="mem-1")
+
+    def test_memory_delete_all_uses_bulk_not_per_memory(self, memory_with_graph):
+        """delete_all() should use delete_all() on graph, not per-memory delete()."""
+        memory, mock_vs, mock_llm, conn = memory_with_graph
+
+        _setup_llm_for_memory_graph(
+            mock_llm,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        memory.graph.add("Alice likes Bob", {"user_id": "user-1"})
+        assert _edge_count(conn) == 1
+
+        # Set up vector store to return memories for deletion
+        mem1 = MockVectorMemory("mem-1", {"data": "Alice likes Bob", "user_id": "user-1"})
+        mock_vs.list.return_value = ([mem1], 1)
+        mock_vs.get.return_value = mem1
+
+        memory.delete_all(user_id="user-1")
+
+        # After delete_all, graph should be empty (via graph.delete_all)
+        assert _edge_count(conn) == 0
+        assert _node_count(conn) == 0
+
+    def test_memory_delete_nonexistent_raises_without_graph_side_effects(self, memory_with_graph):
+        """Deleting a non-existent memory should raise ValueError without touching graph."""
+        memory, mock_vs, mock_llm, conn = memory_with_graph
+
+        # Add some graph data that should NOT be affected
+        _setup_llm_for_memory_graph(
+            mock_llm,
+            entities=[
+                {"entity": "Alice", "entity_type": "person"},
+                {"entity": "Bob", "entity_type": "person"},
+            ],
+            relations=[
+                {"source": "Alice", "destination": "Bob", "relationship": "likes"},
+            ],
+        )
+        memory.graph.add("Alice likes Bob", {"user_id": "user-1"})
+        assert _edge_count(conn) == 1
+
+        # Memory doesn't exist in vector store
+        mock_vs.get.return_value = None
+
+        with pytest.raises(ValueError, match="Memory with id non-existent not found"):
+            memory.delete("non-existent")
+
+        # Graph data should be untouched
+        assert _edge_count(conn) == 1
+
+
+def _setup_llm_for_memory_graph(mock_llm, entities, relations):
+    """Configure mock LLM for the Memory-level graph operations."""
+
+    def generate_response(messages, tools):
+        tool_names = []
+        for t in tools:
+            if isinstance(t, dict):
+                fn = t.get("function", t)
+                tool_names.append(fn.get("name", ""))
+            else:
+                tool_names.append(getattr(t, "name", str(t)))
+
+        if any("extract_entities" in n for n in tool_names):
+            return {
+                "tool_calls": [
+                    {
+                        "name": "extract_entities",
+                        "arguments": {"entities": entities},
+                    }
+                ]
+            }
+        elif any("establish" in n or "relation" in n for n in tool_names):
+            return {
+                "tool_calls": [
+                    {
+                        "name": "establish_nodes_relations",
+                        "arguments": {"entities": relations},
+                    }
+                ]
+            }
+        elif any("delete" in n for n in tool_names):
+            return {"tool_calls": []}
+        return {"tool_calls": []}
+
+    mock_llm.generate_response.side_effect = generate_response

tests/test_main.py

@@ -186,7 +186,9 @@ def test_delete(memory_instance):
 
     result = memory_instance.delete("test_id")
 
-    memory_instance._delete_memory.assert_called_once_with("test_id")
+    # delete() now fetches the memory first and passes it to _delete_memory
+    existing_memory = memory_instance.vector_store.get.return_value
+    memory_instance._delete_memory.assert_called_once_with("test_id", existing_memory)
     assert result["message"] == "Memory deleted successfully!"
 
 

tests/test_memory.py

@@ -320,3 +320,121 @@ def test_normalize_facts_mixed():
 
 def test_normalize_facts_filters_empty_strings():
     assert normalize_facts(["", "valid", ""]) == ["valid"]
+
+
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+def test_delete_nonexistent_memory_raises_error(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Test that delete() raises ValueError when memory_id does not exist
+    and does not attempt to delete from the vector store.
+
+    Issue #3849: memory.delete() fails with AttributeError when memory not found.
+    Should raise a clear ValueError instead.
+    """
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory as MemoryClass
+    config = MemoryConfig()
+    memory = MemoryClass(config)
+
+    mock_vector_store.get.return_value = None
+
+    with pytest.raises(ValueError, match="Memory with id non-existent-id not found"):
+        memory.delete("non-existent-id")
+
+    mock_vector_store.delete.assert_not_called()
+
+
+@pytest.mark.asyncio
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+async def test_async_delete_nonexistent_memory_raises_error(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Test that async delete() raises ValueError when memory_id does not exist
+    and does not attempt to delete from the vector store.
+
+    Issue #3849: memory.delete() fails with AttributeError when memory not found.
+    Should raise a clear ValueError instead.
+    """
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_store.get.return_value = None
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import AsyncMemory
+    config = MemoryConfig()
+    memory = AsyncMemory(config)
+
+    with pytest.raises(ValueError, match="Memory with id non-existent-id not found"):
+        await memory.delete("non-existent-id")
+
+    mock_vector_store.delete.assert_not_called()
+
+
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+def test_update_nonexistent_memory_raises_error(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Test that _update_memory() raises ValueError when memory_id does not exist.
+
+    Same class of bug as #3849 — vector_store.get() returns None and code
+    accesses .payload without a null check.
+    """
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory as MemoryClass
+    config = MemoryConfig()
+    memory = MemoryClass(config)
+
+    mock_vector_store.get.return_value = None
+
+    with pytest.raises(ValueError, match="Memory with id non-existent-id not found"):
+        memory._update_memory("non-existent-id", "new data", {"new data": [0.1, 0.2]})
+
+    mock_vector_store.update.assert_not_called()
+
+
+@pytest.mark.asyncio
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+async def test_async_update_nonexistent_memory_raises_error(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Test that async _update_memory() raises ValueError when memory_id does not exist.
+
+    Same class of bug as #3849 — vector_store.get() returns None and code
+    accesses .payload without a null check.
+    """
+    mock_embedder_factory.return_value = MagicMock()
+    mock_vector_store = MagicMock()
+    mock_vector_store.get.return_value = None
+    mock_vector_factory.return_value = mock_vector_store
+    mock_llm_factory.return_value = MagicMock()
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import AsyncMemory
+    config = MemoryConfig()
+    memory = AsyncMemory(config)
+
+    with pytest.raises(ValueError, match="Memory with id non-existent-id not found"):
+        await memory._update_memory("non-existent-id", "new data", {"new data": [0.1, 0.2]})
+
+    mock_vector_store.update.assert_not_called()

