docs(changelog): correct score_details fields, graph-store removal note, and dead migration link

docs: make example code fences copy-safe (#5833 )
fix(mem0-plugin): store files_touched as a list to stop double JSON-encoding (#5806 )
2026-06-25 11:59:57 +05:30 · 2026-06-24 22:19:40 -07:00 · 2026-06-25 09:06:11 +05:30 · 2026-06-24 15:48:37 -07:00 · 2026-06-24 20:55:14 +05:30 · 2026-06-24 20:21:06 +05:30
61 changed files with 2044 additions and 692 deletions
@@ -12,7 +12,7 @@
      "name": "mem0",
      "source": "./integrations/mem0-plugin",
      "description": "Mem0 memory layer for AI applications. Add persistent memory, personalization, and semantic search to Claude workflows.",
-      "version": "0.2.10"
+      "version": "0.2.11"
    }
  ]
 }
@@ -12,7 +12,7 @@
      "name": "mem0",
      "source": "./integrations/mem0-plugin",
      "description": "Mem0 memory layer for AI applications. Add persistent memory, personalization, and semantic search.",
-      "version": "0.2.10"
+      "version": "0.2.11"
    }
  ]
 }
@@ -113,7 +113,7 @@ Launched a unified Mem0 plugin across three major AI development environments

 Major expansion of the provider ecosystem:

- **Apache AGE** — New graph store support, bringing the total to 4 graph store backends (Neo4j, Memgraph, Kuzu, Apache AGE)
+- **Apache AGE** — New graph store support, bringing the total to 4 graph store backends (Neo4j, Memgraph, Kuzu, Apache AGE). **Note:** All external graph store backends (Neo4j, Memgraph, Kuzu, Apache AGE) were subsequently removed in v2.0.0 (2026-04-14). Graph memory is now built-in entity linking with no external graph store required; see the [v2.0.0 entry above](#mem0-sdk-v2-0-0-v3-0-0).
 - **Turbopuffer** — New vector database provider for Python SDK
 - **MiniMax** — New LLM provider with dedicated AWS Bedrock support
 - **pgvector for Node.js** — PostgreSQL vector support added to the TypeScript OSS SDK
@@ -7,6 +7,65 @@ mode: "wide"
 <Tabs>
 <Tab title="Python">

+<Update label="2026-06-24" description="v2.0.9">
+
+**Bug Fixes:**
+- **Memory (OSS):** Improve entity extraction precision by avoiding sentence-start common noun noise, preserving useful topic phrases, and exact-deduplicating entity links before semantic matching ([#5829](https://github.com/mem0ai/mem0/pull/5829))
+
+</Update>
+
+<Update label="2026-06-24" description="v2.0.8">
+
+**New Features:**
+- **Embeddings:** Add native `embed_batch` to five embedders — LM Studio, Together, HuggingFace, Vertex AI, and Google GenAI — for batched embedding requests ([#5609](https://github.com/mem0ai/mem0/pull/5609))
+
+**Bug Fixes:**
+- **Core:** Guard against malformed `image_url` entries in `parse_vision_messages` to prevent crashes ([#5631](https://github.com/mem0ai/mem0/pull/5631))
+- **Core:** Return `attributed_to` from `get()`, `get_all()`, and `search()` ([#5629](https://github.com/mem0ai/mem0/pull/5629))
+- **Core:** Fix `reset()` only dropping the history table and leaving stale messages behind ([#5541](https://github.com/mem0ai/mem0/pull/5541))
+- **Core:** Guard against an entity `embed_batch` count mismatch in the v3 add pipeline ([#5604](https://github.com/mem0ai/mem0/pull/5604))
+- **Core:** Fix an async `delete_all` race condition that corrupted the entity store's `linked_memory_ids` ([#5553](https://github.com/mem0ai/mem0/pull/5553))
+- **LLMs:** Skip the JSON `response_format` for Groq compound models that reject it ([#5513](https://github.com/mem0ai/mem0/pull/5513))
+- **LLMs:** Preserve reasoning fields during base-to-provider config conversion ([#5638](https://github.com/mem0ai/mem0/pull/5638))
+- **LLMs:** Pass the configured `anthropic_base_url` to the Anthropic client ([#5626](https://github.com/mem0ai/mem0/pull/5626))
+- **LLMs:** Stop the Azure provider from mutating and corrupting caller messages during content rewrite ([#5731](https://github.com/mem0ai/mem0/pull/5731))
+- **LLMs & Embeddings:** Repair HTTP proxy support for `httpx>=0.28` and preserve `proxies` in `LlmFactory` ([#5447](https://github.com/mem0ai/mem0/pull/5447))
+- **Embeddings:** Forward `embedding_dims` to Titan V2 in the AWS Bedrock embedder ([#5671](https://github.com/mem0ai/mem0/pull/5671))
+- **Rerankers:** Log reranking failures instead of swallowing them silently ([#5717](https://github.com/mem0ai/mem0/pull/5717))
+- **Rerankers:** Clamp out-of-range LLM scores instead of mis-parsing them ([#5635](https://github.com/mem0ai/mem0/pull/5635))
+- **Rerankers:** Export all five rerankers from the package root ([#5636](https://github.com/mem0ai/mem0/pull/5636))
+- **Vector Stores:** Point the FastEmbed-missing warning at `mem0ai[extras]` ([#5622](https://github.com/mem0ai/mem0/pull/5622))
+- **Vector Stores:** Preserve empty Azure AI Search update values ([#5524](https://github.com/mem0ai/mem0/pull/5524))
+- **Vector Stores:** Add an `auto_refresh` option for OpenSearch Serverless compatibility ([#3893](https://github.com/mem0ai/mem0/pull/3893))
+- **Vector Stores:** Wrap a scalar `vector_id` in a list for Chroma `delete()` ([#5703](https://github.com/mem0ai/mem0/pull/5703))
+- **Vector Stores:** Wrap Chroma `update()` ids, embeddings, and metadatas in lists ([#5757](https://github.com/mem0ai/mem0/pull/5757))
+- **Vector Stores:** Wrap a scalar `vector_id` in a list for Milvus `delete()` ([#5704](https://github.com/mem0ai/mem0/pull/5704))
+- **Vector Stores:** Map all comparison operators in the Pinecone `_create_filter()` ([#5707](https://github.com/mem0ai/mem0/pull/5707))
+- **Vector Stores:** Return `None` instead of `{}` from Chroma `_generate_where_clause` for empty filters ([#5713](https://github.com/mem0ai/mem0/pull/5713))
+- **Vector Stores:** Return `[[]]` from the OpenSearch `list()` error path to honor the `list()` contract ([#5727](https://github.com/mem0ai/mem0/pull/5727))
+- **Vector Stores:** Return `[[]]` from the Pinecone `list()` error path instead of a dict ([#5706](https://github.com/mem0ai/mem0/pull/5706))
+- **Vector Stores:** Return `[[]]` for an uninitialized FAISS index to honor the `list()` contract ([#5725](https://github.com/mem0ai/mem0/pull/5725))
+- **Vector Stores:** Wrap the MongoDB `list()` return in an outer list to match the interface contract ([#5729](https://github.com/mem0ai/mem0/pull/5729))
+- **Vector Stores:** Deep-copy Redis `DEFAULT_FIELDS` so instances keep distinct dims ([#5633](https://github.com/mem0ai/mem0/pull/5633))
+- **Vector Stores:** Pass the required `vectors` arg in Vertex AI `list()` and similarity search ([#5627](https://github.com/mem0ai/mem0/pull/5627))
+- **Vector Stores:** Return `None` from Redis `get()` for missing IDs ([#5625](https://github.com/mem0ai/mem0/pull/5625))
+- **Vector Stores:** Drop a stray `print` in Weaviate `list_cols` ([#5637](https://github.com/mem0ai/mem0/pull/5637))
+- **Graph:** Keep distinct entities that share a substring prefix ([#5630](https://github.com/mem0ai/mem0/pull/5630))
+- **Client:** Check the HTTP status before parsing the ping response in `_validate_api_key` ([#5639](https://github.com/mem0ai/mem0/pull/5639))
+- **Server:** Fetch filtered dashboard memories beyond the default page ([#5753](https://github.com/mem0ai/mem0/pull/5753))
+- **Server:** Return 404/400 instead of 502 for not-found and invalid input ([#5634](https://github.com/mem0ai/mem0/pull/5634))
+- **Server:** Return 404 instead of 500 for a malformed API key id on revoke ([#5640](https://github.com/mem0ai/mem0/pull/5640))
+- **Server:** Use `127.0.0.1` in the dashboard healthcheck to avoid IPv6 localhost resolution ([#5612](https://github.com/mem0ai/mem0/pull/5612))
+
+**Improvements:**
+- **Vector Stores:** Batch BM25 sparse encoding in Qdrant insert ([#5592](https://github.com/mem0ai/mem0/pull/5592))
+
+**Security:**
+- **Vector Stores:** Sanitize Milvus and Baidu filter values to prevent expression injection ([#5746](https://github.com/mem0ai/mem0/pull/5746))
+- **Vector Stores:** Reject dict filter values in MongoDB to prevent NoSQL operator injection ([#5748](https://github.com/mem0ai/mem0/pull/5748))
+
+</Update>
+
 <Update label="2026-06-17" description="v2.0.7">

 **New Features:**
@@ -62,7 +121,7 @@ mode: "wide"

 **New Features:**
 - **Memory:** Warn at init time when hybrid/BM25 search silently degrades to semantic-only because the configured vector store does not implement `keyword_search`. Affected stores: Chroma, FAISS, Cassandra, LangChain, Neptune Analytics, S3 Vectors, Supabase, TurboPuffer, Valkey ([#5444](https://github.com/mem0ai/mem0/pull/5444))
- **Memory:** Add opt-in `explain=True` parameter to `Memory.search()` and `AsyncMemory.search()`. When enabled, each result includes a `score_breakdown` dict with `semantic`, `keyword` (normalized BM25), `entity_boost`, and `temporal_boost` signals so callers can understand and tune retrieval ranking ([#5102](https://github.com/mem0ai/mem0/pull/5102))
+- **Memory:** Add opt-in `explain=True` parameter to `Memory.search()` and `AsyncMemory.search()`. When enabled, each result includes a `score_details` dict with `semantic_score`, `bm25_score`, `entity_boost`, `raw_score`, `max_possible_score`, `final_score`, and `threshold` so callers can understand and tune retrieval ranking ([#5102](https://github.com/mem0ai/mem0/pull/5102))

 **Bug Fixes:**
 - **Vector Stores:** Normalize similarity scores to `[0, 1]` (higher = better) consistently across all backends. 11 adapters previously returned raw distance metrics (lower = better) — FAISS, Chroma, Milvus, Redis, Cassandra, PGVector, S3 Vectors, Supabase, Valkey, Azure MySQL, and Vertex AI Vector Search — causing incorrect ranking in multi-store setups ([#5391](https://github.com/mem0ai/mem0/pull/5391))
@@ -170,7 +229,7 @@ mode: "wide"
 **Improvements:**
 - **Telemetry:** Sample OSS hot-path events at 10% via PostHog `before_send` hook to reduce event volume ([#4771](https://github.com/mem0ai/mem0/pull/4771))

-See the [OSS v1 to v2 migration guide](https://docs.mem0.ai/migration/oss-v1-to-v2) and [Platform migration guide](https://docs.mem0.ai/migration/platform-v2-to-v3) for upgrade instructions.
+See the [OSS v2 to v3 migration guide](https://docs.mem0.ai/migration/oss-v2-to-v3) and [Platform migration guide](https://docs.mem0.ai/migration/platform-v2-to-v3) for upgrade instructions.