tests/test_server_auth.py

@@ -0,0 +1,496 @@
+"""Comprehensive E2E tests for REST API server authentication.
+
+Tests the actual server/main.py app through FastAPI's TestClient (full ASGI
+round-trip) covering:
+  - Auth disabled mode (ADMIN_API_KEY unset)
+  - Auth enabled mode (ADMIN_API_KEY set)
+  - Edge cases: empty keys, near-miss keys, timing-safe comparison, header
+    casing, response headers, startup logging, and full CRUD flows through auth.
+"""
+
+import importlib
+import logging
+import os
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+pytest.importorskip("fastapi", reason="fastapi not installed")
+
+from fastapi.testclient import TestClient
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+@pytest.fixture
+def _mock_memory():
+    """Patch Memory.from_config so the server imports without a real backend."""
+    mock_instance = MagicMock()
+    # Set up return values so CRUD endpoints return realistic responses
+    mock_instance.get.return_value = {"id": "mem-1", "memory": "test memory", "user_id": "alice"}
+    mock_instance.get_all.return_value = [
+        {"id": "mem-1", "memory": "test memory", "user_id": "alice"},
+    ]
+    mock_instance.add.return_value = {"results": [{"id": "mem-1", "event": "ADD", "memory": "test"}]}
+    mock_instance.search.return_value = [{"id": "mem-1", "memory": "test", "score": 0.9}]
+    mock_instance.update.return_value = {"message": "Memory updated"}
+    mock_instance.history.return_value = [{"id": "mem-1", "old_memory": "a", "new_memory": "b"}]
+    mock_instance.delete.return_value = None
+    mock_instance.delete_all.return_value = {"message": "Memories deleted successfully!"}
+    mock_instance.reset.return_value = None
+
+    with patch.dict(os.environ, {"OPENAI_API_KEY": "fake-key"}):
+        with patch("mem0.Memory.from_config", return_value=mock_instance):
+            yield mock_instance
+
+
+def _load_app(env_overrides: dict):
+    """Reload server/main.py with the given environment and return the FastAPI app."""
+    import server.main as server_main
+
+    with patch.dict(os.environ, env_overrides, clear=False):
+        importlib.reload(server_main)
+    return server_main.app
+
+
+# ---------------------------------------------------------------------------
+# Auth disabled (ADMIN_API_KEY not set)
+# ---------------------------------------------------------------------------
+
+class TestAuthDisabled:
+    """All endpoints should be freely accessible when ADMIN_API_KEY is empty."""
+
+    @pytest.fixture(autouse=True)
+    def _setup(self, _mock_memory):
+        self.app = _load_app({"ADMIN_API_KEY": ""})
+        self.client = TestClient(self.app)
+        self.mock = _mock_memory
+
+    def test_root_redirects_to_docs(self):
+        resp = self.client.get("/", follow_redirects=False)
+        assert resp.status_code == 307
+        assert "/docs" in resp.headers["location"]
+
+    def test_get_memory_without_key(self):
+        resp = self.client.get("/memories/mem-1")
+        assert resp.status_code == 200
+        assert resp.json()["id"] == "mem-1"
+
+    def test_get_all_memories_without_key(self):
+        resp = self.client.get("/memories", params={"user_id": "alice"})
+        assert resp.status_code == 200
+
+    def test_create_memory_without_key(self):
+        resp = self.client.post("/memories", json={
+            "messages": [{"role": "user", "content": "I like pizza"}],
+            "user_id": "alice",
+        })
+        assert resp.status_code == 200
+
+    def test_search_without_key(self):
+        resp = self.client.post("/search", json={"query": "pizza", "user_id": "alice"})
+        assert resp.status_code == 200
+
+    def test_update_memory_without_key(self):
+        resp = self.client.put("/memories/mem-1", json={"data": "updated"})
+        assert resp.status_code == 200
+
+    def test_history_without_key(self):
+        resp = self.client.get("/memories/mem-1/history")
+        assert resp.status_code == 200
+
+    def test_delete_memory_without_key(self):
+        resp = self.client.delete("/memories/mem-1")
+        assert resp.status_code == 200
+
+    def test_delete_all_without_key(self):
+        resp = self.client.delete("/memories", params={"user_id": "alice"})
+        assert resp.status_code == 200
+
+    def test_reset_without_key(self):
+        resp = self.client.post("/reset")
+        assert resp.status_code == 200
+
+    def test_configure_without_key(self):
+        self.mock.from_config = MagicMock()
+        resp = self.client.post("/configure", json={"version": "v1.1"})
+        assert resp.status_code == 200
+
+    def test_supplying_key_still_works_when_auth_disabled(self):
+        """A client that sends X-API-Key should not be penalized when auth is off."""
+        resp = self.client.get(
+            "/memories/mem-1", headers={"X-API-Key": "some-random-key"}
+        )
+        assert resp.status_code == 200
+
+    @pytest.mark.parametrize(
+        "method,path",
+        [
+            ("POST", "/configure"),
+            ("POST", "/memories"),
+            ("GET", "/memories"),
+            ("GET", "/memories/test-id"),
+            ("POST", "/search"),
+            ("PUT", "/memories/test-id"),
+            ("GET", "/memories/test-id/history"),
+            ("DELETE", "/memories/test-id"),
+            ("DELETE", "/memories"),
+            ("POST", "/reset"),
+        ],
+    )
+    def test_no_endpoint_returns_401_when_auth_disabled(self, method, path):
+        resp = self.client.request(method, path)
+        assert resp.status_code != 401, f"{method} {path} should not require auth"
+
+
+# ---------------------------------------------------------------------------
+# Auth enabled (ADMIN_API_KEY set)
+# ---------------------------------------------------------------------------
+
+class TestAuthEnabled:
+    """All protected endpoints must enforce the API key."""
+
+    API_KEY = "test-secret-key-12345"
+
+    @pytest.fixture(autouse=True)
+    def _setup(self, _mock_memory):
+        self.app = _load_app({"ADMIN_API_KEY": self.API_KEY})
+        self.client = TestClient(self.app)
+        self.mock = _mock_memory
+
+    # --- Rejection cases ---
+
+    def test_missing_key_returns_401(self):
+        resp = self.client.get("/memories/mem-1")
+        assert resp.status_code == 401
+
+    def test_missing_key_detail_mentions_header(self):
+        resp = self.client.get("/memories/mem-1")
+        assert "X-API-Key" in resp.json()["detail"]
+
+    def test_wrong_key_returns_401(self):
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": "wrong"})
+        assert resp.status_code == 401
+
+    def test_wrong_key_detail_says_invalid(self):
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": "wrong"})
+        assert "Invalid" in resp.json()["detail"]
+
+    def test_empty_string_key_returns_401(self):
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": ""})
+        assert resp.status_code == 401
+
+    def test_401_includes_www_authenticate_header(self):
+        resp = self.client.get("/memories/mem-1")
+        assert resp.headers.get("www-authenticate") == "ApiKey"
+
+    def test_near_miss_key_rejected(self):
+        """Key that differs by one character should be rejected."""
+        near_miss = self.API_KEY[:-1] + ("6" if self.API_KEY[-1] != "6" else "7")
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": near_miss})
+        assert resp.status_code == 401
+
+    def test_key_with_extra_whitespace_rejected(self):
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": f" {self.API_KEY} "})
+        assert resp.status_code == 401
+
+    def test_key_prefix_rejected(self):
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": self.API_KEY[:5]})
+        assert resp.status_code == 401
+
+    def test_key_with_different_case_rejected(self):
+        resp = self.client.get("/memories/mem-1", headers={"X-API-Key": self.API_KEY.upper()})
+        assert resp.status_code == 401
+
+    @pytest.mark.parametrize(
+        "method,path",
+        [
+            ("POST", "/configure"),
+            ("POST", "/memories"),
+            ("GET", "/memories"),
+            ("GET", "/memories/test-id"),
+            ("POST", "/search"),
+            ("PUT", "/memories/test-id"),
+            ("GET", "/memories/test-id/history"),
+            ("DELETE", "/memories/test-id"),
+            ("DELETE", "/memories"),
+            ("POST", "/reset"),
+        ],
+    )
+    def test_all_endpoints_reject_without_key(self, method, path):
+        resp = self.client.request(method, path)
+        assert resp.status_code == 401, f"{method} {path} should require auth"
+
+    @pytest.mark.parametrize(
+        "method,path",
+        [
+            ("POST", "/configure"),
+            ("POST", "/memories"),
+            ("GET", "/memories"),
+            ("GET", "/memories/test-id"),
+            ("POST", "/search"),
+            ("PUT", "/memories/test-id"),
+            ("GET", "/memories/test-id/history"),
+            ("DELETE", "/memories/test-id"),
+            ("DELETE", "/memories"),
+            ("POST", "/reset"),
+        ],
+    )
+    def test_all_endpoints_reject_wrong_key(self, method, path):
+        resp = self.client.request(method, path, headers={"X-API-Key": "wrong-key"})
+        assert resp.status_code == 401, f"{method} {path} should reject wrong key"
+
+    # --- Acceptance cases ---
+
+    def test_root_does_not_require_key(self):
+        resp = self.client.get("/", follow_redirects=False)
+        assert resp.status_code == 307
+
+    def _authed(self, method, path, **kwargs):
+        headers = kwargs.pop("headers", {})
+        headers["X-API-Key"] = self.API_KEY
+        return self.client.request(method, path, headers=headers, **kwargs)
+
+    def test_get_memory_with_key(self):
+        resp = self._authed("GET", "/memories/mem-1")
+        assert resp.status_code == 200
+        assert resp.json()["id"] == "mem-1"
+
+    def test_get_all_memories_with_key(self):
+        resp = self._authed("GET", "/memories", params={"user_id": "alice"})
+        assert resp.status_code == 200
+
+    def test_create_memory_with_key(self):
+        resp = self._authed("POST", "/memories", json={
+            "messages": [{"role": "user", "content": "I like pizza"}],
+            "user_id": "alice",
+        })
+        assert resp.status_code == 200
+        data = resp.json()
+        assert "results" in data
+
+    def test_search_with_key(self):
+        resp = self._authed("POST", "/search", json={"query": "pizza", "user_id": "alice"})
+        assert resp.status_code == 200
+
+    def test_update_memory_with_key(self):
+        resp = self._authed("PUT", "/memories/mem-1", json={"data": "updated"})
+        assert resp.status_code == 200
+
+    def test_history_with_key(self):
+        resp = self._authed("GET", "/memories/mem-1/history")
+        assert resp.status_code == 200
+
+    def test_delete_memory_with_key(self):
+        resp = self._authed("DELETE", "/memories/mem-1")
+        assert resp.status_code == 200
+
+    def test_delete_all_with_key(self):
+        resp = self._authed("DELETE", "/memories", params={"user_id": "alice"})
+        assert resp.status_code == 200
+
+    def test_reset_with_key(self):
+        resp = self._authed("POST", "/reset")
+        assert resp.status_code == 200
+
+    def test_configure_with_key(self):
+        resp = self._authed("POST", "/configure", json={"version": "v1.1"})
+        assert resp.status_code == 200
+
+
+# ---------------------------------------------------------------------------
+# Full CRUD flow through auth
+# ---------------------------------------------------------------------------
+
+class TestAuthenticatedCRUDFlow:
+    """Verify a complete create → read → search → update → history → delete
+    cycle works end-to-end through the auth layer."""
+
+    API_KEY = "flow-test-key-99"
+
+    @pytest.fixture(autouse=True)
+    def _setup(self, _mock_memory):
+        self.app = _load_app({"ADMIN_API_KEY": self.API_KEY})
+        self.client = TestClient(self.app)
+        self.mock = _mock_memory
+
+    def _authed(self, method, path, **kwargs):
+        headers = kwargs.pop("headers", {})
+        headers["X-API-Key"] = self.API_KEY
+        return self.client.request(method, path, headers=headers, **kwargs)
+
+    def test_full_crud_cycle(self):
+        # 1. Create
+        resp = self._authed("POST", "/memories", json={
+            "messages": [{"role": "user", "content": "I love fresh vegetable pizza"}],
+            "user_id": "alice",
+        })
+        assert resp.status_code == 200
+        self.mock.add.assert_called_once()
+
+        # 2. Read single
+        resp = self._authed("GET", "/memories/mem-1")
+        assert resp.status_code == 200
+        self.mock.get.assert_called_once_with("mem-1")
+
+        # 3. Read all
+        resp = self._authed("GET", "/memories", params={"user_id": "alice"})
+        assert resp.status_code == 200
+        self.mock.get_all.assert_called_once_with(user_id="alice")
+
+        # 4. Search
+        resp = self._authed("POST", "/search", json={"query": "pizza", "user_id": "alice"})
+        assert resp.status_code == 200
+        self.mock.search.assert_called_once()
+
+        # 5. Update
+        resp = self._authed("PUT", "/memories/mem-1", json={"data": "updated content"})
+        assert resp.status_code == 200
+        self.mock.update.assert_called_once()
+
+        # 6. History
+        resp = self._authed("GET", "/memories/mem-1/history")
+        assert resp.status_code == 200
+        self.mock.history.assert_called_once_with(memory_id="mem-1")
+
+        # 7. Delete single
+        resp = self._authed("DELETE", "/memories/mem-1")
+        assert resp.status_code == 200
+        self.mock.delete.assert_called_once_with(memory_id="mem-1")
+
+        # 8. Delete all
+        resp = self._authed("DELETE", "/memories", params={"user_id": "alice"})
+        assert resp.status_code == 200
+        self.mock.delete_all.assert_called_once()
+
+    def test_crud_flow_blocked_without_auth(self):
+        """Same flow should fail at every step without the key."""
+        endpoints = [
+            ("POST", "/memories", {"json": {
+                "messages": [{"role": "user", "content": "test"}], "user_id": "alice"
+            }}),
+            ("GET", "/memories/mem-1", {}),
+            ("GET", "/memories", {"params": {"user_id": "alice"}}),
+            ("POST", "/search", {"json": {"query": "pizza", "user_id": "alice"}}),
+            ("PUT", "/memories/mem-1", {"json": {"data": "x"}}),
+            ("GET", "/memories/mem-1/history", {}),
+            ("DELETE", "/memories/mem-1", {}),
+            ("DELETE", "/memories", {"params": {"user_id": "alice"}}),
+            ("POST", "/reset", {}),
+        ]
+        for method, path, kwargs in endpoints:
+            resp = self.client.request(method, path, **kwargs)
+            assert resp.status_code == 401, f"Unauthenticated {method} {path} should be 401"
+            # Verify the mock was NOT called (auth blocked before reaching handler)
+        self.mock.add.assert_not_called()
+        self.mock.get.assert_not_called()
+        self.mock.search.assert_not_called()
+        self.mock.update.assert_not_called()
+        self.mock.history.assert_not_called()
+        self.mock.delete.assert_not_called()
+        self.mock.delete_all.assert_not_called()
+        self.mock.reset.assert_not_called()
+
+
+# ---------------------------------------------------------------------------
+# Edge cases
+# ---------------------------------------------------------------------------
+
+class TestAuthEdgeCases:
+    """Boundary conditions and unusual inputs."""
+
+    @pytest.fixture(autouse=True)
+    def _setup(self, _mock_memory):
+        self.mock = _mock_memory
+
+    def test_very_long_api_key(self):
+        """Server should handle a very long key without crashing."""
+        long_key = "k" * 4096
+        app = _load_app({"ADMIN_API_KEY": long_key})
+        client = TestClient(app)
+        resp = client.get("/memories/mem-1", headers={"X-API-Key": long_key})
+        assert resp.status_code == 200
+
+    def test_special_characters_in_api_key(self):
+        """Keys with special ASCII characters should work."""
+        special_key = "sk-!@#$%^&*()_+-=[]{}|;:',.<>?/~`"
+        app = _load_app({"ADMIN_API_KEY": special_key})
+        client = TestClient(app)
+
+        resp = client.get("/memories/mem-1", headers={"X-API-Key": special_key})
+        assert resp.status_code == 200
+
+        resp = client.get("/memories/mem-1", headers={"X-API-Key": "wrong"})
+        assert resp.status_code == 401
+
+    def test_key_env_var_not_present_at_all(self):
+        """When the env var is completely absent, auth should be disabled."""
+        import server.main as server_main
+        env = os.environ.copy()
+        env.pop("ADMIN_API_KEY", None)
+        with patch.dict(os.environ, env, clear=True):
+            importlib.reload(server_main)
+        client = TestClient(server_main.app)
+        resp = client.get("/memories/mem-1")
+        assert resp.status_code != 401
+
+    def test_switching_from_enabled_to_disabled(self):
+        """Simulates a server restart with auth toggled off."""
+        # First: auth enabled
+        app1 = _load_app({"ADMIN_API_KEY": "secret"})
+        c1 = TestClient(app1)
+        assert c1.get("/memories/mem-1").status_code == 401
+
+        # Then: auth disabled
+        app2 = _load_app({"ADMIN_API_KEY": ""})
+        c2 = TestClient(app2)
+        assert c2.get("/memories/mem-1").status_code != 401
+
+    def test_openapi_schema_accessible_without_key(self):
+        """The /docs and /openapi.json endpoints should always be reachable."""
+        app = _load_app({"ADMIN_API_KEY": "secret"})
+        client = TestClient(app)
+
+        resp = client.get("/openapi.json")
+        assert resp.status_code == 200
+        schema = resp.json()
+        assert "paths" in schema
+
+        resp = client.get("/docs")
+        assert resp.status_code == 200
+
+    def test_openapi_schema_documents_auth(self):
+        """The OpenAPI schema should mention authentication."""
+        app = _load_app({"ADMIN_API_KEY": "secret"})
+        client = TestClient(app)
+        schema = client.get("/openapi.json").json()
+        assert "Authentication" in schema.get("info", {}).get("description", "")
+
+
+# ---------------------------------------------------------------------------
+# Startup logging
+# ---------------------------------------------------------------------------
+
+class TestStartupLogging:
+    """Verify the server emits the correct log messages at import time."""
+
+    @pytest.fixture(autouse=True)
+    def _setup(self, _mock_memory):
+        pass
+
+    def test_warning_when_auth_disabled(self, caplog):
+        with caplog.at_level(logging.WARNING):
+            _load_app({"ADMIN_API_KEY": ""})
+        assert any("UNSECURED" in r.message for r in caplog.records)
+
+    def test_info_when_auth_enabled(self, caplog):
+        with caplog.at_level(logging.INFO):
+            _load_app({"ADMIN_API_KEY": "a-long-enough-secret-key"})
+        assert any("authentication enabled" in r.message for r in caplog.records)
+
+    def test_warning_when_key_too_short(self, caplog):
+        with caplog.at_level(logging.WARNING):
+            _load_app({"ADMIN_API_KEY": "short"})
+        assert any("shorter than" in r.message for r in caplog.records)

tests/test_server_params.py

@@ -0,0 +1,507 @@
+"""Tests for REST API parameter forwarding.
+
+Verifies that the Pydantic request models in server/main.py correctly accept
+and forward all parameters supported by the underlying Memory class methods,
+including limit, threshold, infer, memory_type, and prompt — which were
+previously silently dropped by Pydantic v2's default extra='ignore' behavior.
+"""
+
+import importlib
+import os
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+pytest.importorskip("fastapi", reason="fastapi not installed")
+
+from fastapi.testclient import TestClient
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+@pytest.fixture
+def _mock_memory():
+    """Patch Memory.from_config so the server imports without a real backend."""
+    mock_instance = MagicMock()
+    mock_instance.add.return_value = {"results": [{"id": "mem-1", "event": "ADD", "memory": "test"}]}
+    mock_instance.search.return_value = [{"id": "mem-1", "memory": "test", "score": 0.9}]
+    mock_instance.get.return_value = {"id": "mem-1", "memory": "test memory"}
+    mock_instance.get_all.return_value = [{"id": "mem-1", "memory": "test memory"}]
+    mock_instance.update.return_value = {"message": "Memory updated"}
+    mock_instance.history.return_value = [{"id": "mem-1", "old_memory": "a", "new_memory": "b"}]
+    mock_instance.delete.return_value = None
+    mock_instance.delete_all.return_value = {"message": "Memories deleted"}
+    mock_instance.reset.return_value = None
+
+    with patch.dict(os.environ, {"OPENAI_API_KEY": "fake-key", "ADMIN_API_KEY": ""}):
+        with patch("mem0.Memory.from_config", return_value=mock_instance):
+            yield mock_instance
+
+
+@pytest.fixture
+def client(_mock_memory):
+    """Return a TestClient wired to the server app with mocked Memory."""
+    import server.main as server_main
+    with patch.dict(os.environ, {"ADMIN_API_KEY": ""}):
+        importlib.reload(server_main)
+    return TestClient(server_main.app)
+
+
+@pytest.fixture
+def mock_memory(_mock_memory):
+    return _mock_memory
+
+
+# ===========================================================================
+# SearchRequest: limit parameter
+# ===========================================================================
+
+class TestSearchLimit:
+    """Verify that the limit parameter is accepted and forwarded to Memory.search()."""
+
+    def test_limit_forwarded(self, client, mock_memory):
+        resp = client.post("/search", json={"query": "food", "user_id": "u1", "limit": 5})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["limit"] == 5
+
+    def test_limit_one(self, client, mock_memory):
+        resp = client.post("/search", json={"query": "food", "user_id": "u1", "limit": 1})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["limit"] == 1
+
+    def test_limit_omitted_uses_memory_default(self, client, mock_memory):
+        """When limit is not sent, it should not appear in the kwargs,
+        allowing Memory.search() to use its own default (100)."""
+        resp = client.post("/search", json={"query": "food", "user_id": "u1"})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert "limit" not in kwargs
+
+
+# ===========================================================================
+# SearchRequest: threshold parameter
+# ===========================================================================
+
+class TestSearchThreshold:
+    """Verify that the threshold parameter is accepted and forwarded."""
+
+    def test_threshold_forwarded(self, client, mock_memory):
+        resp = client.post("/search", json={"query": "food", "user_id": "u1", "threshold": 0.8})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["threshold"] == 0.8
+
+    def test_threshold_zero(self, client, mock_memory):
+        """threshold=0.0 is a valid falsy value that must not be filtered out."""
+        resp = client.post("/search", json={"query": "food", "user_id": "u1", "threshold": 0.0})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["threshold"] == 0.0
+
+    def test_threshold_omitted_uses_memory_default(self, client, mock_memory):
+        resp = client.post("/search", json={"query": "food", "user_id": "u1"})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert "threshold" not in kwargs
+
+
+# ===========================================================================
+# SearchRequest: limit + threshold together
+# ===========================================================================
+
+class TestSearchLimitAndThreshold:
+
+    def test_both_forwarded(self, client, mock_memory):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "limit": 10, "threshold": 0.5
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["limit"] == 10
+        assert kwargs["threshold"] == 0.5
+
+
+# ===========================================================================
+# MemoryCreate: infer parameter
+# ===========================================================================
+
+class TestAddInfer:
+    """Verify that the infer parameter is accepted and forwarded to Memory.add()."""
+
+    def test_infer_false_forwarded(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "Store this exactly"}],
+            "user_id": "u1",
+            "infer": False,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["infer"] is False
+
+    def test_infer_true_forwarded(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "I like pizza"}],
+            "user_id": "u1",
+            "infer": True,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["infer"] is True
+
+    def test_infer_omitted_uses_memory_default(self, client, mock_memory):
+        """When infer is not sent, it should not appear in kwargs,
+        allowing Memory.add() to use its own default (True)."""
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "hello"}],
+            "user_id": "u1",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert "infer" not in kwargs
+
+
+# ===========================================================================
+# MemoryCreate: memory_type parameter
+# ===========================================================================
+
+class TestAddMemoryType:
+    """Verify that the memory_type parameter is accepted and forwarded."""
+
+    def test_memory_type_forwarded(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "I like pizza"}],
+            "user_id": "u1",
+            "memory_type": "core",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["memory_type"] == "core"
+
+    def test_memory_type_omitted(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "hello"}],
+            "user_id": "u1",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert "memory_type" not in kwargs
+
+
+# ===========================================================================
+# MemoryCreate: prompt parameter
+# ===========================================================================
+
+class TestAddPrompt:
+    """Verify that the prompt parameter is accepted and forwarded."""
+
+    def test_prompt_forwarded(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "I like pizza"}],
+            "user_id": "u1",
+            "prompt": "Extract food preferences only.",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["prompt"] == "Extract food preferences only."
+
+    def test_prompt_omitted(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "hello"}],
+            "user_id": "u1",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert "prompt" not in kwargs
+
+
+# ===========================================================================
+# MemoryCreate: all new params together
+# ===========================================================================
+
+class TestAddAllNewParams:
+
+    def test_infer_memory_type_and_prompt_together(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "I like pizza"}],
+            "user_id": "u1",
+            "infer": False,
+            "memory_type": "core",
+            "prompt": "Custom extraction prompt.",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["infer"] is False
+        assert kwargs["memory_type"] == "core"
+        assert kwargs["prompt"] == "Custom extraction prompt."
+
+
+# ===========================================================================
+# Edge cases: falsy-but-valid values must not be filtered out
+# ===========================================================================
+
+class TestFalsyValues:
+    """The handler filters with `v is not None`. Falsy values like False, 0,
+    0.0, and empty string must still be forwarded."""
+
+    def test_infer_false_not_filtered(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "infer": False,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["infer"] is False
+
+    def test_threshold_zero_not_filtered(self, client, mock_memory):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "threshold": 0.0,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["threshold"] == 0.0
+
+
+# ===========================================================================
+# Extra/unknown fields are still silently ignored (existing Pydantic behavior)
+# ===========================================================================
+
+class TestUnknownFieldsIgnored:
+
+    def test_unknown_search_field_ignored(self, client, mock_memory):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "bogus_field": "xyz",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert "bogus_field" not in kwargs
+
+    def test_unknown_add_field_ignored(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "unknown_param": 42,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert "unknown_param" not in kwargs
+
+
+# ===========================================================================
+# Backward compatibility: existing params still work
+# ===========================================================================
+
+class TestExistingParamsUnchanged:
+
+    def test_search_filters_still_forwarded(self, client, mock_memory):
+        resp = client.post("/search", json={
+            "query": "food",
+            "user_id": "u1",
+            "agent_id": "a1",
+            "filters": {"category": "food"},
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["user_id"] == "u1"
+        assert kwargs["agent_id"] == "a1"
+        assert kwargs["filters"] == {"category": "food"}
+
+    def test_add_metadata_still_forwarded(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "agent_id": "a1",
+            "metadata": {"source": "test"},
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["user_id"] == "u1"
+        assert kwargs["agent_id"] == "a1"
+        assert kwargs["metadata"] == {"source": "test"}
+
+
+# ===========================================================================
+# OpenAPI schema: new fields are documented
+# ===========================================================================
+
+class TestOpenAPISchema:
+    """Verify the new fields appear in the auto-generated OpenAPI schema."""
+
+    def test_search_schema_includes_limit(self, client):
+        schema = client.get("/openapi.json").json()
+        search_props = schema["components"]["schemas"]["SearchRequest"]["properties"]
+        assert "limit" in search_props
+        assert search_props["limit"]["description"] == "Maximum number of results to return."
+
+    def test_search_schema_includes_threshold(self, client):
+        schema = client.get("/openapi.json").json()
+        search_props = schema["components"]["schemas"]["SearchRequest"]["properties"]
+        assert "threshold" in search_props
+
+    def test_add_schema_includes_infer(self, client):
+        schema = client.get("/openapi.json").json()
+        add_props = schema["components"]["schemas"]["MemoryCreate"]["properties"]
+        assert "infer" in add_props
+
+    def test_add_schema_includes_memory_type(self, client):
+        schema = client.get("/openapi.json").json()
+        add_props = schema["components"]["schemas"]["MemoryCreate"]["properties"]
+        assert "memory_type" in add_props
+
+    def test_add_schema_includes_prompt(self, client):
+        schema = client.get("/openapi.json").json()
+        add_props = schema["components"]["schemas"]["MemoryCreate"]["properties"]
+        assert "prompt" in add_props
+
+
+# ===========================================================================
+# Pydantic type validation: invalid types return 422
+# ===========================================================================
+
+class TestTypeValidation:
+    """Verify FastAPI/Pydantic rejects invalid types with 422."""
+
+    def test_limit_string_rejected(self, client):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "limit": "not_a_number",
+        })
+        assert resp.status_code == 422
+
+    def test_threshold_string_rejected(self, client):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "threshold": "high",
+        })
+        assert resp.status_code == 422
+
+    def test_infer_string_coerced_by_pydantic(self, client, mock_memory):
+        """Pydantic v2 coerces truthy strings like 'yes' to True for bool fields."""
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "infer": "yes",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert kwargs["infer"] is True
+
+    def test_infer_invalid_value_rejected(self, client):
+        """A value that cannot be coerced to bool should be rejected."""
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "infer": [1, 2, 3],
+        })
+        assert resp.status_code == 422
+
+    def test_limit_float_rejected(self, client):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "limit": 5.7,
+        })
+        assert resp.status_code == 422
+
+    def test_memory_type_int_rejected(self, client):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "memory_type": 123,
+        })
+        assert resp.status_code == 422
+
+
+# ===========================================================================
+# Explicit null values: treated as omitted (filtered by `is not None`)
+# ===========================================================================
+
+class TestExplicitNull:
+    """When a client sends null for an optional field, it should be treated
+    as omitted — the Memory class default should be used."""
+
+    def test_limit_null_uses_memory_default(self, client, mock_memory):
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "limit": None,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert "limit" not in kwargs
+
+    def test_infer_null_uses_memory_default(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "infer": None,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert "infer" not in kwargs
+
+    def test_prompt_null_uses_memory_default(self, client, mock_memory):
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "test"}],
+            "user_id": "u1",
+            "prompt": None,
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        assert "prompt" not in kwargs
+
+
+# ===========================================================================
+# Verify exact call signatures match Memory method params
+# ===========================================================================
+
+class TestCallSignatureMatch:
+    """Ensure forwarded params exactly match Memory.add() and Memory.search()
+    keyword argument names — a typo here would cause a TypeError at runtime."""
+
+    def test_search_kwargs_are_valid(self, client, mock_memory):
+        """All kwargs forwarded to Memory.search() must be in its signature."""
+        resp = client.post("/search", json={
+            "query": "food", "user_id": "u1", "agent_id": "a1",
+            "run_id": "r1", "filters": {"k": "v"},
+            "limit": 10, "threshold": 0.5,
+        })
+        assert resp.status_code == 200
+        # The handler passes query= as a keyword arg, so it appears in kwargs too
+        _, kwargs = mock_memory.search.call_args
+        valid_params = {"query", "user_id", "agent_id", "run_id", "limit", "filters", "threshold", "rerank"}
+        for key in kwargs:
+            assert key in valid_params, f"Unexpected kwarg '{key}' forwarded to Memory.search()"
+
+    def test_add_kwargs_are_valid(self, client, mock_memory):
+        """All kwargs forwarded to Memory.add() must be in its signature."""
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "hi"}],
+            "user_id": "u1", "agent_id": "a1", "run_id": "r1",
+            "metadata": {"k": "v"},
+            "infer": False, "memory_type": "core", "prompt": "custom",
+        })
+        assert resp.status_code == 200
+        # The handler passes messages= as a keyword arg, so it appears in kwargs too
+        _, kwargs = mock_memory.add.call_args
+        valid_params = {"messages", "user_id", "agent_id", "run_id", "metadata", "infer", "memory_type", "prompt"}
+        for key in kwargs:
+            assert key in valid_params, f"Unexpected kwarg '{key}' forwarded to Memory.add()"
+
+    def test_messages_excluded_from_params_dict(self, client, mock_memory):
+        """messages is passed separately via messages= kwarg, not duplicated from model_dump."""
+        resp = client.post("/memories", json={
+            "messages": [{"role": "user", "content": "hi"}],
+            "user_id": "u1",
+        })
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.add.call_args
+        # messages should be present (passed explicitly) and be a list of dicts
+        assert "messages" in kwargs
+        assert isinstance(kwargs["messages"], list)
+        assert kwargs["messages"][0] == {"role": "user", "content": "hi"}
+
+    def test_query_passed_explicitly(self, client, mock_memory):
+        """query is passed as an explicit keyword arg to Memory.search()."""
+        resp = client.post("/search", json={"query": "food", "user_id": "u1"})
+        assert resp.status_code == 200
+        _, kwargs = mock_memory.search.call_args
+        assert kwargs["query"] == "food"

tests/vector_stores/test_chroma.py

@@ -3,6 +3,7 @@ from unittest.mock import Mock, patch
 import pytest
 
 from mem0.vector_stores.chroma import ChromaDB
+from mem0.configs.vector_stores.chroma import ChromaDbConfig
 
 
 @pytest.fixture
@@ -249,3 +250,15 @@ def test_generate_where_clause_non_string_values():
     # ChromaDB accepts non-string values in filters
     expected = {"$and": [{"user_id": {"$eq": "alice"}}, {"count": {"$eq": 5}}, {"active": {"$eq": True}}]}
     assert result == expected
+
+
+def test_chroma_config_accepts_default_tmp_path():
+    """Test that ChromaDbConfig accepts the default /tmp/chroma path."""
+    config = ChromaDbConfig(path="/tmp/chroma")
+    assert config.path == "/tmp/chroma"
+
+
+def test_chroma_config_rejects_no_config():
+    """Test that ChromaDbConfig rejects when no connection config is provided."""
+    with pytest.raises(ValueError):
+        ChromaDbConfig()

tests/vector_stores/test_databricks.py

@@ -1,8 +1,12 @@
 from types import SimpleNamespace
 from unittest.mock import MagicMock, patch
+
+import pytest
+
+pytest.importorskip("databricks", reason="databricks-sdk package not installed")
+
 from databricks.sdk.service.vectorsearch import VectorIndexType, QueryVectorIndexResponse, ResultManifest, ResultData, ColumnInfo
 from mem0.vector_stores.databricks import Databricks
-import pytest
 
 
 # ---------------------- Fixtures ---------------------- #
@@ -205,9 +209,34 @@ def test_search_direct_access_vector(db_instance_direct, mock_workspace_client):
     assert results[0].score == 0.77
 
 
+def test_search_delta_sync_self_managed_vectors(mock_workspace_client):
+    """DELTA_SYNC without embedding model endpoint should use query_vector, not query_text."""
+    mock_workspace_client.tables.exists.return_value = SimpleNamespace(table_exists=True)
+    # DELTA_SYNC without embedding_model_endpoint_name = self-managed vectors
+    inst = Databricks(
+        workspace_url="https://test",
+        access_token="tok",
+        endpoint_name="vs-endpoint",
+        catalog="catalog",
+        schema="schema",
+        table_name="table",
+        warehouse_name="test-warehouse",
+        index_type=VectorIndexType.DELTA_SYNC,
+        embedding_dimension=4,
+        # NOTE: no embedding_model_endpoint_name
+    )
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(data_array=[])
+    )
+    inst.search(query="ignored", vectors=[0.1, 0.2, 0.3, 0.4], limit=5)
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in call_kwargs
+    assert "query_text" not in call_kwargs
+
+
 def test_search_missing_params_raises(db_instance_delta):
     with pytest.raises(ValueError):
-        db_instance_delta.search(query="", vectors=[0.1, 0.2])  # DELTA_SYNC requires query text
+        db_instance_delta.search(query="", vectors=[0.1, 0.2])  # DELTA_SYNC with model endpoint requires query text
 
 
 # ---------------------- Delete Tests ---------------------- #
@@ -275,6 +304,54 @@ def test_get_vector(db_instance_delta, mock_workspace_client):
     assert res.id == "id-get"
     assert res.payload["data"] == "some memory"
     assert res.payload["tag"] == "x"
+    # DELTA_SYNC should use query_text, not query_vector
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_text" in call_kwargs
+    assert "query_vector" not in call_kwargs
+
+
+def test_get_vector_direct_access(db_instance_direct, mock_workspace_client):
+    """get() on a DIRECT_ACCESS index must use query_vector instead of query_text."""
+    mock_workspace_client.vector_search_indexes.query_index.return_value = QueryVectorIndexResponse(
+        manifest=ResultManifest(columns=[
+            ColumnInfo(name="memory_id"),
+            ColumnInfo(name="hash"),
+            ColumnInfo(name="agent_id"),
+            ColumnInfo(name="run_id"),
+            ColumnInfo(name="user_id"),
+            ColumnInfo(name="memory"),
+            ColumnInfo(name="metadata"),
+            ColumnInfo(name="created_at"),
+            ColumnInfo(name="updated_at"),
+            ColumnInfo(name="embedding"),
+            ColumnInfo(name="score"),
+        ]),
+        result=ResultData(
+            data_array=[
+                [
+                    "id-get-da",
+                    "h",
+                    "a",
+                    "r",
+                    "u",
+                    "direct access memory",
+                    '{"tag":"da"}',
+                    "2024-01-01T00:00:00",
+                    "2024-01-01T00:00:00",
+                    [0.1, 0.2, 0.3, 0.4],
+                    "0.88",
+                ]
+            ]
+        )
+    )
+    res = db_instance_direct.get("id-get-da")
+    assert res.id == "id-get-da"
+    assert res.payload["data"] == "direct access memory"
+    # DIRECT_ACCESS should use query_vector, not query_text
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in call_kwargs
+    assert "query_text" not in call_kwargs
+    assert call_kwargs["query_vector"] == [0.0] * 4  # embedding_dimension=4
 