 </Update>

@@ -1011,6 +1070,31 @@ See the [OSS v1 to v2 migration guide](https://docs.mem0.ai/migration/oss-v1-to-

 <Tab title="TypeScript">

+<Update label="2026-06-24" description="v3.0.11">
+
+**Bug Fixes:**
+- **Memory (OSS):** Align entity extraction with Python by reducing generic entity noise, preserving useful topic phrases, and exact-deduplicating entity links before semantic matching ([#5829](https://github.com/mem0ai/mem0/pull/5829))
+
+</Update>
+
+<Update label="2026-06-24" description="v3.0.10">
+
+**Bug Fixes:**
+- **Memory (OSS):** Guard against malformed `image_url` entries in `parseVisionMessages` to prevent crashes ([#5631](https://github.com/mem0ai/mem0/pull/5631))
+- **Memory (OSS):** Return `attributedTo` from `get()`, `search()`, and `getAll()` ([#5675](https://github.com/mem0ai/mem0/pull/5675))
+- **Memory (OSS):** Preserve message roles in the extraction input so assistant facts aren't attributed to the user ([#5643](https://github.com/mem0ai/mem0/pull/5643))
+- **Memory (OSS):** Reject empty or blank messages in `Memory.add()` to prevent hallucinated memories ([#5545](https://github.com/mem0ai/mem0/pull/5545))
+- **Memory (OSS):** Check `message.role` instead of `content` when detecting system messages ([#3921](https://github.com/mem0ai/mem0/pull/3921))
+- **LLMs:** Honor the configured `baseURL` in `AnthropicLLM` ([#5740](https://github.com/mem0ai/mem0/pull/5740))
+- **Client:** Preserve `customCategories` names through key conversion ([#5741](https://github.com/mem0ai/mem0/pull/5741))
+- **Client:** Prevent hallucinated memories on an empty messages payload ([#5613](https://github.com/mem0ai/mem0/pull/5613))
+- **Client:** Preserve user metadata keys across the case-conversion round-trip ([#5515](https://github.com/mem0ai/mem0/pull/5515))
+
+**Security:**
+- **Dependencies:** Upgrade `form-data` to `>=4.0.6` across pnpm workspaces to remediate CVE-2026-12143 ([#5618](https://github.com/mem0ai/mem0/pull/5618))
+
+</Update>
+
 <Update label="2026-06-17" description="v3.0.9">

 **Bug Fixes:**
@@ -7,7 +7,8 @@ To use DeepSeek LLM models, you have to set the `DEEPSEEK_API_KEY` environment v

 ## Usage

-```python
+<CodeGroup>
+```python Python
 import os
 from mem0 import Memory

@@ -36,6 +37,32 @@ messages = [
 m.add(messages, user_id="alice", metadata={"category": "movies"})
 ```

+```typescript TypeScript
+import { Memory } from 'mem0ai/oss';
+
+const config = {
+  llm: {
+    provider: 'deepseek',
+    config: {
+      apiKey: process.env.DEEPSEEK_API_KEY || '',
+      model: 'deepseek-chat',
+      temperature: 0.2,
+      maxTokens: 2000,
+      top_p: 1.0,
+    },
+  },
+};
+const memory = new Memory(config);
+const 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."}
+];
+await memory.add(messages, { userId: 'alice', metadata: { category: 'movies' } });
+``` 
+</CodeGroup>
+
 You can also configure the API base URL in the config:

 ```python
@@ -56,6 +56,30 @@ config = {
 }
 ```

+### Configuration Options
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `collection_name` | string | required | Name of the OpenSearch index |
+| `host` | string | required | OpenSearch endpoint URL |
+| `port` | int | 9200 | Port number |
+| `http_auth` | object | None | Authentication credentials (e.g., AWSV4SignerAuth) |
+| `embedding_model_dims` | int | 1536 | Dimension of embedding vectors |
+| `use_ssl` | bool | False | Enable SSL/TLS connection |
+| `verify_certs` | bool | False | Verify SSL certificates |
+| `auto_refresh` | bool | False | Automatically refresh index after insert. OpenSearch refreshes every ~1 second by default, so this is rarely needed. |
+
+<Note>
+  The defaults above match a local OpenSearch instance. The AWS OpenSearch Serverless
+  example earlier on this page intentionally overrides them with `port=443`, `use_ssl=True`,
+  and `verify_certs=True`, which are required when connecting to a Serverless collection.
+</Note>
+
+<Note>
+  For **AWS OpenSearch Serverless**, keep `auto_refresh=False` (the default).
+  The `indices.refresh()` API is not supported on Serverless collections.
+</Note>
+
 ### Add Memories

 ```python
@@ -345,7 +345,7 @@ When evaluating memory systems, keep these considerations in mind:
  <Card title="Research" icon="flask" href="https://mem0.ai/research">
    Published research papers and technical reports
  </Card>
-  <Card title="Blog Post" icon="newspaper" href="https://mem0.ai/blog/new-algorithm">
+  <Card title="Blog Post" icon="newspaper" href="https://mem0.ai/blog/the-token-efficient-memory-algorithm-now-has-temporal-reasoning">
    Detailed writeup of the new algorithm design and results
  </Card>
  <Card title="Platform Migration" icon="arrow-right" href="/migration/platform-v2-to-v3">
@@ -65,6 +65,7 @@ The plugin uses the same shell scripts as Claude Code, Cursor, and Codex — hoo
 | **User prompt** | `UserPromptSubmit` | Searches relevant memories before each message |
 | **Pre-tool** | `PreToolUse` | Blocks MEMORY.md writes, enforces `user_id`/`app_id` on mem0 tools |
 | **Post-tool** | `PostToolUse` | Tracks stats, scans bash errors for related memories |
+| **Stop** | `Stop` | Stores a session summary when the session ends |

 ## Troubleshooting

@@ -64,7 +64,7 @@ Add the Mem0 MCP server directly with a single command:
 npx mcp-add \
  --name mem0-mcp \
  --type http \
-  --url "https://mcp.mem0.ai/mcp" \
+  --url "https://mcp.mem0.ai/mcp/" \
  --clients "claude code"
 ```

@@ -142,7 +142,8 @@ When installed via the plugin marketplace, Mem0 hooks into Claude Code's lifecyc
 | **User prompt** | `UserPromptSubmit` | Searches relevant memories before each message; skips short prompts |
 | **Pre-tool** | `PreToolUse` | Blocks MEMORY.md writes, enforces `user_id`/`app_id` on mem0 tool calls |
 | **Post-tool** | `PostToolUse` | Tracks stats, scans bash errors for related memories |
-| **Pre-compact** | `PreCompact` | Stores a session summary before context compaction |
+| **Stop** | `Stop` | Stores a session summary when the session ends |
+| **Pre-compact** | `PreCompact` | Stores a summary before the context is compacted |

 ## Example Workflow

@@ -41,7 +41,17 @@ Install the full plugin including MCP server, lifecycle hooks, and SDK skill.
   codex plugin marketplace add mem0ai/mem0
   ```

-2. Restart Codex, open the Plugin Directory, browse the **Mem0 Plugins** marketplace, and install **Mem0**.
+2. Install the plugin:
+
+   ```bash
+   codex plugin add mem0@mem0-plugins
+   ```
+
+   Or, in the app: restart Codex, open the Plugin Directory, browse the **Mem0 Plugins** marketplace, and install **Mem0**.
+
+<Note>
+  Step 1 is required for the app UI. Mem0 isn't in OpenAI's curated directory yet, so **without `codex plugin marketplace add`, Mem0 won't appear in the Codex app's Plugin Directory** — searching for it returns nothing. Adding the marketplace surfaces it (under **Created by you**) and makes it installable.
+</Note>

 <Info>
  Do not combine with Option B. The plugin manifest auto-registers the `mem0` MCP server, so adding both will create a duplicate registration.
@@ -49,27 +59,30 @@ Install the full plugin including MCP server, lifecycle hooks, and SDK skill.

 ### Option B — Direct MCP

-The fastest way to connect Codex to Mem0 — no plugin, no marketplace. Add to `~/.codex/config.toml`:
+The fastest way to connect Codex to Mem0 — no plugin, no marketplace. Add the MCP server with a single command:
+
+```bash
+codex mcp add mem0 --url https://mcp.mem0.ai/mcp/ --bearer-token-env-var MEM0_API_KEY
+```
+
+Or add it manually to `~/.codex/config.toml`:

 ```toml
 [mcp_servers.mem0]
-url = "https://mcp.mem0.ai/mcp"
+url = "https://mcp.mem0.ai/mcp/"
 bearer_token_env_var = "MEM0_API_KEY"
 ```

 Make sure `MEM0_API_KEY` is exported in the shell you launch Codex from, then restart Codex.

-<Info>
-  Codex's `codex mcp add` CLI only supports stdio MCP servers. Because Mem0's MCP is HTTP/streamable, you configure it by editing `config.toml` directly (or via the **Plugins → Connect to a custom MCP → Streamable HTTP** UI in the Codex app).
-</Info>
-
 This gives you the MCP tools but not the lifecycle hooks or SDK skill.

 ### Managing the Plugin

 ```bash
 codex plugin marketplace upgrade               # pull latest plugin versions
-codex plugin marketplace remove mem0-plugins   # unregister the marketplace
+codex plugin remove mem0@mem0-plugins          # uninstall the plugin (keeps the marketplace)
+codex plugin marketplace remove mem0-plugins   # unregister the marketplace entirely
 ```

 To update, run `codex plugin marketplace upgrade` to pull the latest from the Mem0 repo.
@@ -112,7 +125,8 @@ When installed via the plugin marketplace, Mem0 hooks into Codex's lifecycle to
 | **User prompt** | `UserPromptSubmit` | Searches relevant memories before each message |
 | **Pre-tool** | `PreToolUse` | Blocks MEMORY.md writes, enforces `user_id`/`app_id` on mem0 tool calls |
 | **Post-tool** | `PostToolUse` | Tracks stats, scans bash errors for related memories |
-| **Pre-compact** | `PreCompact` | Stores a session summary before context compaction |
+| **Stop** | `Stop` | Stores a session summary when the session ends |
+| **Pre-compact** | `PreCompact` | Stores a summary before the context is compacted |

 ## Example Workflow

@@ -47,7 +47,7 @@ The fastest way to get started. Click the link below to install the Mem0 MCP ser
 npx mcp-add \
  --name mem0-mcp \
  --type http \
-  --url "https://mcp.mem0.ai/mcp" \
+  --url "https://mcp.mem0.ai/mcp/" \
  --clients "cursor"
 ```

@@ -110,7 +110,8 @@ When installed via the Cursor Marketplace, Mem0 hooks into Cursor's lifecycle:
 | **User prompt** | `beforeSubmitPrompt` | Searches relevant memories before each message; skips short prompts |
 | **Pre-tool (2 handlers)** | `preToolUse` | Blocks MEMORY.md writes, enforces `user_id`/`app_id` on mem0 tool calls |
 | **Post-tool (2 handlers)** | `postToolUse` | Tracks stats, scans bash errors for related memories |
-| **Pre-compact** | `preCompact` | Stores a session summary before context compaction |
+| **Stop** | `stop` | Stores a session summary when the session ends |
+| **Pre-compact** | `preCompact` | Stores a summary before the context is compacted |

 ## Example Workflow

@@ -100,12 +100,12 @@ This section:
 Initialize both the ElevenLabs and Mem0 clients:

 ```python
-    # Initialize ElevenLabs client
-    client = ElevenLabs(api_key=API_KEY)
+# Initialize ElevenLabs client
+client = ElevenLabs(api_key=API_KEY)

-    # Initialize memory client and tools
-    client_tools = ClientTools()
-    mem0_client = AsyncMemoryClient()
+# Initialize memory client and tools
+client_tools = ClientTools()
+mem0_client = AsyncMemoryClient()
 ```

 Here we:
@@ -118,36 +118,36 @@ Here we:
 Define the two key memory functions that will be registered as tools:

 ```python
-    # Define memory-related functions for the agent
-    async def add_memories(parameters):
-        """Add a message to the memory store"""
-        message = parameters.get("message")
-        await mem0_client.add(
-            messages=message,
-            user_id=USER_ID
-        )
-        return "Memory added successfully"
+# Define memory-related functions for the agent
+async def add_memories(parameters):
+    """Add a message to the memory store"""
+    message = parameters.get("message")
+    await mem0_client.add(
+        messages=message,
+        user_id=USER_ID
+    )
+    return "Memory added successfully"

-    async def retrieve_memories(parameters):
-        """Retrieve relevant memories based on the input message"""
-        message = parameters.get("message")
+async def retrieve_memories(parameters):
+    """Retrieve relevant memories based on the input message"""
+    message = parameters.get("message")

-        # For Platform API, user_id goes in filters
-        filters = {"user_id": USER_ID}
+    # For Platform API, user_id goes in filters
+    filters = {"user_id": USER_ID}

-        # Search for relevant memories using the message as a query
-        results = await mem0_client.search(
-            query=message,
-            filters=filters
-        )
+    # Search for relevant memories using the message as a query
+    results = await mem0_client.search(
+        query=message,
+        filters=filters
+    )

-        # Extract and join the memory texts
-        memories = ' '.join([result["memory"] for result in results.get('results', [])])
-        print("[ Memories ]", memories)
+    # Extract and join the memory texts
+    memories = ' '.join([result["memory"] for result in results.get('results', [])])
+    print("[ Memories ]", memories)

-        if memories:
-            return memories
-        return "No memories found"
+    if memories:
+        return memories
+    return "No memories found"
 ```

 These functions:
@@ -171,9 +171,9 @@ These functions:
 Register the memory functions with the ElevenLabs ClientTools system:

 ```python
-    # Register the memory functions as tools for the agent
-    client_tools.register("addMemories", add_memories, is_async=True)
-    client_tools.register("retrieveMemories", retrieve_memories, is_async=True)
+# Register the memory functions as tools for the agent
+client_tools.register("addMemories", add_memories, is_async=True)
+client_tools.register("retrieveMemories", retrieve_memories, is_async=True)
 ```

 This allows the ElevenLabs agent to:
@@ -186,19 +186,19 @@ This allows the ElevenLabs agent to:
 Configure the conversation with ElevenLabs:

 ```python
-    # Initialize the conversation
-    conversation = Conversation(
-        client,
-        AGENT_ID,
-        # Assume auth is required when API_KEY is set
-        requires_auth=bool(API_KEY),
-        audio_interface=DefaultAudioInterface(),
-        client_tools=client_tools,
-        callback_agent_response=lambda response: print(f"Agent: {response}"),
-        callback_agent_response_correction=lambda original, corrected: print(f"Agent: {original} -> {corrected}"),
-        callback_user_transcript=lambda transcript: print(f"User: {transcript}"),
-        # callback_latency_measurement=lambda latency: print(f"Latency: {latency}ms"),
-    )
+# Initialize the conversation
+conversation = Conversation(
+    client,
+    AGENT_ID,
+    # Assume auth is required when API_KEY is set
+    requires_auth=bool(API_KEY),
+    audio_interface=DefaultAudioInterface(),
+    client_tools=client_tools,
+    callback_agent_response=lambda response: print(f"Agent: {response}"),
+    callback_agent_response_correction=lambda original, corrected: print(f"Agent: {original} -> {corrected}"),
+    callback_user_transcript=lambda transcript: print(f"User: {transcript}"),
+    # callback_latency_measurement=lambda latency: print(f"Latency: {latency}ms"),
+)
 ```

 This sets up the conversation with:
@@ -217,16 +217,16 @@ This sets up the conversation with:
 Start and manage the conversation:

 ```python
-    # Start the conversation
-    print(f"Starting conversation with user_id: {USER_ID}")
-    conversation.start_session()
+# Start the conversation
+print(f"Starting conversation with user_id: {USER_ID}")
+conversation.start_session()

-    # Handle Ctrl+C to gracefully end the session
-    signal.signal(signal.SIGINT, lambda sig, frame: conversation.end_session())
+# Handle Ctrl+C to gracefully end the session
+signal.signal(signal.SIGINT, lambda sig, frame: conversation.end_session())

-    # Wait for the conversation to end and get the conversation ID
-    conversation_id = conversation.wait_for_session_end()
-    print(f"Conversation ID: {conversation_id}")
+# Wait for the conversation to end and get the conversation ID
+conversation_id = conversation.wait_for_session_end()
+print(f"Conversation ID: {conversation_id}")


 if __name__ == '__main__':
@@ -445,4 +445,3 @@ By integrating ElevenLabs Conversational AI with Mem0, you can create voice agen
    Create voice-first AI applications
  </Card>
 </CardGroup>
-
@@ -1,35 +1,42 @@
 ---
 title: Hermes Agent
-description: "Add long-term memory to Hermes agents using Mem0 as a pluggable memory provider with automatic background sync and zero-latency prefetch."
+description: "Add long-term memory to Hermes agents with Mem0, on managed Mem0 Cloud or fully self-hosted (OSS), with automatic background sync and zero-latency prefetch."
 ---

-Add long-term memory to [Hermes Agent](https://github.com/NousResearch/hermes-agent) — a self-improving AI agent CLI by Nous Research. Hermes has a pluggable memory system, and Mem0 is one of the supported providers. Once enabled, Mem0 automatically learns facts from your conversations and surfaces relevant ones before each turn — all without slowing down the chat.
+Add long-term memory to [Hermes Agent](https://github.com/NousResearch/hermes-agent), a self-improving AI agent CLI by Nous Research. Hermes has a pluggable memory system, and Mem0 is one of the supported providers. Once enabled, Mem0 learns facts from your conversations and surfaces relevant ones before each turn, without slowing down the chat.

-## Overview
+You can run Mem0 in two ways:

-Hermes runs a built-in memory system (file-based `MEMORY.md` and `USER.md`) alongside one external provider. When Mem0 is active, it works additively with the built-in system at three key moments in every conversation turn:
+- **Platform mode** (default): managed Mem0 Cloud. Add your API key and you are ready.
+- **OSS mode**: fully self-hosted with your own LLM, embedder, and vector store. No data leaves your machine.

-### 1. Before the Agent Responds (Prefetch)
+## How It Works

-When you send a message, Hermes checks if it already has cached Mem0 search results from the previous turn. If so, those memories are injected into the system prompt so the LLM can see them. This is **zero-latency** — no waiting for an API call.
+Hermes runs a built-in memory system (file-based `MEMORY.md` and `USER.md`) alongside one external provider. When Mem0 is active, it works additively with the built-in system at three points in every conversation turn.

-### 2. After the Agent Responds (Sync)
+### 1. Before the agent responds (prefetch)

-Once the LLM finishes responding, Hermes sends the `(user message, assistant response)` pair to Mem0's API in a **background thread**. Mem0's server-side LLM automatically extracts facts (e.g., "user prefers Python", "user works at Acme Corp") — you don't have to tell it what to remember.
+When you send a message, Hermes checks for cached Mem0 search results from the previous turn. If they exist, those memories are injected into the system prompt so the model can see them. This is zero-latency, with no waiting on an API call.

-### 3. Background Prefetch for Next Turn
+### 2. After the agent responds (sync)

-At the same time as sync, Hermes kicks off a background search on Mem0 to pre-load relevant memories for the next turn. By the time you type your next message, the memories are already cached.
+Once the model finishes, Hermes sends the `(user message, assistant response)` pair to Mem0 in a background thread. Mem0 extracts facts automatically (for example, "user prefers Python" or "user works at Acme Corp"), so you never have to tell it what to remember. Each write is tagged with the gateway channel it came from.
+
+### 3. Background prefetch for the next turn
+
+At the same time, Hermes runs a background search to pre-load relevant memories for your next message. By the time you type, the results are already cached.

 ## Agent Tools

-When Mem0 is active, the LLM gets three extra tools it can call during conversations:
+When Mem0 is active, the model gets five tools it can call during a conversation:

-| Tool | Description |
-|------|-------------|
-| `mem0_profile` | Fetch all stored memories about the user |
-| `mem0_search` | Semantic search through memories (supports optional reranking via `rerank` and `top_k` parameters) |
-| `mem0_conclude` | Store a specific fact verbatim — uses `infer=False` so no server-side LLM extraction happens |
+| Tool | Description | Parameters |
+|------|-------------|------------|
+| `mem0_list` | List all stored memories, for a full overview | `page`, `page_size` (default 100, max 200) |
+| `mem0_search` | Semantic search by meaning, ranked by relevance | `query` (required), `top_k` (default 10, max 50), `rerank` (default `true`, Platform mode only) |
+| `mem0_add` | Store a fact verbatim, with no LLM extraction | `content` (required) |
+| `mem0_update` | Update a memory's text by ID | `memory_id`, `text` (both required) |
+| `mem0_delete` | Delete a memory by ID | `memory_id` (required) |

 ## Installation

@@ -40,17 +47,19 @@ curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scri
 source ~/.bashrc
 ```

-The `mem0ai` Python package is automatically installed when you enable the Mem0 provider — no manual pip install needed.
+The `mem0ai` package is installed automatically when you enable the Mem0 provider, so there is no manual pip step. OSS providers may need extra packages (for example `qdrant-client`, `psycopg2-binary`, or `ollama`), which the setup flow installs for you when you pick them.

-## Setup
+## Platform Setup

-### Option 1: Interactive Setup Wizard (Recommended)
+Platform mode uses managed Mem0 Cloud and is the fastest way to start.
+
+### Option 1: Interactive wizard (recommended)

 ```bash
 hermes memory setup
 ```

-Select **mem0** as the provider and enter your Mem0 API key when prompted. The wizard writes your config to `~/.hermes/mem0.json`.
+Select **mem0**, choose **Platform**, and paste your API key when prompted. The wizard writes the non-secret settings to `~/.hermes/mem0.json` and keeps the key in `~/.hermes/.env`.

 <Note>Get your API key from <a href="https://app.mem0.ai?utm_source=oss&utm_medium=integration-hermes" rel="nofollow">app.mem0.ai</a>.</Note>

@@ -68,33 +77,151 @@ memory:
  provider: mem0
 ```

-That's it — Mem0 runs automatically from this point.
+That's it. Mem0 runs automatically from here.

-## Configuration Options
+## OSS (Self-Hosted) Setup

-Configuration is stored in `~/.hermes/mem0.json`. Values can also be set via environment variables.
+OSS mode runs Mem0 entirely on your own infrastructure: your LLM, your embedder, and your vector store. No data is sent to Mem0 Cloud, and no Mem0 API key is required.