 
 # ---------------------- Collection Info / Listing Tests ---------------------- #
@@ -330,6 +407,185 @@ def test_list_memories(db_instance_delta, mock_workspace_client):
     assert isinstance(res, list)
     assert len(res[0]) == 1
     assert res[0][0].id == "id-get"
+    # DELTA_SYNC should use query_text
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_text" in call_kwargs
+    assert "query_vector" not in call_kwargs
+
+
+def test_list_memories_direct_access(db_instance_direct, mock_workspace_client):
+    """list() on a DIRECT_ACCESS index must use query_vector instead of query_text."""
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(
+            data_array=[
+                [
+                    "id-da-list",
+                    "h",
+                    "a",
+                    "r",
+                    "u",
+                    "direct memory",
+                    None,
+                    "2024-01-01T00:00:00",
+                    "2024-01-01T00:00:00",
+                    [0.1, 0.2, 0.3, 0.4],
+                ]
+            ]
+        )
+    )
+    res = db_instance_direct.list(limit=5)
+    assert isinstance(res, list)
+    assert len(res[0]) == 1
+    assert res[0][0].id == "id-da-list"
+    assert res[0][0].payload["data"] == "direct memory"
+    # DIRECT_ACCESS should use query_vector
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in call_kwargs
+    assert "query_text" not in call_kwargs
+    assert call_kwargs["query_vector"] == [0.0] * 4
+
+
+def test_get_vector_delta_sync_self_managed(mock_workspace_client):
+    """get() on DELTA_SYNC without model endpoint should use query_vector."""
+    mock_workspace_client.tables.exists.return_value = SimpleNamespace(table_exists=True)
+    inst = Databricks(
+        workspace_url="https://test",
+        access_token="tok",
+        endpoint_name="vs-endpoint",
+        catalog="catalog",
+        schema="schema",
+        table_name="table",
+        warehouse_name="test-warehouse",
+        index_type=VectorIndexType.DELTA_SYNC,
+        embedding_dimension=4,
+        # NOTE: no embedding_model_endpoint_name
+    )
+    mock_workspace_client.vector_search_indexes.query_index.return_value = QueryVectorIndexResponse(
+        manifest=ResultManifest(columns=[
+            ColumnInfo(name="memory_id"), ColumnInfo(name="hash"),
+            ColumnInfo(name="agent_id"), ColumnInfo(name="run_id"),
+            ColumnInfo(name="user_id"), ColumnInfo(name="memory"),
+            ColumnInfo(name="metadata"), ColumnInfo(name="created_at"),
+            ColumnInfo(name="updated_at"),
+        ]),
+        result=ResultData(data_array=[["id-sm", "h", None, None, None, "self-managed mem", None, None, None]]),
+    )
+    res = inst.get("id-sm")
+    assert res.id == "id-sm"
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in call_kwargs
+    assert "query_text" not in call_kwargs
+    assert call_kwargs["query_vector"] == [0.0] * 4
+
+
+def test_list_memories_delta_sync_self_managed(mock_workspace_client):
+    """list() on DELTA_SYNC without model endpoint should use query_vector."""
+    mock_workspace_client.tables.exists.return_value = SimpleNamespace(table_exists=True)
+    inst = Databricks(
+        workspace_url="https://test",
+        access_token="tok",
+        endpoint_name="vs-endpoint",
+        catalog="catalog",
+        schema="schema",
+        table_name="table",
+        warehouse_name="test-warehouse",
+        index_type=VectorIndexType.DELTA_SYNC,
+        embedding_dimension=4,
+        # NOTE: no embedding_model_endpoint_name
+    )
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(data_array=[])
+    )
+    inst.list(limit=5)
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in call_kwargs
+    assert "query_text" not in call_kwargs
+
+
+def test_list_memories_default_limit(db_instance_delta, mock_workspace_client):
+    """list() with no limit should default to 100."""
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(data_array=[])
+    )
+    db_instance_delta.list(limit=None)
+    call_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert call_kwargs["num_results"] == 100
+
+
+# ---------------------- Table Creation Tests ---------------------- #
+
+
+def test_ensure_source_table_uses_dynamic_names(mock_workspace_client):
+    """Verify _ensure_source_table_exists uses self.fully_qualified_table_name and
+    self.table_name for the PK constraint, not hardcoded values."""
+    mock_workspace_client.tables.exists.return_value = SimpleNamespace(table_exists=False)
+    Databricks(
+        workspace_url="https://test",
+        access_token="tok",
+        endpoint_name="vs-endpoint",
+        catalog="my_catalog",
+        schema="my_schema",
+        table_name="my_memories",
+        collection_name="my_index",
+        warehouse_name="test-warehouse",
+        index_type=VectorIndexType.DELTA_SYNC,
+        embedding_model_endpoint_name="embedding-endpoint",
+    )
+    # _ensure_source_table_exists was called during __init__ via create_col
+    constraint_call = mock_workspace_client.table_constraints.create.call_args
+    assert constraint_call.kwargs["full_name_arg"] == "my_catalog.my_schema.my_memories"
+    pk_name = constraint_call.kwargs["constraint"].primary_key_constraint.name
+    assert pk_name == "pk_my_memories"
+
+
+# ---------------------- Config Validation Tests ---------------------- #
+
+
+def test_config_rejects_old_doc_params():
+    """Config should reject the old documentation parameter names like index_name and source_table_name."""
+    from mem0.configs.vector_stores.databricks import DatabricksConfig
+    with pytest.raises(ValueError, match="Extra fields not allowed"):
+        DatabricksConfig(
+            workspace_url="https://test",
+            access_token="tok",
+            endpoint_name="ep",
+            catalog="cat",
+            schema="sch",
+            table_name="tbl",
+            index_name="catalog.schema.index",  # old param from docs
+        )
+
+
+def test_config_rejects_source_table_name():
+    """Config should reject source_table_name which was in old docs."""
+    from mem0.configs.vector_stores.databricks import DatabricksConfig
+    with pytest.raises(ValueError, match="Extra fields not allowed"):
+        DatabricksConfig(
+            workspace_url="https://test",
+            access_token="tok",
+            endpoint_name="ep",
+            catalog="cat",
+            schema="sch",
+            table_name="tbl",
+            source_table_name="catalog.schema.table",  # old param from docs
+        )
+
+
+def test_config_accepts_correct_params():
+    """Config should accept all the correct parameter names."""
+    from mem0.configs.vector_stores.databricks import DatabricksConfig
+    config = DatabricksConfig(
+        workspace_url="https://test",
+        access_token="tok",
+        endpoint_name="ep",
+        catalog="cat",
+        schema="sch",
+        table_name="tbl",
+        collection_name="my_index",
+        embedding_dimension=768,
+    )
+    assert config.collection_name == "my_index"
+    assert config.embedding_dimension == 768
 
 
 # ---------------------- Reset Tests ---------------------- #
@@ -341,3 +597,244 @@ def test_reset(db_instance_delta, mock_workspace_client):
     with patch.object(db_instance_delta, "create_col", wraps=db_instance_delta.create_col) as create_spy:
         db_instance_delta.reset()
         assert create_spy.called
+
+
+# ---------------------- End-to-End Config → Factory → CRUD Tests ---------------------- #
+
+
+def test_e2e_config_to_factory_delta_sync(mock_workspace_client):
+    """End-to-end: VectorStoreConfig validates docs-correct params, factory creates Databricks instance."""
+    from mem0.vector_stores.configs import VectorStoreConfig
+    from mem0.utils.factory import VectorStoreFactory
+
+    # Step 1: Config validation (simulates what Memory.from_config does)
+    vs_config = VectorStoreConfig(
+        provider="databricks",
+        config={
+            "workspace_url": "https://my-workspace.databricks.com",
+            "access_token": "my-token",
+            "endpoint_name": "my-endpoint",
+            "catalog": "prod_catalog",
+            "schema": "ai_schema",
+            "table_name": "memories_table",
+            "collection_name": "my_index",
+            "embedding_dimension": 768,
+            "warehouse_name": "test-warehouse",
+        },
+    )
+    assert vs_config.config.collection_name == "my_index"
+    assert vs_config.config.catalog == "prod_catalog"
+
+    # Step 2: Factory instantiation (same as MemoryBase.__init__)
+    instance = VectorStoreFactory.create("databricks", vs_config.config)
+    assert isinstance(instance, Databricks)
+    assert instance.fully_qualified_table_name == "prod_catalog.ai_schema.memories_table"
+    assert instance.fully_qualified_index_name == "prod_catalog.ai_schema.my_index"
+    assert instance.embedding_dimension == 768
+
+
+def test_e2e_config_to_factory_direct_access(mock_workspace_client):
+    """End-to-end: DIRECT_ACCESS via config → factory creates correct instance."""
+    from mem0.vector_stores.configs import VectorStoreConfig
+    from mem0.utils.factory import VectorStoreFactory
+
+    mock_workspace_client.tables.exists.return_value = SimpleNamespace(table_exists=True)
+
+    vs_config = VectorStoreConfig(
+        provider="databricks",
+        config={
+            "workspace_url": "https://my-workspace.databricks.com",
+            "access_token": "my-token",
+            "endpoint_name": "my-endpoint",
+            "catalog": "cat",
+            "schema": "sch",
+            "table_name": "tbl",
+            "index_type": "DIRECT_ACCESS",
+            "embedding_dimension": 4,
+            "warehouse_name": "test-warehouse",
+        },
+    )
+    instance = VectorStoreFactory.create("databricks", vs_config.config)
+    assert isinstance(instance, Databricks)
+    assert "embedding" in instance.column_names
+
+
+def test_e2e_old_docs_config_rejected():
+    """End-to-end: Config from old docs (with index_name, source_table_name) is rejected at validation."""
+    from mem0.vector_stores.configs import VectorStoreConfig
+
+    with pytest.raises(ValueError, match="Extra fields not allowed"):
+        VectorStoreConfig(
+            provider="databricks",
+            config={
+                "workspace_url": "https://my-workspace.databricks.com",
+                "access_token": "my-token",
+                "endpoint_name": "my-endpoint",
+                "index_name": "catalog.schema.index_name",
+                "source_table_name": "catalog.schema.source_table",
+                "embedding_dimension": 1536,
+            },
+        )
+
+
+def test_e2e_crud_lifecycle_delta_sync(mock_workspace_client):
+    """End-to-end CRUD lifecycle: insert → search → get → list → update → delete."""
+    from mem0.vector_stores.configs import VectorStoreConfig
+    from mem0.utils.factory import VectorStoreFactory
+
+    vs_config = VectorStoreConfig(
+        provider="databricks",
+        config={
+            "workspace_url": "https://test",
+            "access_token": "tok",
+            "endpoint_name": "ep",
+            "catalog": "cat",
+            "schema": "sch",
+            "table_name": "tbl",
+            "warehouse_name": "test-warehouse",
+            "embedding_model_endpoint_name": "emb-ep",
+        },
+    )
+    db = VectorStoreFactory.create("databricks", vs_config.config)
+
+    # INSERT
+    db.insert(
+        vectors=[[0.1, 0.2]],
+        payloads=[{"data": "test memory", "user_id": "u1", "hash": "h1"}],
+        ids=["mem-001"],
+    )
+    insert_sql = mock_workspace_client.statement_execution.execute_statement.call_args.kwargs["statement"]
+    assert "INSERT INTO cat.sch.tbl" in insert_sql
+    assert "mem-001" in insert_sql
+
+    # SEARCH
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(
+            data_array=[["mem-001", "h1", None, None, "u1", "test memory", None, None, None, 0.95]]
+        )
+    )
+    results = db.search(query="test", vectors=None, limit=5)
+    assert len(results) == 1
+    assert results[0].id == "mem-001"
+    assert results[0].payload["data"] == "test memory"
+    search_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert search_kwargs["query_text"] == "test"
+
+    # GET
+    mock_workspace_client.vector_search_indexes.query_index.return_value = QueryVectorIndexResponse(
+        manifest=ResultManifest(columns=[
+            ColumnInfo(name="memory_id"), ColumnInfo(name="hash"),
+            ColumnInfo(name="agent_id"), ColumnInfo(name="run_id"),
+            ColumnInfo(name="user_id"), ColumnInfo(name="memory"),
+            ColumnInfo(name="metadata"), ColumnInfo(name="created_at"),
+            ColumnInfo(name="updated_at"),
+        ]),
+        result=ResultData(data_array=[["mem-001", "h1", None, None, "u1", "test memory", None, None, None]]),
+    )
+    got = db.get("mem-001")
+    assert got.id == "mem-001"
+    assert got.payload["data"] == "test memory"
+    get_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_text" in get_kwargs
+    assert "query_vector" not in get_kwargs
+
+    # LIST
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(
+            data_array=[["mem-001", "h1", None, None, "u1", "test memory", None, None, None]]
+        )
+    )
+    listed = db.list(filters={"user_id": "u1"}, limit=10)
+    assert len(listed[0]) == 1
+    assert listed[0][0].id == "mem-001"
+    list_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_text" in list_kwargs
+    assert list_kwargs["num_results"] == 10
+
+    # UPDATE
+    db.update(vector_id="mem-001", payload={"memory": "updated memory"})
+    update_sql = mock_workspace_client.statement_execution.execute_statement.call_args.kwargs["statement"]
+    assert "UPDATE cat.sch.tbl" in update_sql
+    assert "mem-001" in update_sql
+    assert "updated memory" in update_sql
+
+    # DELETE
+    db.delete("mem-001")
+    delete_sql = mock_workspace_client.statement_execution.execute_statement.call_args.kwargs["statement"]
+    assert "DELETE FROM cat.sch.tbl" in delete_sql
+    assert "mem-001" in delete_sql
+
+
+def test_e2e_crud_lifecycle_direct_access(mock_workspace_client):
+    """End-to-end CRUD lifecycle for DIRECT_ACCESS: insert → search → get → list."""
+    from mem0.vector_stores.configs import VectorStoreConfig
+    from mem0.utils.factory import VectorStoreFactory
+
+    mock_workspace_client.tables.exists.return_value = SimpleNamespace(table_exists=True)
+
+    vs_config = VectorStoreConfig(
+        provider="databricks",
+        config={
+            "workspace_url": "https://test",
+            "access_token": "tok",
+            "endpoint_name": "ep",
+            "catalog": "cat",
+            "schema": "sch",
+            "table_name": "tbl",
+            "index_type": "DIRECT_ACCESS",
+            "embedding_dimension": 4,
+            "warehouse_name": "test-warehouse",
+            "embedding_model_endpoint_name": "emb-ep",
+        },
+    )
+    db = VectorStoreFactory.create("databricks", vs_config.config)
+    assert "embedding" in db.column_names
+
+    # INSERT with vector
+    db.insert(
+        vectors=[[0.1, 0.2, 0.3, 0.4]],
+        payloads=[{"data": "direct memory", "user_id": "u1", "hash": "h1"}],
+        ids=["mem-da-001"],
+    )
+    insert_sql = mock_workspace_client.statement_execution.execute_statement.call_args.kwargs["statement"]
+    assert "array(0.1, 0.2, 0.3, 0.4)" in insert_sql
+
+    # SEARCH with vector
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(
+            data_array=[["mem-da-001", "h1", None, None, "u1", "direct memory", None, None, None, [0.1, 0.2, 0.3, 0.4], 0.9]]
+        )
+    )
+    results = db.search(query="", vectors=[0.1, 0.2, 0.3, 0.4], limit=5)
+    assert len(results) == 1
+    search_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in search_kwargs
+    assert "query_text" not in search_kwargs
+
+    # GET — must use query_vector for DIRECT_ACCESS
+    mock_workspace_client.vector_search_indexes.query_index.return_value = QueryVectorIndexResponse(
+        manifest=ResultManifest(columns=[
+            ColumnInfo(name="memory_id"), ColumnInfo(name="hash"),
+            ColumnInfo(name="agent_id"), ColumnInfo(name="run_id"),
+            ColumnInfo(name="user_id"), ColumnInfo(name="memory"),
+            ColumnInfo(name="metadata"), ColumnInfo(name="created_at"),
+            ColumnInfo(name="updated_at"), ColumnInfo(name="embedding"),
+        ]),
+        result=ResultData(data_array=[["mem-da-001", "h1", None, None, "u1", "direct memory", None, None, None, [0.1, 0.2, 0.3, 0.4]]]),
+    )
+    got = db.get("mem-da-001")
+    assert got.id == "mem-da-001"
+    get_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in get_kwargs
+    assert get_kwargs["query_vector"] == [0.0] * 4
+
+    # LIST — must use query_vector for DIRECT_ACCESS
+    mock_workspace_client.vector_search_indexes.query_index.return_value = SimpleNamespace(
+        result=SimpleNamespace(
+            data_array=[["mem-da-001", "h1", None, None, "u1", "direct memory", None, None, None, [0.1, 0.2, 0.3, 0.4]]]
+        )
+    )
+    db.list(limit=5)
+    list_kwargs = mock_workspace_client.vector_search_indexes.query_index.call_args.kwargs
+    assert "query_vector" in list_kwargs
+    assert "query_text" not in list_kwargs

tests/vector_stores/test_mongodb.py

@@ -48,17 +48,16 @@ def test_initalize_create_col(mongo_vector_fixture):
     search_index_model = args[0].document
     assert search_index_model == {
         "name": "test_collection_vector_index",
+        "type": "vectorSearch",
         "definition": {
-            "mappings": {
-                "dynamic": False,
-                "fields": {
-                    "embedding": {
-                        "type": "knnVector",
-                        "dimensions": 1536,
-                        "similarity": "cosine",
-                    }
-                },
-            }
+            "fields": [
+                {
+                    "type": "vector",
+                    "path": "embedding",
+                    "numDimensions": 1536,
+                    "similarity": "cosine",
+                }
+            ]
         },
     }
     assert mongo_vector.collection == mock_collection
@@ -95,7 +94,7 @@ def test_search(mongo_vector_fixture):
                 "$vectorSearch": {
                     "index": "test_collection_vector_index",
                     "limit": 2,
-                    "numCandidates": 2,
+                    "numCandidates": 40,
                     "queryVector": query_vector,
                     "path": "embedding",
                 },
@@ -204,6 +203,10 @@ def test_delete(mongo_vector_fixture):
 
 
 def test_update(mongo_vector_fixture):
+    """
+    Test that update() uses dot notation for payload fields instead of replacing
+    the entire payload document.
+    """
     mongo_vector, mock_collection, _ = mongo_vector_fixture
     vector_id = "id1"
     updated_vector = [0.3] * 1536
@@ -213,11 +216,51 @@ def test_update(mongo_vector_fixture):
 
     mongo_vector.update(vector_id=vector_id, vector=updated_vector, payload=updated_payload)
 
+    # Should use dot notation (payload.name) instead of full replacement (payload)
     mock_collection.update_one.assert_called_once_with(
-        {"_id": vector_id}, {"$set": {"embedding": updated_vector, "payload": updated_payload}}
+        {"_id": vector_id}, {"$set": {"embedding": updated_vector, "payload.name": "updated_vector"}}
     )
 
 
+def test_update_payload_only_uses_dot_notation(mongo_vector_fixture):
+    """
+    Test that updating only the payload uses dot notation for each field,
+    preserving existing metadata fields not included in the update.
+    """
+    mongo_vector, mock_collection, _ = mongo_vector_fixture
+
+    mock_collection.update_one.return_value = MagicMock(matched_count=1)
+
+    mongo_vector.update(
+        vector_id="id1",
+        payload={"data": "updated text", "hash": "def456", "updated_at": "2025-06-01"},
+    )
+
+    set_arg = mock_collection.update_one.call_args[0][1]["$set"]
+
+    # Only the specified fields should be in $set, using dot notation
+    assert set_arg == {
+        "payload.data": "updated text",
+        "payload.hash": "def456",
+        "payload.updated_at": "2025-06-01",
+    }
+    # "payload" key itself should NOT appear (that would replace the whole document)
+    assert "payload" not in set_arg
+
+
+def test_update_vector_only_does_not_touch_payload(mongo_vector_fixture):
+    """Test that updating only the vector does not touch any payload fields."""
+    mongo_vector, mock_collection, _ = mongo_vector_fixture
+
+    mock_collection.update_one.return_value = MagicMock(matched_count=1)
+
+    mongo_vector.update(vector_id="id1", vector=[0.7, 0.8, 0.9])
+
+    set_arg = mock_collection.update_one.call_args[0][1]["$set"]
+    assert set_arg == {"embedding": [0.7, 0.8, 0.9]}
+    assert not any(k.startswith("payload.") for k in set_arg)
+
+
 def test_get(mongo_vector_fixture):
     mongo_vector, mock_collection, _ = mongo_vector_fixture
     vector_id = "id1"

tests/vector_stores/test_opensearch.py

@@ -18,25 +18,25 @@ from mem0.vector_stores.opensearch import OpenSearchDB
 # Mock classes for testing OpenSearch with AWS authentication
 class MockFieldInfo:
     """Mock pydantic field info."""
+
     def __init__(self, default=None):
         self.default = default
 
 
 class MockOpenSearchConfig:
-    
     model_fields = {
-        'collection_name': MockFieldInfo(default="default_collection"),
-        'host': MockFieldInfo(default="localhost"),
-        'port': MockFieldInfo(default=9200),
-        'embedding_model_dims': MockFieldInfo(default=1536),
-        'http_auth': MockFieldInfo(default=None),
-        'auth': MockFieldInfo(default=None),
-        'credentials': MockFieldInfo(default=None),
-        'connection_class': MockFieldInfo(default=None),
-        'use_ssl': MockFieldInfo(default=False),
-        'verify_certs': MockFieldInfo(default=False),
+        "collection_name": MockFieldInfo(default="default_collection"),
+        "host": MockFieldInfo(default="localhost"),
+        "port": MockFieldInfo(default=9200),
+        "embedding_model_dims": MockFieldInfo(default=1536),
+        "http_auth": MockFieldInfo(default=None),
+        "auth": MockFieldInfo(default=None),
+        "credentials": MockFieldInfo(default=None),
+        "connection_class": MockFieldInfo(default=None),
+        "use_ssl": MockFieldInfo(default=False),
+        "verify_certs": MockFieldInfo(default=False),
     }
-    
+
     def __init__(self, collection_name="test_collection", include_auth=True, **kwargs):
         self.collection_name = collection_name
         self.host = kwargs.get("host", "localhost")
@@ -44,7 +44,7 @@ class MockOpenSearchConfig:
         self.embedding_model_dims = kwargs.get("embedding_model_dims", 1536)
         self.use_ssl = kwargs.get("use_ssl", True)
         self.verify_certs = kwargs.get("verify_certs", True)
-        
+
         if any(field in kwargs for field in ["http_auth", "auth", "credentials", "connection_class"]):
             self.http_auth = kwargs.get("http_auth")
             self.auth = kwargs.get("auth")
@@ -63,20 +63,18 @@ class MockOpenSearchConfig:
 
 
 class MockAWSAuth:
-    
     def __init__(self):
         self._lock = threading.Lock()
         self.region = "us-east-1"
-    
+
     def __deepcopy__(self, memo):
         raise TypeError("cannot pickle '_thread.lock' object")
 
 
 class MockConnectionClass:
-    
     def __init__(self):
         self._state = {"connected": False}
-    
+
     def __deepcopy__(self, memo):
         raise TypeError("cannot pickle connection state")
 
@@ -255,6 +253,104 @@ class TestOpenSearchDB(unittest.TestCase):
         self.os_db.delete_col()
         self.client_mock.indices.delete.assert_called_once_with(index="test_collection")
 
+    def test_insert_rejects_null_vectors(self):
+        """Vectors that are None should raise ValueError before hitting OpenSearch."""
+        vectors = [None]
+        payloads = [{"key": "value"}]
+        ids = ["id1"]
+
+        with self.assertRaises(ValueError) as ctx:
+            self.os_db.insert(vectors=vectors, payloads=payloads, ids=ids)
+
+        self.assertIn("null", str(ctx.exception).lower())
+        self.client_mock.index.assert_not_called()
+
+    def test_insert_rejects_empty_vectors(self):
+        """Empty vector lists should raise ValueError."""
+        vectors = [[]]
+        payloads = [{"key": "value"}]
+        ids = ["id1"]
+
+        with self.assertRaises(ValueError) as ctx:
+            self.os_db.insert(vectors=vectors, payloads=payloads, ids=ids)
+
+        self.assertIn("empty", str(ctx.exception).lower())
+        self.client_mock.index.assert_not_called()
+
+    def test_insert_rejects_dimension_mismatch(self):
+        """Vectors with wrong dimensions should raise ValueError with a clear message."""
+        vectors = [[0.1] * 768]
+        payloads = [{"key": "value"}]
+        ids = ["id1"]
+
+        with self.assertRaises(ValueError) as ctx:
+            self.os_db.insert(vectors=vectors, payloads=payloads, ids=ids)
+
+        error_msg = str(ctx.exception)
+        self.assertIn("768", error_msg)
+        self.assertIn("1536", error_msg)
+        self.client_mock.index.assert_not_called()
+
+    def test_update_rejects_empty_vector(self):
+        """Update with an empty vector should raise ValueError."""
+        with self.assertRaises(ValueError) as ctx:
+            self.os_db.update("id1", vector=[], payload={"key": "value"})
+
+        self.assertIn("empty", str(ctx.exception).lower())
+        self.client_mock.search.assert_not_called()
+
+    def test_update_rejects_dimension_mismatch(self):
+        """Update with wrong vector dimensions should raise ValueError."""
+        vector = [0.1] * 768
+        payload = {"key": "value"}
+
+        with self.assertRaises(ValueError) as ctx:
+            self.os_db.update("id1", vector=vector, payload=payload)
+
+        error_msg = str(ctx.exception)
+        self.assertIn("768", error_msg)
+        self.assertIn("1536", error_msg)
+        self.client_mock.search.assert_not_called()
+
+    @patch("mem0.vector_stores.opensearch.logger")
+    def test_update_error_logs_with_exc_info(self, mock_logger):
+        """Update errors should log with exc_info and re-raise."""
+        mock_search_response = {"hits": {"hits": [{"_id": "doc1", "_source": {"id": "id1"}}]}}
+        self.client_mock.search.return_value = mock_search_response
+        self.client_mock.update.side_effect = Exception("Update failed")
+
+        with self.assertRaises(Exception):
+            self.os_db.update("id1", vector=[0.1] * 1536, payload={"key": "value"})
+
+        mock_logger.error.assert_called_once()
+        call_kwargs = mock_logger.error.call_args
+        self.assertTrue(call_kwargs[1].get("exc_info"), "logger.error must be called with exc_info=True")
+
+    @patch("mem0.vector_stores.opensearch.logger")
+    def test_insert_error_logs_with_exc_info(self, mock_logger):
+        """Error logging should include exc_info for full stack trace."""
+        vectors = [[0.1] * 1536]
+        payloads = [{"key": "value"}]
+        ids = ["id1"]
+        self.client_mock.index.side_effect = Exception("Connection refused")
+
+        with self.assertRaises(Exception):
+            self.os_db.insert(vectors=vectors, payloads=payloads, ids=ids)
+
+        mock_logger.error.assert_called_once()
+        call_kwargs = mock_logger.error.call_args
+        self.assertTrue(call_kwargs[1].get("exc_info"), "logger.error must be called with exc_info=True")
+
+    @patch("mem0.vector_stores.opensearch.logger")
+    def test_search_error_logs_with_exc_info(self, mock_logger):
+        """Search error logging should include exc_info for full stack trace."""
+        self.client_mock.search.side_effect = Exception("Search failed")
+        results = self.os_db.search(query="", vectors=[[0.1] * 1536], limit=5)
+        self.assertEqual(results, [])
+        mock_logger.error.assert_called_once()
+        call_kwargs = mock_logger.error.call_args
+        self.assertTrue(call_kwargs[1].get("exc_info"), "logger.error must be called with exc_info=True")
+
     def test_init_with_http_auth(self):
         mock_credentials = MagicMock()
         mock_signer = AWSV4SignerAuth(mock_credentials, "us-east-1", "es")
@@ -282,11 +378,13 @@ class TestOpenSearchDB(unittest.TestCase):
 
 
 # Tests for OpenSearch config deepcopy with AWS authentication (Issue #3464)
-@patch('mem0.utils.factory.EmbedderFactory.create')
-@patch('mem0.utils.factory.VectorStoreFactory.create')
-@patch('mem0.utils.factory.LlmFactory.create')
-@patch('mem0.memory.storage.SQLiteManager')
-def test_safe_deepcopy_config_handles_opensearch_auth(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_safe_deepcopy_config_handles_opensearch_auth(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
     """Test that _safe_deepcopy_config handles OpenSearch configs with AWS auth objects gracefully."""
     mock_embedder_factory.return_value = MagicMock()
     mock_vector_store = MagicMock()
@@ -295,9 +393,9 @@ def test_safe_deepcopy_config_handles_opensearch_auth(mock_sqlite, mock_llm_fact
     mock_sqlite.return_value = MagicMock()
 
     from mem0.memory.main import _safe_deepcopy_config
-    
+
     config_with_auth = MockOpenSearchConfig(collection_name="opensearch_test", include_auth=True)
-    
+
     safe_config = _safe_deepcopy_config(config_with_auth)
 
     # Runtime auth objects must be preserved (Issue #3580)
@@ -306,7 +404,7 @@ def test_safe_deepcopy_config_handles_opensearch_auth(mock_sqlite, mock_llm_fact
     assert safe_config.connection_class is not None
     # Credentials dict is a sensitive secret and should be redacted
     assert safe_config.credentials is None
-    
+
     assert safe_config.collection_name == "opensearch_test"
     assert safe_config.host == "localhost"
     assert safe_config.port == 9200
@@ -315,10 +413,10 @@ def test_safe_deepcopy_config_handles_opensearch_auth(mock_sqlite, mock_llm_fact
     assert safe_config.verify_certs is True
 
 
-@patch('mem0.utils.factory.EmbedderFactory.create')
-@patch('mem0.utils.factory.VectorStoreFactory.create') 
-@patch('mem0.utils.factory.LlmFactory.create')
-@patch('mem0.memory.storage.SQLiteManager')
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
 def test_safe_deepcopy_config_normal_configs(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
     """Test that _safe_deepcopy_config handles normal OpenSearch configs without auth."""
     mock_embedder_factory.return_value = MagicMock()
@@ -328,12 +426,12 @@ def test_safe_deepcopy_config_normal_configs(mock_sqlite, mock_llm_factory, mock
     mock_sqlite.return_value = MagicMock()
 
     from mem0.memory.main import _safe_deepcopy_config
-    
+
     config_without_auth = MockOpenSearchConfig(collection_name="normal_test", include_auth=False)
-    
+
     safe_config = _safe_deepcopy_config(config_without_auth)
-    
-    assert safe_config.collection_name == "normal_test" 
+
+    assert safe_config.collection_name == "normal_test"
     assert safe_config.host == "localhost"
     assert safe_config.port == 9200
     assert safe_config.embedding_model_dims == 1536
@@ -341,13 +439,15 @@ def test_safe_deepcopy_config_normal_configs(mock_sqlite, mock_llm_factory, mock
     assert safe_config.verify_certs is True
 
 
-@patch('mem0.utils.factory.EmbedderFactory.create')
-@patch('mem0.utils.factory.VectorStoreFactory.create')
-@patch('mem0.utils.factory.LlmFactory.create')
-@patch('mem0.memory.storage.SQLiteManager')
-def test_memory_initialization_opensearch_aws_auth(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+@patch("mem0.utils.factory.EmbedderFactory.create")
+@patch("mem0.utils.factory.VectorStoreFactory.create")
+@patch("mem0.utils.factory.LlmFactory.create")
+@patch("mem0.memory.storage.SQLiteManager")
+def test_memory_initialization_opensearch_aws_auth(
+    mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory
+):
     """Test that Memory initialization works with OpenSearch configs containing AWS auth."""
-    
+
     mock_embedder_factory.return_value = MagicMock()
     mock_vector_store = MagicMock()
     mock_vector_factory.return_value = mock_vector_store

tests/vector_stores/test_qdrant.py

@@ -1,13 +1,21 @@
+import os
+import tempfile
 import unittest
 import uuid
-from unittest.mock import MagicMock
+from unittest.mock import MagicMock, patch
 
 from qdrant_client import QdrantClient
 from qdrant_client.models import (
     Distance,
+    FieldCondition,
     Filter,
+    MatchAny,
+    MatchExcept,
+    MatchText,
+    MatchValue,
     PointIdsList,
     PointStruct,
+    Range,
     VectorParams,
 )
 