-| Key | Env Variable | Default | Description |
-|-----|-------------|---------|-------------|
-| `api_key` | `MEM0_API_KEY` | — | **Required.** Mem0 Platform API key |
-| `user_id` | `MEM0_USER_ID` | `hermes-user` | User identifier for scoping memories |
-| `agent_id` | `MEM0_AGENT_ID` | `hermes` | Agent identifier |
-| `rerank` | — | `true` | Enable reranking for memory recall |
+### Interactive
+
+```bash
+hermes memory setup
+# Select "mem0", then "Open Source (self-hosted)"
+# Follow the prompts for LLM, embedder, and vector store
+```
+
+### With flags
+
+```bash
+hermes memory setup mem0 --mode oss \
+  --oss-llm openai --oss-llm-key sk-... \
+  --oss-vector qdrant
+```
+
+### Supported providers
+
+| Component | Providers |
+|-----------|-----------|
+| LLM | `openai` (default model `gpt-5-mini`), `ollama` (local, default `llama3.1:8b`) |
+| Embedder | `openai` (default `text-embedding-3-small`), `ollama` (local, default `nomic-embed-text`) |
+| Vector store | `qdrant` (local path or server), `pgvector` |
+
+### Flag reference
+
+| Flag | Description |
+|------|-------------|
+| `--mode` | `platform` or `oss` |
+| `--oss-llm` | LLM provider (`openai` or `ollama`, default `openai`) |
+| `--oss-llm-key` | LLM API key (for `openai`) |
+| `--oss-llm-model` | Override the LLM model |
+| `--oss-llm-url` | LLM base URL (for `ollama` or a custom endpoint) |
+| `--oss-embedder` | Embedder provider (default `openai`) |
+| `--oss-embedder-key` | Embedder API key |
+| `--oss-vector` | Vector store (`qdrant` or `pgvector`, default `qdrant`) |
+| `--oss-vector-path` | Local Qdrant storage path |
+| `--oss-vector-host`, `--oss-vector-port` | PGVector or remote Qdrant host and port |
+| `--oss-vector-user`, `--oss-vector-password`, `--oss-vector-dbname` | PGVector connection details |
+| `--user-id` | Canonical user identifier |
+| `--dry-run` | Preview the resolved config without writing it |
+
+## Switching Modes
+
+You can move between Platform and OSS at any time. Run the setup command again, or edit `~/.hermes/mem0.json` directly.
+
+```bash
+# Platform to OSS
+hermes memory setup mem0 --mode oss --oss-llm-key sk-...
+
+# OSS to Platform
+hermes memory setup mem0 --mode platform --api-key sk-...
+
+# Preview without writing anything
+hermes memory setup mem0 --mode oss --oss-llm-key sk-... --dry-run
+```
+
+A self-hosted `~/.hermes/mem0.json` looks like this:
+
+```json
+{
+  "mode": "oss",
+  "oss": {
+    "llm": {"provider": "openai", "config": {"model": "gpt-5-mini"}},
+    "embedder": {"provider": "openai", "config": {"model": "text-embedding-3-small"}},
+    "vector_store": {"provider": "qdrant", "config": {"path": "~/.hermes/mem0_qdrant"}}
+  }
+}
+```
+
+## Configuration
+
+Behavioral settings live in `~/.hermes/mem0.json` and are written for you by `hermes memory setup`. Only the secret `MEM0_API_KEY` belongs in `~/.hermes/.env`.
+
+| Key | Default | Description |
+|-----|---------|-------------|
+| `mode` | `platform` | `platform` (Mem0 Cloud) or `oss` (self-hosted) |
+| `api_key` | none | Mem0 Platform API key, required in Platform mode. Stored in `.env` as `MEM0_API_KEY` |
+| `user_id` | `hermes-user` | Identifier that scopes memories. See cross-channel behavior below |
+| `agent_id` | `hermes` | Agent identifier attached to writes |
+| `rerank` | `true` | Rerank search results for relevance (Platform mode only) |
+
+### Cross-channel memories
+
+Hermes can run from the CLI and from gateways like Telegram, Slack, and Discord. The `user_id` setting controls how memories are scoped across them:
+
+- **Set a `user_id`** and it applies to every gateway, so one person gets a single merged memory store no matter where they talk to the agent.
+- **Leave it unset** (or at the default `hermes-user`) and each gateway uses its own native id, keeping per-platform memories separate.
+
+Either way, every write is tagged with `metadata.channel` (for example `telegram` or `cli`), so per-channel views are still possible at query time.


 ## Reliability

- **Circuit Breaker** — If Mem0's API fails 5 times in a row, Hermes stops calling it for 2 minutes, then retries. The agent keeps working fine without memory during that time.
- **Non-blocking** — All Mem0 API calls happen in background daemon threads. A slow or failed API call never blocks your conversation.
- **Thread-safe** — The Mem0 client uses lazy initialization with locking, safe for concurrent access.
+- **Circuit breaker**: if Mem0 fails five times in a row, Hermes pauses calls for two minutes, then retries. The agent keeps working without memory during that window. Expected client errors, like a 404 on a missing memory id, do not count toward tripping the breaker.
+- **Non-blocking**: every Mem0 call runs in a background daemon thread, so a slow or failed call never blocks your conversation.
+- **Thread-safe**: the client uses lazy initialization with locking, and the background sync and prefetch threads are guarded so concurrent gateway messages cannot produce duplicate memories.
+
+## Troubleshooting
+
+### "Mem0 temporarily unavailable"
+
+The circuit breaker tripped after five consecutive failures and resets after two minutes.
+
+- **Platform mode**: check your API key and internet connection.
+- **OSS mode**: make sure your vector store (Qdrant or PGVector) is running and reachable.
+
+### OSS: vector store connection refused
+
+```bash
+# Local Qdrant: confirm the storage path is writable
+ls -la ~/.hermes/mem0_qdrant
+
+# Qdrant server: confirm it is reachable
+curl http://localhost:6333/healthz
+
+# PGVector: confirm PostgreSQL is accepting connections
+pg_isready -h localhost -p 5432
+```
+
+### OSS: Ollama not reachable
+
+```bash
+curl http://localhost:11434/api/tags
+```
+
+### Memories not appearing
+
+- `mem0_add` stores text verbatim with no extraction. Ordinary conversation turns are extracted automatically by the background sync.
+- Search is semantic, so try a broader query.
+- Confirm `user_id` is the same across sessions (check `~/.hermes/mem0.json`).

 ## Key Features

-1. **Zero-Latency Recall** — Memories are prefetched in the background and cached, ready before you type
-2. **Server-side Extraction** — Mem0's API automatically extracts and deduplicates facts from each exchange
-3. **Non-blocking** — All API calls run in background daemon threads
-4. **Fault Tolerant** — Circuit breaker ensures the agent works even if Mem0 is temporarily unreachable
-5. **Additive Memory** — Works alongside Hermes' built-in file-based memory system (MEMORY.md, USER.md)
+1. **Two ways to run**: managed Platform or fully self-hosted OSS, switchable at any time.
+2. **Zero-latency recall**: memories are prefetched in the background and cached before you type.
+3. **Automatic extraction**: Mem0 extracts and deduplicates facts from each exchange for you.
+4. **Non-blocking and fault tolerant**: background threads plus a circuit breaker keep the agent responsive even when Mem0 is unreachable.
+5. **Additive memory**: works alongside Hermes' built-in file memory (`MEMORY.md`, `USER.md`).

 <CardGroup cols={2}>
  <Card title="OpenClaw Integration" icon={<svg width="24" height="24" viewBox="0 0 500 500" fill="none" xmlns="http://www.w3.org/2000/svg"><path fill-rule="evenodd" d="m153.5 173.5q24.62 1.46 46 13.5 12.11 8.1 17.5 21.5 0.74 2.45 0.5 5 0.09 0.81 1 1 1.48-4.9 1-10 5.04 10.48 1.5 22-9.81 27.86-35.5 42.5-26.17 14.97-56 19.5-2.77-0.4-2 1 2.86 1.27 6 1 25.64 1.53 48.5-10 0.34 10.08 2 20 1.08 5.76 5 10 1 1.5 0 3-31.11 20.84-68.5 17.5-23.7-5.7-32.5-28.5-4.39-9.18-3.5-19 15.41 6.23 32 4.5-20.68-6.39-39-18-34.81-27.22-12.5-65.5 11.84-14.83 29-23 4.21 7.66 11.5 12.5 3 1 6 0-26.04-34.62-29-78-0.13-8.46 2-16.5 1 6.5 2 13 3.43 39.53 24.5 73 2.03 2.28 4.5 4 0.5-1.25 1-2.5-1.27-6.54-5-12 0.5-0.75 1-1.5 9.72-3.43 20-4 0.55 10.34 8 17.5 1.94 0.74 4 0.5-17.8-64.6 16.5-122 0.98-1.79 1.5 0-28.21 56.64-13.5 118 1.08 1.43 2.5 0.5 2.21-4.98 2-10.5z" fill="currentColor"/><path fill-rule="evenodd" d="m454.5 97.5q-1.33 11.18-8.5 20-21.81 26.28-55.5 32-1.11-0.2-2 0.5 2.31 2.82 5.5 4.5 1 2 0 4-9.56 11.3-19.5 20 19.71-8.72 31-27 2.68-0.43 5 1-14.24 30.97-48 36.5-9.93 1.71-20 1.5-6.8-0.48-13 1 5.81 6.92 14 11-10.78 16.03-27 26.5 27.16-7.4 38-33.5 4.34 1.35 9 1-9.08 23.84-33 33.5-18.45 6.41-38 7 22.59 8.92 45-1 12.05-5.52 24-11 9.01-1.79 17 2.5 5.28-4.38 11-8 12.8-6.07 27-5 0 0.5 0 1-19.34 2.69-34 15.5 0.5 0.25 1 0.5 17.79-8.09 36-15 2.71-0.79 5-2 2.5-1 5-2 5.53-4.04 11-8 11.7-4.18 24-6.5 7.78-1.36 15 1.5-2.97 18.45-13.5 34-34.92 49.37-94.5 62.5-59.27 12.45-108-23-15.53-12.52-21.5-31.5-2.47-14.26 4-27-3.15 24.41 14 42-4.92-10.28-7-22-1.97-17.63 7-33 47.28-69.5 125.5-100 15.86-3.42 32-5.5 18.63-1.47 37 1.5z" fill="currentColor"/><path fill-rule="evenodd" d="m231.5 238.5q1.31-0.2 2 1-3.13 28.62 15 51-16.25 6.75-27-7.5-1-1-2 0 14.73 29.34 46 18.5 1.79 0.52 0 1.5-37.63 16.82-50.5-22.5-5.1-26.48 16.5-42z" fill="currentColor"/><path fill-rule="evenodd" d="m203.5 266.5q1.31-0.2 2 1-2.48 22.08 12 39-6.99 1.35-14 0.5 4.59 4.08 10 7-8.71 0.28-14.5-6.5-16.98-22.76 4.5-41z" fill="currentColor"/><path fill-rule="evenodd" d="m58.5 284.5q9.6-2.17 14.5 6 5.15 14.18-1 28-11.05-13.14-27.5-17.5 5.15-9.9 14-16.5z" fill="currentColor"/><path fill-rule="evenodd" d="m56.5 313.5q3.43 5.43 8 10-4.88 0.44-8 4-1.11-0.2-2 0.5 28.91 1.65 38 28.5 0.45 3.16-1 6-11.02-7.01-23-12.5-4.75-3.75-9.5-7.5 1.47 7.42 7 13 8.34 27.18 32 43 0.99 2.41-1.5 3.5-40.25 5.58-66.5-25.5-15.67-22.01-8-48 10.46-23.87 34.5-15z" fill="currentColor"/><path fill-rule="evenodd" d="m198.5 319.5q1.44 0.68 2.5 2 2.41 8.23 6 16 1.2 2.64-0.5 5-30.65 21.41-68 18.5-25.16-6.17-32.5-30.5 6.96 4.99 15.5 6.5 8.99 0.75 18 0.5 16.25 2.38 32-2.5 15.9-3.94 27-15.5z" fill="currentColor"/><path fill-rule="evenodd" d="m239.5 342.5q7.02-0.25 14 0.5 4.46 1.06 8 3.5-5.2 2.35-10 5.5-3.88 4.65-9 7.5-9.89-3.09-9.5-13 2.36-3.63 6.5-4z" fill="currentColor"/><path fill-rule="evenodd" d="m214.5 349.5q5.96 7.2 13.5 13 1 1 0 2-28.58 23.34-65.5 20.5-18.15-4.24-27.5-19.5 1.13 0.94 2.5 1.5 14.7 1.42 29-1.5 26.57-0.52 48-16z" fill="currentColor"/><path fill-rule="evenodd" d="m302.5 373.5q0.21 2.44-2 3.5-28.69 7.6-50.5-12.5-0.06-6.71 6.5-9 4.45-0.75 9-1 22.26 2.27 37 19z" fill="currentColor"/><path fill-rule="evenodd" d="m232.5 365.5q17.6 6.19 10.5 23-10.6 10.42-25.5 11.5-25.94 3.21-49-9 36.75-1.65 64-25.5z" fill="currentColor"/><path fill-rule="evenodd" d="m113.5 367.5q7.7-0.01 9.5 7-9.69 7.19-18.5 15.5-7.23 5.76-5.5-3.5 3.12-12.84 14.5-19z" fill="currentColor"/><path fill-rule="evenodd" d="m126.5 380.5q7.88-0.4 12 6.5-8.5 7.25-17 14.5-5.62-12.55 5-21z" fill="currentColor"/><path fill-rule="evenodd" d="m283.5 385.5q3.22 2.95 7 5.5 2.8 4.03 6 7.5 0.42 2.77-2 4-15.5-9.75-31-19.5-1.79-0.98 0-1.5 9.96 2.49 20 4z" fill="currentColor"/></svg>} href="/integrations/openclaw">
@@ -62,7 +62,8 @@ def add(
    filters: dict = None,
    output_format: str = None,  # ❌ REMOVED
    version: str = None         # ❌ REMOVED
-) -> Union[List[dict], dict]
+) -> Union[List[dict], dict]:
+    ...
 ```

 #### v1.0.0  Signature
@@ -76,7 +77,8 @@ def add(
    metadata: dict = None,
    filters: dict = None,
    infer: bool = True          # ✅ NEW: Control memory inference
-) -> dict  # Always returns dict with "results" key
+) -> dict:  # Always returns dict with "results" key
+    ...
 ```