@@ -25,6 +33,23 @@ class TestQdrant(unittest.TestCase):
             on_disk=True,
         )
 
+    def test_local_path_on_disk_false_preserves_existing_directory(self):
+        """#4473: local path must not be removed when on_disk is False."""
+        with tempfile.TemporaryDirectory() as tmp:
+            sentinel = os.path.join(tmp, "sentinel")
+            with open(sentinel, "w", encoding="utf-8") as f:
+                f.write("keep")
+            mock_client = MagicMock()
+            mock_client.get_collections.return_value = MagicMock(collections=[])
+            with patch("mem0.vector_stores.qdrant.QdrantClient", return_value=mock_client):
+                Qdrant(
+                    collection_name="c",
+                    embedding_model_dims=128,
+                    path=tmp,
+                    on_disk=False,
+                )
+            self.assertTrue(os.path.isfile(sentinel))
+
     def test_create_col(self):
         self.client_mock.get_collections.return_value = MagicMock(collections=[])
 
@@ -298,3 +323,456 @@ class TestQdrant(unittest.TestCase):
 
     def tearDown(self):
         del self.qdrant
+
+
+class TestQdrantEnhancedFilters(unittest.TestCase):
+    """Tests for enhanced metadata filtering operators (issue #3975)."""
+
+    def setUp(self):
+        self.client_mock = MagicMock(spec=QdrantClient)
+        self.qdrant = Qdrant(
+            collection_name="test_collection",
+            embedding_model_dims=128,
+            client=self.client_mock,
+        )
+
+    # ------------------------------------------------------------------ #
+    # _build_field_condition                                               #
+    # ------------------------------------------------------------------ #
+
+    def test_simple_equality(self):
+        """Plain value maps to MatchValue."""
+        cond = self.qdrant._build_field_condition("category", "programming")
+        self.assertIsInstance(cond, FieldCondition)
+        self.assertEqual(cond.key, "category")
+        self.assertIsInstance(cond.match, MatchValue)
+        self.assertEqual(cond.match.value, "programming")
+
+    def test_eq_operator(self):
+        """{\"eq\": v} maps to MatchValue."""
+        cond = self.qdrant._build_field_condition("category", {"eq": "programming"})
+        self.assertIsInstance(cond.match, MatchValue)
+        self.assertEqual(cond.match.value, "programming")
+
+    def test_ne_operator(self):
+        """{\"ne\": v} maps to MatchExcept with a single-element list."""
+        cond = self.qdrant._build_field_condition("category", {"ne": "spam"})
+        self.assertIsInstance(cond.match, MatchExcept)
+        self.assertIn("spam", getattr(cond.match, "except_", None) or getattr(cond.match, "except"))
+
+    def test_in_operator(self):
+        """{\"in\": [...]} maps to MatchAny."""
+        cond = self.qdrant._build_field_condition("category", {"in": ["a", "b", "c"]})
+        self.assertIsInstance(cond.match, MatchAny)
+        self.assertEqual(cond.match.any, ["a", "b", "c"])
+
+    def test_nin_operator(self):
+        """{\"nin\": [...]} maps to MatchExcept."""
+        cond = self.qdrant._build_field_condition("status", {"nin": ["deleted", "banned"]})
+        self.assertIsInstance(cond.match, MatchExcept)
+        excluded = getattr(cond.match, "except_", None) or getattr(cond.match, "except")
+        self.assertIn("deleted", excluded)
+        self.assertIn("banned", excluded)
+
+    def test_contains_operator(self):
+        """{\"contains\": x} maps to MatchText."""
+        cond = self.qdrant._build_field_condition("bio", {"contains": "python"})
+        self.assertIsInstance(cond.match, MatchText)
+        self.assertEqual(cond.match.text, "python")
+
+    def test_icontains_operator(self):
+        """{\"icontains\": x} maps to MatchText (same as contains for Qdrant)."""
+        cond = self.qdrant._build_field_condition("bio", {"icontains": "Python"})
+        self.assertIsInstance(cond.match, MatchText)
+        self.assertEqual(cond.match.text, "Python")
+
+    def test_gt_operator(self):
+        """{\"gt\": v} maps to Range with gt only."""
+        cond = self.qdrant._build_field_condition("priority", {"gt": 5})
+        self.assertIsInstance(cond.range, Range)
+        self.assertEqual(cond.range.gt, 5)
+        self.assertIsNone(cond.range.gte)
+        self.assertIsNone(cond.range.lt)
+        self.assertIsNone(cond.range.lte)
+
+    def test_gte_operator(self):
+        """{\"gte\": v} maps to Range with gte only."""
+        cond = self.qdrant._build_field_condition("priority", {"gte": 5})
+        self.assertIsInstance(cond.range, Range)
+        self.assertEqual(cond.range.gte, 5)
+        self.assertIsNone(cond.range.gt)
+
+    def test_lt_operator(self):
+        """{\"lt\": v} maps to Range with lt only."""
+        cond = self.qdrant._build_field_condition("priority", {"lt": 10})
+        self.assertIsInstance(cond.range, Range)
+        self.assertEqual(cond.range.lt, 10)
+
+    def test_lte_operator(self):
+        """{\"lte\": v} maps to Range with lte only."""
+        cond = self.qdrant._build_field_condition("priority", {"lte": 10})
+        self.assertIsInstance(cond.range, Range)
+        self.assertEqual(cond.range.lte, 10)
+
+    def test_range_gte_lte(self):
+        """Combined {\"gte\": x, \"lte\": y} maps to Range with both bounds."""
+        cond = self.qdrant._build_field_condition("priority", {"gte": 5, "lte": 10})
+        self.assertIsInstance(cond.range, Range)
+        self.assertEqual(cond.range.gte, 5)
+        self.assertEqual(cond.range.lte, 10)
+
+    def test_range_gt_lt(self):
+        """Open interval {\"gt\": x, \"lt\": y} maps to Range with both bounds."""
+        cond = self.qdrant._build_field_condition("score", {"gt": 0.5, "lt": 0.9})
+        self.assertIsInstance(cond.range, Range)
+        self.assertEqual(cond.range.gt, 0.5)
+        self.assertEqual(cond.range.lt, 0.9)
+
+    # ------------------------------------------------------------------ #
+    # _create_filter — comparison and list operators                       #
+    # ------------------------------------------------------------------ #
+
+    def test_create_filter_eq(self):
+        result = self.qdrant._create_filter({"category": {"eq": "programming"}})
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must), 1)
+        self.assertIsInstance(result.must[0].match, MatchValue)
+
+    def test_create_filter_ne(self):
+        result = self.qdrant._create_filter({"category": {"ne": "spam"}})
+        self.assertIsInstance(result, Filter)
+        self.assertIsInstance(result.must[0].match, MatchExcept)
+
+    def test_create_filter_in(self):
+        result = self.qdrant._create_filter({"category": {"in": ["prog", "data"]}})
+        self.assertIsInstance(result, Filter)
+        self.assertIsInstance(result.must[0].match, MatchAny)
+        self.assertEqual(result.must[0].match.any, ["prog", "data"])
+
+    def test_create_filter_nin(self):
+        result = self.qdrant._create_filter({"status": {"nin": ["deleted", "banned"]}})
+        self.assertIsInstance(result, Filter)
+        self.assertIsInstance(result.must[0].match, MatchExcept)
+
+    def test_create_filter_gt(self):
+        result = self.qdrant._create_filter({"priority": {"gt": 7}})
+        self.assertIsInstance(result, Filter)
+        self.assertIsInstance(result.must[0].range, Range)
+        self.assertEqual(result.must[0].range.gt, 7)
+
+    def test_create_filter_gte_lte_range(self):
+        result = self.qdrant._create_filter({"priority": {"gte": 5, "lte": 9}})
+        self.assertIsInstance(result, Filter)
+        r = result.must[0].range
+        self.assertEqual(r.gte, 5)
+        self.assertEqual(r.lte, 9)
+
+    def test_create_filter_contains(self):
+        result = self.qdrant._create_filter({"bio": {"contains": "python"}})
+        self.assertIsInstance(result, Filter)
+        self.assertIsInstance(result.must[0].match, MatchText)
+
+    # ------------------------------------------------------------------ #
+    # _create_filter — logical operators                                   #
+    # ------------------------------------------------------------------ #
+
+    def test_and_operator(self):
+        """AND populates Filter.must with nested Filter objects."""
+        filters = {"AND": [{"category": "programming"}, {"priority": 10}]}
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must), 2)
+        # Each item is a Filter that wraps a single FieldCondition
+        for item in result.must:
+            self.assertIsInstance(item, Filter)
+
+    def test_or_operator(self):
+        """OR populates Filter.should with nested Filter objects."""
+        filters = {"OR": [{"category": "programming"}, {"category": "data"}]}
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.should), 2)
+        self.assertIsNone(result.must)
+
+    def test_not_operator(self):
+        """NOT populates Filter.must_not with nested Filter objects."""
+        filters = {"NOT": [{"category": "spam"}]}
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must_not), 1)
+        self.assertIsNone(result.must)
+        self.assertIsNone(result.should)
+
+    def test_mixed_field_and_and(self):
+        """Top-level field conditions and AND can coexist in must."""
+        filters = {
+            "user_id": "alice",
+            "AND": [{"priority": {"gte": 5}}, {"category": "programming"}],
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        # user_id FieldCondition + 2 nested AND Filters = 3 items in must
+        self.assertEqual(len(result.must), 3)
+
+    def test_nested_and_or(self):
+        """AND containing an OR sub-condition produces correct nesting."""
+        filters = {
+            "AND": [
+                {"OR": [{"category": "prog"}, {"category": "data"}]},
+                {"priority": {"gt": 3}},
+            ]
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must), 2)
+        # First item is a Filter with should (the OR)
+        or_filter = result.must[0]
+        self.assertIsInstance(or_filter, Filter)
+        self.assertEqual(len(or_filter.should), 2)
+
+    # ------------------------------------------------------------------ #
+    # Edge cases                                                           #
+    # ------------------------------------------------------------------ #
+
+    def test_empty_filter_returns_none(self):
+        self.assertIsNone(self.qdrant._create_filter({}))
+        self.assertIsNone(self.qdrant._create_filter(None))
+
+    def test_backward_compat_simple_equality(self):
+        """Plain scalar equality still works as before."""
+        result = self.qdrant._create_filter({"user_id": "alice", "run_id": "r1"})
+        self.assertIsInstance(result, Filter)
+        keys = [c.key for c in result.must]
+        self.assertIn("user_id", keys)
+        self.assertIn("run_id", keys)
+
+    def test_backward_compat_gte_lte_range(self):
+        """Original gte+lte range filter still produces a Range condition."""
+        result = self.qdrant._create_filter({"count": {"gte": 5, "lte": 10}})
+        self.assertIsInstance(result, Filter)
+        self.assertIsInstance(result.must[0].range, Range)
+        self.assertEqual(result.must[0].range.gte, 5)
+        self.assertEqual(result.must[0].range.lte, 10)
+
+    def test_unknown_operator_raises_error(self):
+        """Unknown operator dict should raise ValueError, not ValidationError."""
+        with self.assertRaises(ValueError) as ctx:
+            self.qdrant._build_field_condition("field", {"unknown_op": "foo"})
+        self.assertIn("Unsupported", str(ctx.exception))
+
+    def test_mixed_range_and_non_range_raises_error(self):
+        """Mixing range ops with non-range ops should raise ValueError."""
+        with self.assertRaises(ValueError):
+            self.qdrant._build_field_condition("priority", {"gte": 5, "ne": 10})
+
+    def test_mixed_range_and_eq_raises_error(self):
+        """Mixing range ops with eq should raise ValueError."""
+        with self.assertRaises(ValueError):
+            self.qdrant._build_field_condition("score", {"gt": 0.5, "eq": 1.0})
+
+    def test_wildcard_returns_none(self):
+        """Wildcard '*' should return None (skip filter — match any)."""
+        result = self.qdrant._build_field_condition("category", "*")
+        self.assertIsNone(result)
+
+    def test_create_filter_with_wildcard_skips_it(self):
+        """Wildcard fields should be skipped in the final filter."""
+        result = self.qdrant._create_filter({"category": "*", "user_id": "alice"})
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must), 1)
+        self.assertEqual(result.must[0].key, "user_id")
+
+    def test_create_filter_only_wildcard_returns_none(self):
+        """Filter with only wildcard should return None."""
+        result = self.qdrant._create_filter({"category": "*"})
+        self.assertIsNone(result)
+
+    def test_and_with_non_list_raises_error(self):
+        """AND with non-list value should raise ValueError."""
+        with self.assertRaises(ValueError):
+            self.qdrant._create_filter({"AND": "not_a_list"})
+
+    def test_or_with_non_list_raises_error(self):
+        """OR with non-list value should raise ValueError."""
+        with self.assertRaises(ValueError):
+            self.qdrant._create_filter({"OR": {"category": "work"}})
+
+    def test_not_with_non_list_raises_error(self):
+        """NOT with non-list value should raise ValueError."""
+        with self.assertRaises(ValueError):
+            self.qdrant._create_filter({"NOT": "invalid"})
+
+    def test_empty_and_list(self):
+        """AND with empty list should return None."""
+        result = self.qdrant._create_filter({"AND": []})
+        self.assertIsNone(result)
+
+    def test_empty_or_list(self):
+        """OR with empty list should return None."""
+        result = self.qdrant._create_filter({"OR": []})
+        self.assertIsNone(result)
+
+    def test_empty_not_list(self):
+        """NOT with empty list should return None."""
+        result = self.qdrant._create_filter({"NOT": []})
+        self.assertIsNone(result)
+
+    def test_deeply_nested_logical(self):
+        """3-level nesting: AND > OR > NOT."""
+        filters = {
+            "AND": [
+                {
+                    "OR": [
+                        {"category": "work"},
+                        {"category": "personal"}
+                    ]
+                },
+                {"priority": {"gte": 5}},
+                {
+                    "NOT": [
+                        {"status": "archived"}
+                    ]
+                }
+            ]
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must), 3)
+
+    def test_boolean_equality(self):
+        """Boolean values should work with MatchValue."""
+        cond = self.qdrant._build_field_condition("active", True)
+        self.assertIsInstance(cond.match, MatchValue)
+        self.assertEqual(cond.match.value, True)
+
+    def test_integer_equality(self):
+        """Integer values should work with MatchValue."""
+        cond = self.qdrant._build_field_condition("count", 42)
+        self.assertIsInstance(cond.match, MatchValue)
+        self.assertEqual(cond.match.value, 42)
+
+    def test_eq_with_boolean(self):
+        """eq operator with boolean should work."""
+        cond = self.qdrant._build_field_condition("active", {"eq": False})
+        self.assertIsInstance(cond.match, MatchValue)
+        self.assertEqual(cond.match.value, False)
+
+    def test_in_with_integers(self):
+        """in operator with integer list should use MatchAny."""
+        cond = self.qdrant._build_field_condition("priority", {"in": [1, 2, 3]})
+        self.assertIsInstance(cond.match, MatchAny)
+        self.assertEqual(cond.match.any, [1, 2, 3])
+
+    def test_wildcard_inside_and(self):
+        """Wildcard inside AND should be skipped, other conditions preserved."""
+        result = self.qdrant._create_filter({
+            "AND": [{"category": "*"}, {"user_id": "alice"}]
+        })
+        self.assertIsInstance(result, Filter)
+        # AND produces nested Filters; the wildcard sub-filter returns None and is skipped
+        # Only the user_id sub-filter remains
+        self.assertEqual(len(result.must), 1)
+
+    def test_list_value_treated_as_match_any(self):
+        """List value shorthand should be treated as in-operator (MatchAny)."""
+        cond = self.qdrant._build_field_condition("tags", ["a", "b", "c"])
+        self.assertIsInstance(cond.match, MatchAny)
+        self.assertEqual(cond.match.any, ["a", "b", "c"])
+
+    def test_list_value_in_create_filter(self):
+        """List value shorthand should work through _create_filter too."""
+        result = self.qdrant._create_filter({"tags": ["python", "rust"]})
+        self.assertIsInstance(result, Filter)
+        self.assertEqual(len(result.must), 1)
+        self.assertIsInstance(result.must[0].match, MatchAny)
+
+    def test_non_dict_item_in_and_raises_error(self):
+        """Non-dict item inside AND list should raise clear ValueError."""
+        with self.assertRaises(ValueError) as ctx:
+            self.qdrant._create_filter({"AND": ["not_a_dict", {"field": "val"}]})
+        self.assertIn("index 0", str(ctx.exception))
+        self.assertIn("must be a dict", str(ctx.exception))
+
+    def test_non_dict_item_in_or_raises_error(self):
+        """Non-dict item inside OR list should raise clear ValueError."""
+        with self.assertRaises(ValueError) as ctx:
+            self.qdrant._create_filter({"OR": [{"field": "val"}, 42]})
+        self.assertIn("index 1", str(ctx.exception))
+
+    def test_empty_dict_value_raises_error(self):
+        """Empty dict as filter value should raise ValueError."""
+        with self.assertRaises(ValueError):
+            self.qdrant._build_field_condition("field", {})
+
+    def test_ne_alone_does_not_trigger_mixed_error(self):
+        """ne as sole operator should NOT trigger mixed-operator error (regression guard)."""
+        cond = self.qdrant._build_field_condition("status", {"ne": "deleted"})
+        self.assertIsInstance(cond.match, MatchExcept)
+
+    def test_eq_with_literal_star(self):
+        """eq operator with literal '*' should match the string '*', not wildcard."""
+        cond = self.qdrant._build_field_condition("category", {"eq": "*"})
+        self.assertIsInstance(cond.match, MatchValue)
+        self.assertEqual(cond.match.value, "*")
+
+    # ------------------------------------------------------------------ #
+    # $or / $not normalization (Memory._process_metadata_filters injects  #
+    # these keys alongside the original OR/NOT)                           #
+    # ------------------------------------------------------------------ #
+
+    def test_dollar_or_handled_as_or(self):
+        """$or injected by Memory middleware should be treated as OR."""
+        filters = {
+            "user_id": "alice",
+            "$or": [{"category": "programming"}, {"category": "data"}],
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertIsNotNone(result.should)
+        self.assertEqual(len(result.should), 2)
+
+    def test_dollar_not_handled_as_not(self):
+        """$not injected by Memory middleware should be treated as NOT."""
+        filters = {
+            "user_id": "alice",
+            "$not": [{"category": "spam"}],
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertIsNotNone(result.must_not)
+        self.assertEqual(len(result.must_not), 1)
+
+    def test_memory_search_or_shape(self):
+        """Simulate exact shape Memory.search() sends for OR filters.
+
+        effective_filters keeps the original OR key (via deepcopy of
+        input_filters) and _process_metadata_filters adds $or with the
+        same content.  _create_filter should deduplicate so only the
+        first occurrence (OR) is used — exactly 2 should entries.
+        """
+        filters = {
+            "OR": [{"category": "programming"}, {"category": "data"}],
+            "user_id": "test_user",
+            "$or": [{"category": "programming"}, {"category": "data"}],
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertIsNotNone(result.should)
+        # Deduplicated: OR wins, $or is skipped — exactly 2 entries
+        self.assertEqual(len(result.should), 2)
+
+    def test_memory_search_not_shape(self):
+        """Simulate exact shape Memory.search() sends for NOT filters.
+
+        Same deduplication as OR: NOT wins, $not is skipped.
+        """
+        filters = {
+            "NOT": [{"category": "spam"}],
+            "user_id": "test_user",
+            "$not": [{"category": "spam"}],
+        }
+        result = self.qdrant._create_filter(filters)
+        self.assertIsInstance(result, Filter)
+        self.assertIsNotNone(result.must_not)
+        # Deduplicated: NOT wins, $not is skipped — exactly 1 entry
+        self.assertEqual(len(result.must_not), 1)