 #### Changes Summary
@@ -147,7 +149,8 @@ def search(
    filters: dict = None,       # Basic key-value only
    output_format: str = None,  # ❌ REMOVED
    version: str = None         # ❌ REMOVED
-) -> Union[List[dict], dict]
+) -> Union[List[dict], dict]:
+    ...
 ```

 #### v1.0.0  Signature
@@ -161,7 +164,8 @@ def search(
    limit: int = 100,
    filters: dict = None,       # ✅ ENHANCED: Advanced operators
    rerank: bool = True         # ✅ NEW: Reranking support
-) -> dict  # Always returns dict with "results" key
+) -> dict:  # Always returns dict with "results" key
+    ...
 ```

 #### Enhanced Filtering
@@ -216,7 +220,8 @@ def get_all(
    filters: dict = None,
    output_format: str = None,  # ❌ REMOVED
    version: str = None         # ❌ REMOVED
-) -> Union[List[dict], dict]
+) -> Union[List[dict], dict]:
+    ...
 ```

 #### v1.0.0  Signature
@@ -227,7 +232,8 @@ def get_all(
    agent_id: str = None,
    run_id: str = None,
    filters: dict = None        # ✅ ENHANCED: Advanced operators
-) -> dict  # Always returns dict with "results" key
+) -> dict:  # Always returns dict with "results" key
+    ...
 ```

 ### update() Method
@@ -239,7 +245,8 @@ def update(
    self,
    memory_id: str,
    data: str
-) -> dict
+) -> dict:
+    ...
 ```

 ### delete() Method
@@ -250,7 +257,8 @@ def update(
 def delete(
    self,
    memory_id: str
-) -> dict
+) -> dict:
+    ...
 ```

 ### delete_all() Method
@@ -344,7 +352,7 @@ config = {
 ### New Configuration Options

 #### Reranker Configuration
-```python
+```text
 # Cohere reranker
 "reranker": {
    "provider": "cohere",
@@ -563,4 +571,4 @@ results = m.search(

 <Info>
 Use this reference to systematically update your codebase. Test each change thoroughly before deploying to production.
-</Info>
+</Info>
@@ -117,7 +117,7 @@ results_without_criteria = client.search(
 ### Compare Results

 ### Search Results (with Criteria)
-```python
+```text
 [
    {"memory": "User feels refreshed and ready to take on anything on a beautiful sunny day", "score": 0.666, ...},
    {"memory": "User finally has time to draw something after a long time", "score": 0.616, ...},
@@ -128,7 +128,7 @@ results_without_criteria = client.search(
 ```

 ### Search Results (without Criteria)
-```python
+```text
 [
    {"memory": "User is happy today", "score": 0.607, ...},
    {"memory": "User feels refreshed and ready to take on anything on a beautiful sunny day", "score": 0.512, ...},
@@ -190,7 +190,7 @@ messages = [
 client.add(messages, user_id='alice')
 ```