tests/vector_stores/test_turbopuffer.py

@@ -0,0 +1,725 @@
+from datetime import datetime
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+pytest.importorskip("turbopuffer", reason="turbopuffer not installed")
+
+from turbopuffer.types import Row
+
+from mem0.vector_stores.turbopuffer import OutputData, TurbopufferDB
+
+
+def _make_row(id, dist=None, vector=None, **attributes):
+    """Helper to create a turbopuffer Row with extra attributes."""
+    data = {"id": id, "vector": vector}
+    if dist is not None:
+        data["$dist"] = dist
+    data.update(attributes)
+    return Row.model_validate(data)
+
+
+@pytest.fixture
+def mock_client():
+    with patch("mem0.vector_stores.turbopuffer.TurbopufferClient") as MockClient:
+        client_instance = MagicMock()
+        namespace_instance = MagicMock()
+        client_instance.namespace.return_value = namespace_instance
+        MockClient.return_value = client_instance
+        yield client_instance, namespace_instance
+
+
+@pytest.fixture
+def db(mock_client):
+    _, _ = mock_client
+    return TurbopufferDB(
+        collection_name="test_ns",
+        embedding_model_dims=4,
+        api_key="tpuf_test_key",
+        region="gcp-us-central1",
+        distance_metric="cosine_distance",
+        batch_size=2,
+    )
+
+
+# ── Initialization ──────────────────────────────────────────────────
+
+
+class TestInit:
+    def test_init_with_api_key(self, mock_client):
+        client_instance, namespace_instance = mock_client
+        db = TurbopufferDB(
+            collection_name="my_ns",
+            embedding_model_dims=128,
+            api_key="tpuf_key",
+            region="gcp-us-central1",
+        )
+        assert db.collection_name == "my_ns"
+        assert db.embedding_model_dims == 128
+        assert db.distance_metric == "cosine_distance"
+        assert db.batch_size == 100
+        client_instance.namespace.assert_called_with("my_ns")
+
+    def test_init_with_env_var(self, mock_client):
+        with patch.dict("os.environ", {"TURBOPUFFER_API_KEY": "tpuf_env_key"}):
+            db = TurbopufferDB(
+                collection_name="test",
+                embedding_model_dims=4,
+                region="gcp-us-central1",
+            )
+            assert db.collection_name == "test"
+
+    def test_init_without_api_key_raises(self, mock_client):
+        with patch.dict("os.environ", {}, clear=True):
+            # Remove TURBOPUFFER_API_KEY if it exists
+            import os
+            os.environ.pop("TURBOPUFFER_API_KEY", None)
+            with pytest.raises(ValueError, match="API key must be provided"):
+                TurbopufferDB(
+                    collection_name="test",
+                    embedding_model_dims=4,
+                    region="gcp-us-central1",
+                )
+
+    def test_init_with_extra_params(self, mock_client):
+        client_instance, _ = mock_client
+        with patch("mem0.vector_stores.turbopuffer.TurbopufferClient") as MockClient:
+            MockClient.return_value = client_instance
+            TurbopufferDB(
+                collection_name="test",
+                embedding_model_dims=4,
+                api_key="key",
+                region="aws-us-west-2",
+                extra_params={"compression": True},
+            )
+            MockClient.assert_called_with(
+                api_key="key",
+                region="aws-us-west-2",
+                compression=True,
+            )
+
+    def test_init_default_region(self, mock_client):
+        client_instance, _ = mock_client
+        with patch("mem0.vector_stores.turbopuffer.TurbopufferClient") as MockClient:
+            MockClient.return_value = client_instance
+            TurbopufferDB(
+                collection_name="test",
+                embedding_model_dims=4,
+                api_key="key",
+            )
+            MockClient.assert_called_with(
+                api_key="key",
+                region="gcp-us-central1",
+            )
+
+
+# ── create_col ───────────────────────────────────────────────────────
+
+
+class TestCreateCol:
+    def test_create_col_is_noop(self, db):
+        # Should not raise or call anything
+        result = db.create_col()
+        assert result is None
+
+    def test_create_col_with_args_is_noop(self, db):
+        result = db.create_col(name="x", vector_size=128, distance="cosine")
+        assert result is None
+
+
+# ── insert ───────────────────────────────────────────────────────────
+
+
+class TestInsert:
+    def test_insert_with_ids_and_payloads(self, db):
+        vectors = [[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8]]
+        payloads = [{"data": "hello", "user_id": "u1"}, {"data": "world", "user_id": "u2"}]
+        ids = ["id1", "id2"]
+
+        db.insert(vectors, payloads, ids)
+
+        db.namespace.write.assert_called_once()
+        call_kwargs = db.namespace.write.call_args[1]
+        rows = call_kwargs["upsert_rows"]
+        assert len(rows) == 2
+        assert rows[0]["id"] == "id1"
+        assert rows[0]["vector"] == [0.1, 0.2, 0.3, 0.4]
+        assert rows[0]["data"] == "hello"
+        assert rows[0]["user_id"] == "u1"
+        assert rows[1]["id"] == "id2"
+        assert call_kwargs["distance_metric"] == "cosine_distance"
+
+    def test_insert_without_ids_generates_ids(self, db):
+        vectors = [[0.1, 0.2, 0.3, 0.4]]
+        db.insert(vectors)
+
+        db.namespace.write.assert_called_once()
+        call_kwargs = db.namespace.write.call_args[1]
+        assert call_kwargs["upsert_rows"][0]["id"] == "0"
+
+    def test_insert_without_payloads(self, db):
+        vectors = [[0.1, 0.2, 0.3, 0.4]]
+        ids = ["id1"]
+        db.insert(vectors, ids=ids)
+
+        call_kwargs = db.namespace.write.call_args[1]
+        row = call_kwargs["upsert_rows"][0]
+        assert row["id"] == "id1"
+        assert row["vector"] == [0.1, 0.2, 0.3, 0.4]
+        assert len(row) == 2  # only id and vector
+
+    def test_insert_payload_does_not_overwrite_id_or_vector(self, db):
+        """Payload with 'id' or 'vector' keys must not overwrite the actual values."""
+        vectors = [[0.1, 0.2, 0.3, 0.4]]
+        payloads = [{"id": "fake_id", "vector": "fake_vector", "data": "hello"}]
+        ids = ["real_id"]
+        db.insert(vectors, payloads, ids)
+
+        call_kwargs = db.namespace.write.call_args[1]
+        row = call_kwargs["upsert_rows"][0]
+        assert row["id"] == "real_id"
+        assert row["vector"] == [0.1, 0.2, 0.3, 0.4]
+        assert row["data"] == "hello"
+
+    def test_insert_batching(self, db):
+        """batch_size=2, so 3 vectors should produce 2 write calls."""
+        vectors = [[0.1] * 4, [0.2] * 4, [0.3] * 4]
+        ids = ["a", "b", "c"]
+        db.insert(vectors, ids=ids)
+
+        assert db.namespace.write.call_count == 2
+        first_call = db.namespace.write.call_args_list[0][1]
+        second_call = db.namespace.write.call_args_list[1][1]
+        assert len(first_call["upsert_rows"]) == 2
+        assert len(second_call["upsert_rows"]) == 1
+
+
+# ── _parse_output ────────────────────────────────────────────────────
+
+
+class TestParseOutput:
+    def test_parse_rows_with_dist_and_attributes(self, db):
+        rows = [
+            _make_row("id1", dist=0.1, data="hello", user_id="u1"),
+            _make_row("id2", dist=0.3, data="world", user_id="u2"),
+        ]
+        results = db._parse_output(rows)
+
+        assert len(results) == 2
+        assert results[0].id == "id1"
+        assert results[0].score == pytest.approx(0.9)
+        assert results[0].payload == {"data": "hello", "user_id": "u1"}
+        assert results[1].id == "id2"
+        assert results[1].score == pytest.approx(0.7)
+        assert results[1].payload == {"data": "world", "user_id": "u2"}
+
+    def test_parse_rows_without_dist(self, db):
+        rows = [_make_row("id1", data="hello")]
+        results = db._parse_output(rows)
+
+        assert results[0].score is None
+        assert results[0].payload == {"data": "hello"}
+
+    def test_parse_rows_strips_vector_and_id(self, db):
+        rows = [_make_row("id1", dist=0.0, vector=[1.0, 2.0, 3.0, 4.0], data="test")]
+        results = db._parse_output(rows)
+
+        assert "vector" not in results[0].payload
+        assert "id" not in results[0].payload
+        assert "$dist" not in results[0].payload
+        assert results[0].id == "id1"
+
+    def test_parse_empty_rows(self, db):
+        assert db._parse_output([]) == []
+
+
+# ── _convert_filters ─────────────────────────────────────────────────
+
+
+class TestConvertFilters:
+    def test_none_filters(self, db):
+        assert db._convert_filters(None) is None
+
+    def test_empty_filters(self, db):
+        assert db._convert_filters({}) is None
+
+    def test_single_eq_filter(self, db):
+        result = db._convert_filters({"user_id": "u1"})
+        assert result == ("user_id", "Eq", "u1")
+
+    def test_multiple_eq_filters(self, db):
+        result = db._convert_filters({"user_id": "u1", "agent_id": "a1"})
+        assert result[0] == "And"
+        conditions = result[1]
+        assert ("user_id", "Eq", "u1") in conditions
+        assert ("agent_id", "Eq", "a1") in conditions
+
+    def test_range_filter_gte_lte(self, db):
+        result = db._convert_filters({"score": {"gte": 0.5, "lte": 1.0}})
+        assert result[0] == "And"
+        conditions = result[1]
+        assert ("score", "Gte", 0.5) in conditions
+        assert ("score", "Lte", 1.0) in conditions
+
+    def test_range_filter_gte_only(self, db):
+        result = db._convert_filters({"score": {"gte": 0.5}})
+        assert result == ("score", "Gte", 0.5)
+
+    def test_range_filter_lte_only(self, db):
+        result = db._convert_filters({"score": {"lte": 1.0}})
+        assert result == ("score", "Lte", 1.0)
+
+    def test_mixed_eq_and_range_filters(self, db):
+        result = db._convert_filters({"user_id": "u1", "score": {"gte": 0.5}})
+        assert result[0] == "And"
+        conditions = result[1]
+        assert ("user_id", "Eq", "u1") in conditions
+        assert ("score", "Gte", 0.5) in conditions
+
+
+# ── search ───────────────────────────────────────────────────────────
+
+
+class TestSearch:
+    def test_search_basic(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = [
+            _make_row("id1", dist=0.1, data="hello"),
+            _make_row("id2", dist=0.2, data="world"),
+        ]
+        db.namespace.query.return_value = mock_response
+
+        results = db.search("test query", [0.1, 0.2, 0.3, 0.4], limit=2)
+
+        db.namespace.query.assert_called_once_with(
+            rank_by=("vector", "ANN", [0.1, 0.2, 0.3, 0.4]),
+            top_k=2,
+            include_attributes=True,
+        )
+        assert len(results) == 2
+        assert results[0].id == "id1"
+        assert results[0].score == pytest.approx(0.9)
+
+    def test_search_with_filters(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = [_make_row("id1", dist=0.1, data="hello")]
+        db.namespace.query.return_value = mock_response
+
+        results = db.search(
+            "query", [0.1, 0.2, 0.3, 0.4], limit=5, filters={"user_id": "u1"}
+        )
+
+        call_kwargs = db.namespace.query.call_args[1]
+        assert call_kwargs["filters"] == ("user_id", "Eq", "u1")
+        assert len(results) == 1
+
+    def test_search_no_results(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = None
+        db.namespace.query.return_value = mock_response
+
+        results = db.search("query", [0.1, 0.2, 0.3, 0.4])
+        assert results == []
+
+    def test_search_empty_rows(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = []
+        db.namespace.query.return_value = mock_response
+
+        results = db.search("query", [0.1, 0.2, 0.3, 0.4])
+        assert results == []
+
+
+# ── delete ───────────────────────────────────────────────────────────
+
+
+class TestDelete:
+    def test_delete_by_string_id(self, db):
+        db.delete("id1")
+        db.namespace.write.assert_called_once_with(deletes=["id1"])
+
+    def test_delete_by_int_id(self, db):
+        db.delete(123)
+        db.namespace.write.assert_called_once_with(deletes=["123"])
+
+
+# ── update ───────────────────────────────────────────────────────────
+
+
+class TestUpdate:
+    def test_update_with_vector_and_payload(self, db):
+        db.update("id1", vector=[0.5, 0.6, 0.7, 0.8], payload={"data": "updated"})
+
+        db.namespace.write.assert_called_once()
+        call_kwargs = db.namespace.write.call_args[1]
+        row = call_kwargs["upsert_rows"][0]
+        assert row["id"] == "id1"
+        assert row["vector"] == [0.5, 0.6, 0.7, 0.8]
+        assert row["data"] == "updated"
+        assert call_kwargs["distance_metric"] == "cosine_distance"
+
+    def test_update_vector_only(self, db):
+        db.update("id1", vector=[0.5, 0.6, 0.7, 0.8])
+
+        call_kwargs = db.namespace.write.call_args[1]