-```python Memories with categories
+```text Memories with categories
 # Following categories will be created for the memories added
 Sometimes draws and sketches in free time (hobbies)
 Is quite athletic (sports)
@@ -108,10 +108,10 @@ print(response)

 ```javascript JavaScript
 // Basic Export request
-const filters = {"user_id": "alice"};
+const basicFilters = {"user_id": "alice"};
 const response = await client.createMemoryExport({
    schema: json_schema,
-    filters: filters
+    filters: basicFilters
 });

 // Export with custom instructions and additional filters
@@ -124,16 +124,16 @@ const export_instructions = `
 `;

 // For create operation, using only user_id filter as requested
-const filters = {
+const exportFilters = {
    "AND": [
        {"user_id": "alex"},
        {"created_at": {"gte": "2024-01-01"}}
    ]
-}
+};

 const responseWithInstructions = await client.createMemoryExport({
    schema: json_schema,
-    filters: filters,
+    filters: exportFilters,
    exportInstructions: export_instructions
 });

@@ -1,6 +1,6 @@
 {
  "name": "mem0",
-  "version": "0.2.10",
+  "version": "0.2.11",
  "description": "Persistent memory for Claude Code. Remembers decisions, patterns, and preferences across sessions.",
  "author": {
    "name": "Mem0",
@@ -1,6 +1,6 @@
 {
  "name": "mem0",
-  "version": "0.2.10",
+  "version": "0.2.11",
  "description": "Persistent memory for Codex. Remembers decisions, patterns, and preferences across sessions.",
  "author": {
    "name": "Mem0",
@@ -1,6 +1,6 @@
 {
  "name": "mem0",
-  "version": "0.2.10",
+  "version": "0.2.11",
  "description": "Mem0 memory layer for AI applications. Add persistent memory, personalization, and semantic search using the Mem0 Platform MCP server.",
  "author": {
    "name": "Mem0",
@@ -2,6 +2,18 @@

 All notable changes to the Mem0 plugin will be documented in this file.

+## 0.2.11 — Session-summary metadata fix + rerank auto-injected context by default
+
+> Versions: Claude Code / Cursor / Codex `0.2.11`; Antigravity `0.1.3`. All four editors share `scripts/`, so the fix below applies to every editor.
+
+### Fixed
+
+- **`files_touched` was double-JSON-encoded in session summaries (`scripts/capture_session_summary.py`):** the Stop-hook summary set `metadata["files_touched"] = json.dumps(files[:20])` — a pre-serialized JSON string — and then serialized the whole request body again with `json.dumps(body)`. The stored memory therefore carried an escaped string blob (`"[\"mem0/memory/main.py\", \"src/client/index.ts\"]"`) instead of a real array, so file paths surfaced as backslash- and slash-heavy escaped text when those memories were returned by `search_memories`/`get_memories` and shown in Claude Code, Cursor, Codex, and Antigravity. The fix stores the list directly (`metadata["files_touched"] = files[:20]`) so the body is encoded exactly once. New `tests/test_capture_session_summary.py` asserts the posted body contains a JSON array and no escaped-string artifact.
+
+### Changed
+
+- **Auto-injected memory context is now reranked by default (`scripts/_search.py`, `scripts/file_context.py`, `scripts/on_bash_output.sh`, `scripts/on_user_prompt.sh`):** the REST search endpoint does not rerank when `rerank` is omitted, so hook-injected context (file-context, bash-error lookup, session-resume prefetch) was ordered by raw vector similarity and the single most relevant memory could fall outside the injected `top_k` window. A new `should_rerank()` helper turns reranking on for every auto-injection path; the extra ~150–200 ms stays within the hook's curl budget. Opt out with `MEM0_RERANK=0` (also accepts `false`/`no`/`off`). (#5690)
+
 ## 0.2.10 — Accurate per-editor telemetry attribution

 ### Fixed
@@ -1,7 +1,7 @@
 {
  "id": "mem0",
  "name": "mem0",
-  "version": "0.1.2",
+  "version": "0.1.3",
  "description": "Persistent semantic memory for Antigravity agents. Cross-session, user-level recall via the Mem0 Platform MCP server. 16 slash commands, lifecycle hooks for auto-capture and metadata enforcement.",
  "author": { "name": "Mem0", "email": "support@mem0.ai" },
  "publisher": "mem0ai",
@@ -173,7 +173,7 @@ def store_summary(
    if branch:
        metadata["branch"] = branch
    if files:
-        metadata["files_touched"] = json.dumps(files[:20])
+        metadata["files_touched"] = files[:20]

    body = {
        "messages": [{"role": "user", "content": summary_prompt}],
@@ -0,0 +1,79 @@
+"""Regression tests for capture_session_summary.py request body construction.
+
+Guards against the double-JSON-encoding bug where ``files_touched`` was stored
+as a pre-serialized JSON string and then encoded a second time with the rest of
+the request body — surfacing as escaped, slash-heavy blobs in the memories shown
+inside Claude Code / Cursor / Codex / Antigravity (all four editors share this
+script).
+"""
+
+from __future__ import annotations
+
+import json
+
+
+class _FakeResp:
+    status = 200
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *_):
+        return False
+
+
+def _capture_request_body(monkeypatch):
+    """Patch urlopen so store_summary posts nowhere; capture the request body."""
+    import capture_session_summary as css
+
+    captured: dict = {}
+
+    def fake_urlopen(req, timeout=0):
+        captured["raw"] = req.data.decode("utf-8")
+        captured["body"] = json.loads(captured["raw"])
+        return _FakeResp()
+
+    monkeypatch.setattr(css.urllib.request, "urlopen", fake_urlopen)
+    return captured, css
+
+
+def test_files_touched_is_json_array_not_double_encoded(monkeypatch):
+    """files_touched must be a real JSON array, encoded exactly once."""
+    captured, css = _capture_request_body(monkeypatch)
+    files = ["mem0/memory/main.py", "src/client/index.ts"]
+
+    css.store_summary(
+        api_key="test-key",
+        summary_prompt="did some work",
+        user_id="u1",
+        session_id="s1",
+        project_id="p1",
+        branch="main",
+        files=files,
+    )
+
+    files_touched = captured["body"]["metadata"]["files_touched"]
+    assert isinstance(files_touched, list), (
+        "files_touched must be a JSON array, not a double-encoded string; "
+        f"got {type(files_touched).__name__}: {files_touched!r}"
+    )
+    assert files_touched == files
+    # The file paths must not appear as an escaped JSON string inside the body.
+    assert '\\"' not in captured["raw"]
+
+
+def test_files_touched_omitted_when_no_files(monkeypatch):
+    """No files touched -> no files_touched key (unchanged behaviour)."""
+    captured, css = _capture_request_body(monkeypatch)
+
+    css.store_summary(
+        api_key="test-key",
+        summary_prompt="did some work",
+        user_id="u1",
+        session_id="s1",
+        project_id="p1",
+        branch="main",
+        files=[],
+    )
+
+    assert "files_touched" not in captured["body"]["metadata"]
@@ -1,6 +1,6 @@
 {
  "name": "mem0ai",
-  "version": "3.0.9",
+  "version": "3.0.11",
  "description": "The Memory Layer For Your AI Apps",
  "main": "./dist/index.js",
  "module": "./dist/index.mjs",
@@ -169,7 +169,9 @@ export interface PaginatedMemories {

 export interface ProjectResponse {
  customInstructions?: string;
-  customCategories?: string[];
+  // The API returns category objects (`[{ "<name>": "<description>" }]`),
+  // not bare strings (see issue #5738).
+  customCategories?: custom_categories[];
  [key: string]: any;
 }

@@ -99,4 +99,50 @@ describe("camelToSnakeKeys / snakeToCamelKeys", () => {
      });
    });
  });
+
+  describe("user-controlled customCategories names (issue #5738)", () => {
+    it("converts the outer key but leaves multi-word category names on write", () => {
+      expect(
+        camelToSnakeKeys({
+          customCategories: [
+            { work_life_balance: "desc" },
+            { AIResearch: "desc" },
+          ],
+        }),
+      ).toEqual({
+        // outer SDK key is snake_cased, user-defined category names are not
+        custom_categories: [
+          { work_life_balance: "desc" },
+          { AIResearch: "desc" },
+        ],
+      });
+    });
+
+    it("converts the outer key but leaves category names verbatim on read", () => {
+      expect(
+        snakeToCamelKeys({
+          custom_categories: [
+            { work_life_balance: "desc" },
+            { AIResearch: "desc" },
+          ],
+        }),
+      ).toEqual({
+        customCategories: [
+          { work_life_balance: "desc" },
+          { AIResearch: "desc" },
+        ],
+      });
+    });
+
+    it("round-trips category names losslessly (write then read)", () => {
+      const customCategories = [
+        { work_life_balance: "balance between work and life" },
+        { AIResearch: "artificial intelligence research" },
+      ];
+      const roundTripped = snakeToCamelKeys(
+        camelToSnakeKeys({ customCategories }),
+      );
+      expect(roundTripped.customCategories).toEqual(customCategories);
+    });
+  });
 });
@@ -29,6 +29,11 @@ const OPAQUE_VALUE_KEYS = new Set([
  "metadata",
  "structuredDataSchema",
  "structured_data_schema",
+  // Custom-category names are user-controlled keys (`[{ "<name>": "<desc>" }]`).
+  // Listed in both casings so they round-trip verbatim in both directions
+  // (see issue #5738; same class as `metadata`/`structuredDataSchema`).
+  "customCategories",
+  "custom_categories",
 ]);

 /**
@@ -1,43 +0,0 @@
-import { OpenAILLM } from "./openai";
-import { LLMConfig, Message } from "../types";
-import { LLMResponse } from "./base";
-
-export class TogetherLLM extends OpenAILLM {
-  constructor(config: LLMConfig) {
-    const apiKey = config.apiKey || process.env.TOGETHER_API_KEY;
-    if (!apiKey) {
-      throw new Error("Together API key is required");
-    }
-    super({
-      ...config,
-      apiKey,
-      baseURL:
-        config.baseURL ||
-        process.env.TOGETHER_API_BASE ||
-        "https://api.together.xyz/v1",
-      model: config.model || "mistralai/Mixtral-8x7B-Instruct-v0.1",
-    });
-  }
-
-  async generateResponse(
-    messages: Message[],
-    responseFormat?: { type: string },
-    tools?: any[],
-  ): Promise<string | LLMResponse> {
-    try {
-      return await super.generateResponse(messages, responseFormat, tools);
-    } catch (err) {
-      const message = err instanceof Error ? err.message : String(err);
-      throw new Error(`Together LLM failed: ${message}`);
-    }
-  }
-
-  async generateChat(messages: Message[]): Promise<LLMResponse> {
-    try {
-      return await super.generateChat(messages);
-    } catch (err) {
-      const message = err instanceof Error ? err.message : String(err);
-      throw new Error(`Together LLM failed: ${message}`);
-    }
-  }
-}
@@ -296,6 +296,44 @@ export class Memory {
    return filters;
  }

+  private _normalizeEntityText(value: string): string {
+    return value.trim().toLowerCase().replace(/\s+/g, " ");
+  }
+
+  private async _existingEntitiesByText(
+    entityStore: VectorStore,
+    filters: Record<string, any>,
+  ): Promise<Map<string, { id: string; payload: Record<string, any> }>> {
+    const rowsByText = new Map<
+      string,
+      { id: string; payload: Record<string, any> }
+    >();
+    let rows: Array<{ id: string; payload: Record<string, any> }> = [];
+    try {
+      const listed = await entityStore.list(filters, 10000);
+      rows = (
+        Array.isArray(listed) && Array.isArray(listed[0])
+          ? listed[0]
+          : (listed as any)
+      ) as Array<{ id: string; payload: Record<string, any> }>;
+    } catch (e) {
+      console.debug(
+        `Exact entity lookup failed, falling back to semantic dedup: ${e}`,
+      );
+      return rowsByText;
+    }
+
+    for (const row of rows) {
+      const text = row.payload?.data;
+      if (typeof text !== "string") continue;
+      const key = this._normalizeEntityText(text);
+      if (key && !rowsByText.has(key)) {
+        rowsByText.set(key, row);
+      }
+    }
+    return rowsByText;
+  }
+
  /**
   * Remove `memoryId` from every entity record scoped to `filters`.
   * If an entity's `linkedMemoryIds` becomes empty after removal, the
@@ -393,6 +431,10 @@ export class Memory {
      if (entities.length === 0) return;

      const entityStore = await this.getEntityStore();
+      const exactMatches = await this._existingEntitiesByText(
+        entityStore,
+        filters,
+      );

      for (const entity of entities) {
        try {
@@ -409,12 +451,21 @@ export class Memory {
            score?: number;
            payload: Record<string, any>;
          }> = [];
-          try {
-            matches = await entityStore.search(entityVec, 1, filters);
-          } catch {}
+          const exactMatch = exactMatches.get(
+            this._normalizeEntityText(entity.text),
+          );
+          if (!exactMatch) {
+            try {
+              matches = await entityStore.search(entityVec, 1, filters);
+            } catch {}
+          }

-          if (matches.length > 0 && (matches[0].score ?? 0) >= 0.95) {
-            const match = matches[0];
+          const semanticMatch =
+            matches.length > 0 && (matches[0].score ?? 0) >= 0.95
+              ? matches[0]
+              : undefined;
+          const match = exactMatch ?? semanticMatch;
+          if (match) {
            const payload = match.payload || {};
            const linked = new Set<string>(
              Array.isArray(payload.linkedMemoryIds)
@@ -1062,6 +1113,10 @@ export class Memory {

        if (valid.length > 0) {
          const entityStore = await this.getEntityStore();
+          const exactMatches = await this._existingEntitiesByText(
+            entityStore,
+            filters,
+          );