+        row = call_kwargs["upsert_rows"][0]
+        assert row["id"] == "id1"
+        assert row["vector"] == [0.5, 0.6, 0.7, 0.8]
+        assert "data" not in row
+
+    def test_update_payload_only_uses_patch(self, db):
+        """Payload-only updates should use patch_rows, not upsert_rows."""
+        db.update("id1", vector=None, payload={"data": "patched", "user_id": "u1"})
+
+        call_kwargs = db.namespace.write.call_args[1]
+        row = call_kwargs["patch_rows"][0]
+        assert row["id"] == "id1"
+        assert row["data"] == "patched"
+        assert row["user_id"] == "u1"
+        assert "upsert_rows" not in call_kwargs
+
+    def test_update_payload_does_not_overwrite_id(self, db):
+        """Payload with an 'id' key must not overwrite the actual vector ID."""
+        db.update("real_id", vector=None, payload={"id": "fake_id", "data": "test"})
+        call_kwargs = db.namespace.write.call_args[1]
+        row = call_kwargs["patch_rows"][0]
+        assert row["id"] == "real_id"
+
+    def test_update_nothing(self, db):
+        """Neither vector nor payload: should not call write."""
+        db.update("id1")
+        db.namespace.write.assert_not_called()
+
+
+# ── get ──────────────────────────────────────────────────────────────
+
+
+class TestGet:
+    def test_get_found(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = [_make_row("id1", dist=0.0, data="hello", user_id="u1")]
+        db.namespace.query.return_value = mock_response
+
+        result = db.get("id1")
+
+        db.namespace.query.assert_called_once_with(
+            top_k=1,
+            rank_by=("vector", "ANN", [0.0, 0.0, 0.0, 0.0]),
+            filters=("id", "Eq", "id1"),
+            include_attributes=True,
+        )
+        assert result is not None
+        assert result.id == "id1"
+        assert result.payload["data"] == "hello"
+        assert result.payload["user_id"] == "u1"
+
+    def test_get_not_found(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = []
+        db.namespace.query.return_value = mock_response
+
+        result = db.get("nonexistent")
+        assert result is None
+
+    def test_get_none_rows(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = None
+        db.namespace.query.return_value = mock_response
+
+        result = db.get("id1")
+        assert result is None
+
+    def test_get_handles_exception(self, db):
+        db.namespace.query.side_effect = Exception("API error")
+        result = db.get("id1")
+        assert result is None
+
+
+# ── list_cols ────────────────────────────────────────────────────────
+
+
+class TestListCols:
+    def test_list_cols(self, db):
+        ns1 = MagicMock()
+        ns1.id = "ns1"
+        ns2 = MagicMock()
+        ns2.id = "ns2"
+        db.client.namespaces.return_value = [ns1, ns2]
+
+        result = db.list_cols()
+        assert len(result) == 2
+        db.client.namespaces.assert_called_once()
+
+
+# ── delete_col ───────────────────────────────────────────────────────
+
+
+class TestDeleteCol:
+    def test_delete_col(self, db):
+        db.delete_col()
+        db.namespace.delete_all.assert_called_once()
+
+    def test_delete_col_handles_error(self, db):
+        db.namespace.delete_all.side_effect = Exception("API error")
+        # Should not raise
+        db.delete_col()
+
+
+# ── col_info ─────────────────────────────────────────────────────────
+
+
+class TestColInfo:
+    def test_col_info(self, db):
+        mock_metadata = MagicMock()
+        mock_metadata.approx_row_count = 42
+        mock_metadata.approx_logical_bytes = 1024
+        mock_metadata.created_at = datetime(2025, 1, 1)
+        mock_metadata.updated_at = datetime(2025, 6, 1)
+        db.namespace.metadata.return_value = mock_metadata
+
+        info = db.col_info()
+        assert info["name"] == "test_ns"
+        assert info["approx_row_count"] == 42
+        assert info["approx_logical_bytes"] == 1024
+
+    def test_col_info_handles_error(self, db):
+        db.namespace.metadata.side_effect = Exception("not found")
+        info = db.col_info()
+        assert info == {"name": "test_ns"}
+
+
+# ── list ─────────────────────────────────────────────────────────────
+
+
+class TestList:
+    def test_list_returns_wrapped_format(self, db):
+        """list() must return [[results]] for compatibility with main.py."""
+        mock_response = MagicMock()
+        mock_response.rows = [
+            _make_row("id1", dist=0.1, data="hello"),
+            _make_row("id2", dist=0.2, data="world"),
+        ]
+        db.namespace.query.return_value = mock_response
+
+        result = db.list()
+
+        # Must be wrapped: result[0] is the actual list
+        assert isinstance(result, list)
+        assert isinstance(result[0], list)
+        assert len(result[0]) == 2
+        assert result[0][0].id == "id1"
+        assert result[0][1].id == "id2"
+
+    def test_list_with_filters(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = [_make_row("id1", dist=0.1, data="hello")]
+        db.namespace.query.return_value = mock_response
+
+        db.list(filters={"user_id": "u1"}, limit=50)
+
+        call_kwargs = db.namespace.query.call_args[1]
+        assert call_kwargs["filters"] == ("user_id", "Eq", "u1")
+        assert call_kwargs["top_k"] == 50
+
+    def test_list_empty(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = None
+        db.namespace.query.return_value = mock_response
+
+        result = db.list()
+        assert result == [[]]
+
+    def test_list_uses_zero_vector(self, db):
+        mock_response = MagicMock()
+        mock_response.rows = []
+        db.namespace.query.return_value = mock_response
+
+        db.list()
+        call_kwargs = db.namespace.query.call_args[1]
+        assert call_kwargs["rank_by"] == ("vector", "ANN", [0.0, 0.0, 0.0, 0.0])
+
+    def test_list_compatible_with_main_py_get_all(self, db):
+        """Simulate how main.py _get_all_from_vector_store unwraps list()."""
+        mock_response = MagicMock()
+        mock_response.rows = [_make_row("id1", dist=0.1, data="hello")]
+        db.namespace.query.return_value = mock_response
+
+        memories_result = db.list(filters={"user_id": "u1"})
+
+        # Reproduce main.py unwrapping logic
+        first_element = memories_result[0]
+        if isinstance(first_element, (list, tuple)):
+            actual_memories = first_element
+        else:
+            actual_memories = memories_result
+
+        assert isinstance(actual_memories, list)
+        assert len(actual_memories) == 1
+        assert actual_memories[0].id == "id1"
+        assert actual_memories[0].payload["data"] == "hello"
+
+    def test_list_compatible_with_main_py_delete_all(self, db):
+        """Simulate how main.py delete_all uses list()[0]."""
+        mock_response = MagicMock()
+        mock_response.rows = [
+            _make_row("id1", dist=0.1, data="hello"),
+            _make_row("id2", dist=0.2, data="world"),
+        ]
+        db.namespace.query.return_value = mock_response
+
+        result = db.list(filters={"user_id": "u1"})
+        memories = result[0]
+
+        assert isinstance(memories, list)
+        for mem in memories:
+            assert hasattr(mem, "id")
+            assert hasattr(mem, "payload")
+
+
+# ── count ────────────────────────────────────────────────────────────
+
+
+class TestCount:
+    def test_count(self, db):
+        mock_metadata = MagicMock()
+        mock_metadata.approx_row_count = 100
+        db.namespace.metadata.return_value = mock_metadata
+
+        assert db.count() == 100
+
+    def test_count_handles_error(self, db):
+        db.namespace.metadata.side_effect = Exception("error")
+        assert db.count() == 0
+
+
+# ── reset ────────────────────────────────────────────────────────────
+
+
+class TestReset:
+    def test_reset_calls_delete_all(self, db):
+        db.reset()
+        db.namespace.delete_all.assert_called_once()
+
+
+# ── Config ───────────────────────────────────────────────────────────
+
+
+class TestConfig:
+    def test_config_valid(self):
+        with patch.dict("os.environ", {"TURBOPUFFER_API_KEY": "key"}):
+            from mem0.configs.vector_stores.turbopuffer import TurbopufferConfig
+
+            config = TurbopufferConfig()
+            assert config.collection_name == "mem0"
+            assert config.embedding_model_dims == 1536
+            assert config.distance_metric == "cosine_distance"
+            assert config.batch_size == 100
+            assert config.region == "gcp-us-central1"
+
+    def test_config_custom_values(self):
+        from mem0.configs.vector_stores.turbopuffer import TurbopufferConfig
+
+        config = TurbopufferConfig(
+            collection_name="custom",
+            embedding_model_dims=768,
+            api_key="tpuf_key",
+            region="aws-us-west-2",
+            distance_metric="euclidean_squared",
+            batch_size=50,
+        )
+        assert config.collection_name == "custom"
+        assert config.embedding_model_dims == 768
+        assert config.api_key == "tpuf_key"
+        assert config.region == "aws-us-west-2"
+        assert config.distance_metric == "euclidean_squared"
+        assert config.batch_size == 50
+
+    def test_config_rejects_extra_fields(self):
+        from mem0.configs.vector_stores.turbopuffer import TurbopufferConfig
+
+        with pytest.raises(ValueError, match="Extra fields not allowed"):
+            TurbopufferConfig(api_key="key", unknown_field="value")
+
+    def test_config_requires_api_key(self):
+        from mem0.configs.vector_stores.turbopuffer import TurbopufferConfig
+
+        with patch.dict("os.environ", {}, clear=True):
+            import os
+            os.environ.pop("TURBOPUFFER_API_KEY", None)
+            with pytest.raises(ValueError, match="api_key"):
+                TurbopufferConfig()
+
+    def test_config_accepts_env_var(self):
+        from mem0.configs.vector_stores.turbopuffer import TurbopufferConfig
+
+        with patch.dict("os.environ", {"TURBOPUFFER_API_KEY": "env_key"}):
+            config = TurbopufferConfig()
+            assert config.api_key is None  # not set explicitly, but env var is present
+
+
+# ── Factory Registration ─────────────────────────────────────────────
+
+
+class TestFactoryRegistration:
+    def test_vector_store_factory_has_turbopuffer(self):
+        from mem0.utils.factory import VectorStoreFactory
+
+        assert "turbopuffer" in VectorStoreFactory.provider_to_class
+        assert VectorStoreFactory.provider_to_class["turbopuffer"] == "mem0.vector_stores.turbopuffer.TurbopufferDB"
+
+    def test_vector_store_config_has_turbopuffer(self):
+        from mem0.vector_stores.configs import VectorStoreConfig
+
+        config = VectorStoreConfig.__private_attributes__["_provider_configs"].default
+        assert "turbopuffer" in config
+        assert config["turbopuffer"] == "TurbopufferConfig"
+
+    def test_config_validation_pipeline(self):
+        """Test that VectorStoreConfig correctly resolves turbopuffer config."""
+        from mem0.vector_stores.configs import VectorStoreConfig
+
+        with patch.dict("os.environ", {"TURBOPUFFER_API_KEY": "key"}):
+            config = VectorStoreConfig(
+                provider="turbopuffer",
+                config={"collection_name": "test", "region": "gcp-us-central1"},
+            )
+            assert config.config.collection_name == "test"
+            assert config.config.region == "gcp-us-central1"
+            assert config.config.embedding_model_dims == 1536
+
+
+# ── OutputData ───────────────────────────────────────────────────────
+
+
+class TestOutputData:
+    def test_output_data_has_required_fields(self):
+        od = OutputData(id="test", score=0.9, payload={"data": "hello"})
+        assert od.id == "test"
+        assert od.score == 0.9
+        assert od.payload == {"data": "hello"}
+
+    def test_output_data_nullable_fields(self):
+        od = OutputData(id=None, score=None, payload=None)
+        assert od.id is None
+        assert od.score is None
+        assert od.payload is None
+
+    def test_output_data_payload_access_pattern(self):
+        """Test the exact access pattern used by main.py."""
+        od = OutputData(
+            id="mem1",
+            score=0.95,
+            payload={
+                "data": "User likes sci-fi movies",
+                "hash": "abc123",
+                "created_at": "2025-01-01",
+                "updated_at": "2025-06-01",
+                "user_id": "alice",
+                "agent_id": "agent1",
+            },
+        )
+        assert od.payload.get("data", "") == "User likes sci-fi movies"
+        assert od.payload.get("hash") == "abc123"
+        assert od.payload.get("created_at") == "2025-01-01"
+        assert od.payload.get("user_id") == "alice"
+        assert od.payload.get("run_id") is None  # not present, should return None