          // 7c: Search for existing entities one by one (no batch search)
          const toInsertVectors: number[][] = [];
@@ -1077,13 +1132,20 @@ export class Memory {
              score?: number;
              payload: Record<string, any>;
            }> = [];
-            try {
-              matches = await entityStore.search(entityVec, 1, filters);
-            } catch {}
+            const exactMatch = exactMatches.get(key);
+            if (!exactMatch) {
+              try {
+                matches = await entityStore.search(entityVec, 1, filters);
+              } catch {}
+            }

-            if (matches.length > 0 && (matches[0].score ?? 0) >= 0.95) {
+            const semanticMatch =
+              matches.length > 0 && (matches[0].score ?? 0) >= 0.95
+                ? matches[0]
+                : undefined;
+            const match = exactMatch ?? semanticMatch;
+            if (match) {
              // Update existing entity
-              const match = matches[0];
              const payload = match.payload || {};
              const linked = new Set<string>(payload.linkedMemoryIds ?? []);
              for (const mid of memoryIds) linked.add(mid);
@@ -4,8 +4,8 @@
 * Extracts four types of entities from text:
 * - PROPER:   Capitalized multi-word sequences (person names, places, brands)
 * - QUOTED:   Text in single or double quotes (titles, specific terms)
- * - COMPOUND: Multi-word noun phrases with specific modifiers (e.g., "machine learning")
- * - NOUN:     Single nouns from circumstantial compound patterns
+ * - TOPIC:    Multi-word noun/topic phrases with specific modifiers
+ * - IDENTIFIER: Dotted technical identifiers such as person.properties.email
 *
 * Uses the `compromise` npm package for NLP-based extraction when available.
 * Falls back to regex-only extraction if `compromise` is not installed.
@@ -196,6 +196,25 @@ const NON_SPECIFIC_ADJ: Set<string> = new Set([
  "final",
  "initial",
  "side",
+  "top",
+]);
+
+/** Leading words that frame a topic but are not part of the topic itself. */
+const TOPIC_PREFIX_WORDS: Set<string> = new Set([
+  "a",
+  "an",
+  "the",
+  "my",
+  "your",
+  "our",
+  "their",
+  "his",
+  "her",
+  "its",
+  "this",
+  "that",
+  "these",
+  "those",
 ]);

 /** Generic tail words to strip from compound entities. */
@@ -267,6 +286,25 @@ const GENERIC_CAPS: Set<string> = new Set([
  "disadvantages",
 ]);

+/** Generic role/title words that should not become single-token entities. */
+const GENERIC_SINGLE_ENTITY_TERMS: Set<string> = new Set([
+  "user",
+  "assistant",
+  "agent",
+  "customer",
+  "client",
+  "person",
+  "people",
+  "human",
+  "memory",
+  "message",
+  "conversation",
+  "chat",
+  "session",
+  "system",
+  "top",
+]);
+
 /** Markdown/formatting markers to skip during extraction. */
 const FORMATTING_MARKERS: Set<string> = new Set([
  "*",
@@ -287,7 +325,7 @@ const FORMATTING_MARKERS: Set<string> = new Set([
 // ---------------------------------------------------------------------------

 export interface ExtractedEntity {
-  type: "PROPER" | "QUOTED" | "COMPOUND" | "NOUN";
+  type: "PROPER" | "QUOTED" | "TOPIC" | "IDENTIFIER";
  text: string;
 }

@@ -338,32 +376,96 @@ function stripGenericEnding(words: string[]): string[] {
  return words;
 }

-/**
- * Determine if a token position is at the start of a sentence.
- * Simple heuristic: index 0, or preceded by sentence-ending punctuation
- * or formatting markers.
- */
-function isSentenceStart(
-  tokens: string[],
-  idx: number,
-  rawText: string,
-): boolean {
-  if (idx === 0) {
-    return true;
+function stripTopicPrefix(words: string[]): string[] {
+  let start = 0;
+  while (
+    start < words.length &&
+    TOPIC_PREFIX_WORDS.has(words[start].toLowerCase())
+  ) {
+    start++;
  }
-  const prev = tokens[idx - 1];
-  if (/[.!?:]$/.test(prev)) {
+  return words.slice(start);
+}
+
+function cleanToken(token: string): string {
+  return token.replace(/^[^\w.]+|[^\w.]+$/g, "");
+}
+
+function tokenize(text: string): string[] {
+  return (
+    text.match(
+      /[A-Za-z_][\w-]*(?:\.[A-Za-z_][\w-]*)*|\d[\d,]*(?:\.\d+)?|[,:;.!?&]/g,
+    ) ?? []
+  );
+}
+
+function isCapitalized(token: string): boolean {
+  return /^[A-Z]/.test(token) && /[A-Za-z]/.test(token);
+}
+
+function hasInternalCapOrDigit(token: string): boolean {
+  return (
+    /\d/.test(token) ||
+    /[A-Z]/.test(token.slice(1)) ||
+    /^[A-Z]{2,}$/.test(token)
+  );
+}
+
+function isBadSingleNameToken(token: string): boolean {
+  const lower = token.toLowerCase();
+  return GENERIC_SINGLE_ENTITY_TERMS.has(lower) || GENERIC_CAPS.has(lower);
+}
+
+function looksLikeMetricCount(token: string): boolean {
+  return /^\d[\d,]*(?:\.\d+)?$/.test(token);
+}
+
+function isMetricListContext(tokens: string[], idx: number): boolean {
+  const prev = idx > 0 ? tokens[idx - 1] : "";
+  const next = idx + 1 < tokens.length ? tokens[idx + 1] : "";
+  return [":", ",", ";"].includes(prev) || [",", ";"].includes(next);
+}
+
+function isSentenceStart(tokens: string[], idx: number): boolean {
+  if (idx === 0) return true;
+  return (
+    [".", "!", "?", ":"].includes(tokens[idx - 1]) ||
+    FORMATTING_MARKERS.has(tokens[idx - 1])
+  );
+}
+
+function isListItemNameToken(tokens: string[], idx: number): boolean {
+  const token = cleanToken(tokens[idx]);
+  if (!isCapitalized(token) || isBadSingleNameToken(token)) return false;
+  const next = idx + 1 < tokens.length ? cleanToken(tokens[idx + 1]) : "";
+  if (!looksLikeMetricCount(next)) return false;
+  return (
+    isMetricListContext(tokens, idx) || isMetricListContext(tokens, idx + 1)
+  );
+}
+
+function isNameToken(tokens: string[], idx: number): boolean {
+  const token = cleanToken(tokens[idx]);
+  if (!token || !isCapitalized(token) || isBadSingleNameToken(token))
+    return false;
+  if (hasInternalCapOrDigit(token) || isListItemNameToken(tokens, idx))
    return true;
-  }
-  if (FORMATTING_MARKERS.has(prev)) {
-    return true;
-  }
-  // Check for newline before this token in the raw text
-  const tokenStart = rawText.indexOf(tokens[idx]);
-  if (tokenStart > 0 && rawText.charAt(tokenStart - 1) === "\n") {
-    return true;
-  }
-  return false;
+  return !isSentenceStart(tokens, idx);
+}
+
+function cleanEntityText(text: string): string {
+  return text
+    .replace(/^\*+\s*|\s*\*+$/g, "")
+    .replace(/\s*:+$/g, "")
+    .replace(/^\d+\s*\.\s*/, "")
+    .replace(/\s+\d[\d,]*(?:\.\d+)?$/g, "")
+    .replace(/[.,;!?]+$/, "")
+    .trim()
+    .replace(/\s+/g, " ");
+}
+
+function isCoordinatedNameTopic(text: string): boolean {
+  return /\b[A-Z][\w-]+\s+and\s+[A-Z][\w-]+\b/.test(text);
 }

 // ---------------------------------------------------------------------------
@@ -397,86 +499,79 @@ function extractQuoted(text: string): ExtractedEntity[] {
 }

 /**
- * Extract proper noun sequences using capitalization heuristics.
- * Finds sequences of capitalized words that are not at sentence starts.
+ * Extract dotted technical identifiers such as person.properties.email.
+ */
+function extractIdentifiers(text: string): ExtractedEntity[] {
+  const entities: ExtractedEntity[] = [];
+  const identifierRe = /\b[A-Za-z_][\w-]*(?:\.[A-Za-z_][\w-]*)+\b/g;
+  let match: RegExpExecArray | null;
+  while ((match = identifierRe.exec(text)) !== null) {
+    entities.push({ type: "IDENTIFIER", text: match[0] });
+  }
+  return entities;
+}
+
+/**
+ * Extract proper names using capitalization and list-context heuristics.
 */
 function extractProper(text: string): ExtractedEntity[] {
  const entities: ExtractedEntity[] = [];
-  // Tokenize on whitespace, preserving order
-  const tokens = text.split(/\s+/).filter(Boolean);
-  const functionWords = new Set([
-    "'s",
-    "of",
-    "the",
-    "in",
-    "and",
-    "for",
-    "at",
-    "is",
-  ]);
+  const tokens = tokenize(text);
+  const innerConnectors = new Set(["of", "the", "in", "for", "at"]);

  let i = 0;
  while (i < tokens.length) {
-    const tok = tokens[i];
-    // Skip formatting markers
-    if (FORMATTING_MARKERS.has(tok)) {
+    const token = cleanToken(tokens[i]);
+    const next = i + 1 < tokens.length ? tokens[i + 1] : "";
+    const afterNext = i + 2 < tokens.length ? cleanToken(tokens[i + 2]) : "";
+    if (
+      token &&
+      next === "&" &&
+      afterNext &&
+      isCapitalized(token) &&
+      isCapitalized(afterNext) &&
+      !isBadSingleNameToken(token) &&
+      !isBadSingleNameToken(afterNext)
+    ) {
+      entities.push({
+        type: "PROPER",
+        text: cleanEntityText(`${token} & ${afterNext}`),
+      });
+      i += 3;
+      continue;
+    }
+
+    if (!isNameToken(tokens, i)) {
      i++;
      continue;
    }

-    const isLabel = i + 1 < tokens.length && tokens[i + 1] === ":";
-    const isCap =
-      tok.length > 0 &&
-      tok.charAt(0) === tok.charAt(0).toUpperCase() &&
-      /[A-Z]/.test(tok.charAt(0));
-
-    if (isCap && !isLabel) {
-      const seq: Array<{ token: string; idx: number }> = [
-        { token: tok, idx: i },
-      ];
-      let j = i + 1;
-      while (j < tokens.length) {
-        const t = tokens[j];
-        const tIsCap =
-          t.length > 0 &&
-          t.charAt(0) === t.charAt(0).toUpperCase() &&
-          /[A-Z]/.test(t.charAt(0));
-        if (tIsCap || functionWords.has(t.toLowerCase())) {
-          seq.push({ token: t, idx: j });
-          j++;
-        } else {
-          break;
-        }
+    const span = [cleanToken(tokens[i])];
+    let j = i + 1;
+    while (j < tokens.length) {
+      const current = cleanToken(tokens[j]);
+      if (isNameToken(tokens, j)) {
+        span.push(current);
+        j++;
+        continue;
      }
-
-      // Strip trailing function words
-      while (
-        seq.length > 0 &&
-        functionWords.has(seq[seq.length - 1].token.toLowerCase())
+      if (
+        innerConnectors.has(current.toLowerCase()) &&
+        j + 1 < tokens.length &&
+        isNameToken(tokens, j + 1)
      ) {
-        seq.pop();
+        span.push(current, cleanToken(tokens[j + 1]));
+        j += 2;
+        continue;
      }
-
-      if (seq.length > 0) {
-        // Check for at least one mid-sentence capitalized word
-        const hasMidCap = seq.some(({ token, idx: tokenIdx }) => {
-          const isCapWord =
-            /[A-Z]/.test(token.charAt(0)) &&
-            !functionWords.has(token.toLowerCase());
-          return isCapWord && !isSentenceStart(tokens, tokenIdx, text);
-        });
-
-        if (hasMidCap) {
-          const phrase = seq.map((s) => s.token).join(" ");
-          if (phrase.length > 2) {
-            entities.push({ type: "PROPER", text: phrase });
-          }
-        }
-      }
-      i = j;
-    } else {
-      i++;
+      break;
    }
+
+    const phrase = cleanEntityText(span.join(" "));
+    if (phrase.length > 2) {
+      entities.push({ type: "PROPER", text: phrase });
+    }
+    i = Math.max(j, i + 1);
  }

  return entities;
@@ -484,7 +579,7 @@ function extractProper(text: string): ExtractedEntity[] {

 /**
 * Extract compound noun phrases using the `compromise` NLP library.
- * Returns COMPOUND and NOUN entities derived from noun chunks.
+ * Returns TOPIC entities derived from noun chunks.
 */
 function extractCompoundsWithNlp(text: string): ExtractedEntity[] {
  if (!nlp) {
@@ -524,12 +619,12 @@ function extractCompoundsWithNlp(text: string): ExtractedEntity[] {
    const filtered = words.filter(
      (w) => !NON_SPECIFIC_ADJ.has(w.toLowerCase()),
    );
-    const cleaned = stripGenericEnding(filtered);
+    const cleaned = stripGenericEnding(stripTopicPrefix(filtered));

    if (cleaned.length >= 2) {
-      const phrase = cleaned.join(" ");
+      const phrase = cleanEntityText(cleaned.join(" "));
      if (phrase.length > 3) {
-        entities.push({ type: "COMPOUND", text: phrase });
+        entities.push({ type: "TOPIC", text: phrase });
      }
    }
  }
@@ -547,7 +642,7 @@ function extractCompoundsRegex(text: string): ExtractedEntity[] {
  // Multi-word sequences with at least one non-trivial word
  // Match sequences like "machine learning", "New York", "data science"
  const compoundRe =
-    /\b([A-Z][a-z]+(?:\s+(?:of|and|the|for|in)\s+)?[A-Z][a-z]+(?:\s+[A-Z][a-z]+)*)\b/g;
+    /\b([A-Z][a-z]+(?:\s+(?:of|the|for|in)\s+)?[A-Z][a-z]+(?:\s+[A-Z][a-z]+)*)\b/g;
  let match: RegExpExecArray | null;
  while ((match = compoundRe.exec(text)) !== null) {
    const phrase = match[1].trim();
@@ -558,9 +653,12 @@ function extractCompoundsRegex(text: string): ExtractedEntity[] {
        const filtered = words.filter(
          (w) => !NON_SPECIFIC_ADJ.has(w.toLowerCase()),
        );
-        const cleaned = stripGenericEnding(filtered);
+        const cleaned = stripGenericEnding(stripTopicPrefix(filtered));
        if (cleaned.length >= 2) {
-          entities.push({ type: "COMPOUND", text: cleaned.join(" ") });
+          entities.push({
+            type: "TOPIC",
+            text: cleanEntityText(cleaned.join(" ")),
+          });
        }
      }
    }
@@ -590,9 +688,12 @@ function extractCompoundsRegex(text: string): ExtractedEntity[] {
          const filtered = words.filter(
            (w) => !NON_SPECIFIC_ADJ.has(w.toLowerCase()),
          );
-          const cleaned = stripGenericEnding(filtered);
+          const cleaned = stripGenericEnding(stripTopicPrefix(filtered));
          if (cleaned.length >= 2) {
-            entities.push({ type: "COMPOUND", text: cleaned.join(" ") });
+            entities.push({
+              type: "TOPIC",
+              text: cleanEntityText(cleaned.join(" ")),
+            });
          }
        }
      }
@@ -614,9 +715,9 @@ function extractCompoundsRegex(text: string): ExtractedEntity[] {
 *
 * Entity types (in priority order for deduplication):
 *   PROPER   - Capitalized multi-word sequences not at sentence start
- *   COMPOUND - Multi-word noun phrases with specific modifiers
+ *   IDENTIFIER - Dotted technical identifiers
 *   QUOTED   - Text in single or double quotes (min 3 chars)
- *   NOUN     - Single nouns from circumstantial patterns
+ *   TOPIC    - Multi-word noun/topic phrases with specific modifiers
 *
 * @param text - Input text to extract entities from.
 * @returns Deduplicated list of extracted entities.
@@ -630,7 +731,10 @@ export function extractEntities(text: string): ExtractedEntity[] {
  // 2. PROPER entities (capitalization heuristics)
  raw.push(...extractProper(text));

-  // 3. COMPOUND entities (NLP or regex fallback)
+  // 3. IDENTIFIER entities
+  raw.push(...extractIdentifiers(text));
+
+  // 4. TOPIC entities (NLP or regex fallback)
  if (nlp) {
    raw.push(...extractCompoundsWithNlp(text));
  } else {
@@ -654,19 +758,17 @@ export function extractEntities(text: string): ExtractedEntity[] {
  const cleaned: ExtractedEntity[] = [];
  for (const entity of deduped) {
    let txt = entity.text.trim();
-    // Strip leading/trailing asterisks
-    txt = txt.replace(/^\*+\s*|\s*\*+$/g, "");
-    // Strip trailing colons
-    txt = txt.replace(/\s*:+$/, "");
-    // Strip leading numbered list markers
-    txt = txt.replace(/^\d+\s*\.\s*/, "");
-    // Strip trailing sentence punctuation (".", ",", ";", "!", "?") — otherwise
-    // "Paris." and "Paris" produce different embeddings and break entity dedup.
-    txt = txt.replace(/[.,;!?]+$/, "").trim();
+    txt = cleanEntityText(txt);

    if (!txt || txt.length <= 2 || hasArtifacts(txt)) {
      continue;
    }
+    if (
+      entity.type === "TOPIC" &&
+      (/^\d/.test(txt) || isCoordinatedNameTopic(txt))
+    ) {
+      continue;
+    }

    // Filter generic single-word PROPER nouns
    if (
@@ -680,12 +782,12 @@ export function extractEntities(text: string): ExtractedEntity[] {
    cleaned.push({ type: entity.type, text: txt });
  }

-  // Keep best type per entity (PROPER > COMPOUND > QUOTED > NOUN)
+  // Keep best type per entity (PROPER > IDENTIFIER > QUOTED > TOPIC)
  const typePriority: Record<string, number> = {
    PROPER: 0,
-    COMPOUND: 1,
+    IDENTIFIER: 1,
    QUOTED: 2,
-    NOUN: 3,
+    TOPIC: 3,
  };
  const best = new Map<string, ExtractedEntity>();
  for (const entity of cleaned) {
@@ -700,14 +802,17 @@ export function extractEntities(text: string): ExtractedEntity[] {
  }
  const bestEntities = Array.from(best.values());

-  // Remove entities that are substrings of longer entities
-  const allLower = bestEntities.map((e) => e.text.toLowerCase());
+  // Remove entities that are token substrings of longer entities.
  return bestEntities.filter(
    (entity) =>
-      !allLower.some(
+      !bestEntities.some(
        (other) =>
-          entity.text.toLowerCase() !== other &&
-          other.includes(entity.text.toLowerCase()),
+          entity.text.toLowerCase() !== other.text.toLowerCase() &&
+          (typePriority[entity.type] ?? 99) >=
+            (typePriority[other.type] ?? 99) &&
+          new RegExp(
+            `(^|\\s)${entity.text.toLowerCase().replace(/[.*+?^${}()|[\]\\]/g, "\\$&")}(\\s|$)`,
+          ).test(other.text.toLowerCase()),
      ),
  );
 }
@@ -22,7 +22,6 @@ import { RedisDB } from "../vector_stores/redis";
 import { OllamaLLM } from "../llms/ollama";
 import { LMStudioLLM } from "../llms/lmstudio";
 import { DeepSeekLLM } from "../llms/deepseek";
-import { TogetherLLM } from "../llms/together";
 import { SupabaseDB } from "../vector_stores/supabase";
 import { SQLiteManager } from "../storage/SQLiteManager";
 import { MemoryHistoryManager } from "../storage/MemoryHistoryManager";
@@ -86,8 +85,6 @@ export class LLMFactory {
        return new LangchainLLM(config);
      case "deepseek":
        return new DeepSeekLLM(config);
-      case "together":
-        return new TogetherLLM(config);
      default:
        throw new Error(`Unsupported LLM provider: ${provider}`);
    }
@@ -31,9 +31,11 @@ const parse_vision_messages = async (messages: Message[]) => {
        typeof message.content === "object" &&
        message.content.type === "image_url"
      ) {
-        const description = await get_image_description(
-          message.content.image_url.url,
-        );
+        const imageUrl = message.content.image_url?.url;
+        if (!imageUrl) {
+          throw new Error("image_url content part is missing image_url.url");
+        }
+        const description = await get_image_description(imageUrl);
        new_message.content =
          typeof description === "string"
            ? description
@@ -0,0 +1,45 @@
+import { extractEntities } from "../src/utils/entity_extraction";
+
+describe("extractEntities", () => {
+  it("handles product lists, coordinated names, and identifiers", () => {
+    const text =
+      "User reported top inbound integration pages: OpenClaw 25,443, " +
+      "Claude Code 8,916, Codex 2,573, Dify 656. " +
+      "User compared Cartesia and Deepgram. " +
+      "The email field for Mem0 lives at person.properties.email. " +
+      "The qwen endpoint uses person.properties.email. " +
+      "Johnson & Johnson was mentioned. " +
+      "Glasses around my window. " +
+      "On 2026-05-27 there were 90 days of stats.";
+
+    const entityTexts = new Set(
+      extractEntities(text).map((entity) => entity.text),
+    );
+    const normalized = new Set(
+      [...entityTexts].map((entityText) => entityText.toLowerCase()),
+    );
+
+    for (const expected of [
+      "OpenClaw",
+      "Claude Code",
+      "Codex",
+      "Dify",
+      "Cartesia",
+      "Deepgram",
+      "Mem0",
+    ]) {
+      expect(entityTexts.has(expected)).toBe(true);
+    }
+    expect(entityTexts.has("person.properties.email")).toBe(true);
+    expect(entityTexts.has("qwen endpoint")).toBe(true);
+    expect(entityTexts.has("Johnson & Johnson")).toBe(true);
+    expect(entityTexts.has("Johnson")).toBe(false);
+    expect(normalized.has("top")).toBe(false);
+    expect(normalized.has("glasses")).toBe(false);
+    expect(entityTexts.has("Cartesia and Deepgram")).toBe(false);
+    expect(entityTexts.has("Claude Code 8,916")).toBe(false);
+    for (const rejected of ["8,916", "2,573", "656", "2026-05-27", "90"]) {
+      expect(entityTexts.has(rejected)).toBe(false);
+    }
+  });
+});
@@ -1,124 +0,0 @@
-/// <reference types="jest" />
-/**
- * Together LLM — unit tests (mocked OpenAI).
- */
-
-import { TogetherLLM } from "../src/llms/together";
-
-const mockCreate = jest.fn();
-
-jest.mock("openai", () => {
-  return jest.fn().mockImplementation(() => ({
-    chat: { completions: { create: mockCreate } },
-  }));
-});
-
-describe("TogetherLLM (unit)", () => {
-  beforeEach(() => mockCreate.mockClear());
-
-  it("throws when no API key is provided", () => {
-    const original = process.env.TOGETHER_API_KEY;
-    delete process.env.TOGETHER_API_KEY;
-    try {
-      expect(() => new TogetherLLM({})).toThrow("Together API key is required");
-    } finally {
-      if (original !== undefined) process.env.TOGETHER_API_KEY = original;
-    }
-  });
-
-  it("generateResponse() returns a text response", async () => {
-    mockCreate.mockResolvedValueOnce({
-      choices: [
-        {
-          message: {
-            content: "Hello, world!",
-            role: "assistant",
-            tool_calls: null,
-          },
-        },
-      ],
-    });
-
-    const llm = new TogetherLLM({ apiKey: "test-key" });
-    const result = await llm.generateResponse([
-      { role: "user", content: "Hi" },
-    ]);
-
-    expect(mockCreate).toHaveBeenCalledTimes(1);
-    expect(result).toBe("Hello, world!");
-  });
-
-  it("generateResponse() handles tool calls", async () => {
-    mockCreate.mockResolvedValueOnce({
-      choices: [
-        {
-          message: {
-            content: "",
-            role: "assistant",
-            tool_calls: [
-              {
-                function: {
-                  name: "get_weather",
-                  arguments: '{"city": "London"}',
-                },
-              },
-            ],
-          },
-        },
-      ],
-    });
-
-    const llm = new TogetherLLM({ apiKey: "test-key" });
-    const result = await llm.generateResponse(
-      [{ role: "user", content: "What is the weather?" }],
-      undefined,
-      [{ type: "function", function: { name: "get_weather" } }],
-    );
-
-    expect(result).toEqual({
-      content: "",
-      role: "assistant",
-      toolCalls: [{ name: "get_weather", arguments: '{"city": "London"}' }],
-    });
-  });
-
-  it("generateResponse() wraps API errors with a clear message", async () => {
-    mockCreate.mockRejectedValueOnce(new Error("Connection refused"));
-
-    const llm = new TogetherLLM({ apiKey: "test-key" });
-
-    await expect(
-      llm.generateResponse([{ role: "user", content: "Hi" }]),
-    ).rejects.toThrow("Together LLM failed: Connection refused");
-  });
-
-  it("generateChat() returns LLMResponse shape", async () => {
-    mockCreate.mockResolvedValueOnce({
-      choices: [
-        {
-          message: { content: "I can help with that.", role: "assistant" },
-        },
-      ],
-    });
-
-    const llm = new TogetherLLM({ apiKey: "test-key" });
-    const result = await llm.generateChat([
-      { role: "user", content: "Help me" },
-    ]);
-
-    expect(result).toEqual({
-      content: "I can help with that.",
-      role: "assistant",
-    });
-  });
-
-  it("generateChat() wraps API errors with a clear message", async () => {
-    mockCreate.mockRejectedValueOnce(new Error("Timeout"));
-
-    const llm = new TogetherLLM({ apiKey: "test-key" });
-
-    await expect(
-      llm.generateChat([{ role: "user", content: "Hi" }]),
-    ).rejects.toThrow("Together LLM failed: Timeout");
-  });
-});
@@ -18,6 +18,12 @@ class OpenSearchConfig(BaseModel):
        "RequestsHttpConnection", description="Connection class for OpenSearch"
    )
    pool_maxsize: int = Field(20, description="Maximum number of connections in the pool")
+    auto_refresh: bool = Field(
+        False,
+        description="Automatically refresh index after insert operations to make documents "
+        "immediately searchable. Disabled by default for OpenSearch Serverless compatibility. "
+        "OpenSearch automatically refreshes indices every ~1 second, so most users don't need this.",
+    )

    @model_validator(mode="before")
    @classmethod
@@ -1,6 +1,7 @@
 import json
+import logging
 import os
-from typing import Dict, List, Optional
+from typing import Dict, List, Optional, Union

 try:
    from groq import Groq
@@ -11,6 +12,8 @@ from mem0.configs.llms.base import BaseLlmConfig
 from mem0.llms.base import LLMBase
 from mem0.memory.utils import extract_json

+logger = logging.getLogger(__name__)
+

 class GroqLLM(LLMBase):
    def __init__(self, config: Optional[BaseLlmConfig] = None):
@@ -22,6 +25,24 @@ class GroqLLM(LLMBase):
        api_key = self.config.api_key or os.getenv("GROQ_API_KEY")
        self.client = Groq(api_key=api_key)

+    @staticmethod
+    def _supports_json_mode(model: Optional[Union[str, Dict]]) -> bool:
+        """
+        Groq's compound agentic systems (e.g. ``groq/compound``, ``groq/compound-mini``)
+        do not support the JSON ``response_format`` and return empty or non-JSON content
+        when it is requested. See https://console.groq.com/docs/structured-outputs.
+
+        Non-string models (the config allows a dict) are assumed to support JSON mode,
+        preserving prior behavior.
+        """
+        if not isinstance(model, str):
+            return True
+        # Strip provider prefixes (e.g. "groq/compound-mini" -> "compound-mini"),
+        # mirroring the _is_reasoning_model heuristic in LLMBase, so the match
+        # targets the compound family rather than any name containing the substring.
+        base_model = model.lower().rsplit("/", 1)[-1]
+        return not base_model.startswith("compound")
+
    def _parse_response(self, response, tools):
        """
        Process the response based on whether tools are used or not.
@@ -79,7 +100,17 @@ class GroqLLM(LLMBase):
            "top_p": self.config.top_p,
        }
        if response_format:
-            params["response_format"] = response_format
+            requests_json = isinstance(response_format, dict) and response_format.get("type") in (
+                "json_object",
+                "json_schema",
+            )
+            if requests_json and not self._supports_json_mode(self.config.model):
+                logger.debug(
+                    f"Model '{self.config.model}' does not support JSON response_format; "
+                    "sending the request without it."
+                )
+            else:
+                params["response_format"] = response_format
        if tools:
            params["tools"] = tools
            params["tool_choice"] = tool_choice
@@ -81,6 +81,14 @@ warnings.filterwarnings("ignore", category=DeprecationWarning, message=".*swigva
 logger = logging.getLogger(__name__)


+def _vector_store_list_rows(listed):
+    if isinstance(listed, (list, tuple)) and listed and isinstance(listed[0], list):
+        return listed[0]
+    if isinstance(listed, (list, tuple)):
+        return listed
+    return []
+
+
 # Fields that hold runtime auth/connection objects and must be preserved.
 # These are non-serializable objects (e.g. AWSV4SignerAuth, RequestsHttpConnection)
 # needed by clients like OpenSearch — not sensitive strings to redact.
@@ -499,22 +507,49 @@ class Memory(MemoryBase):
            )
        return self._entity_store

+    @staticmethod
+    def _normalize_entity_text(value: str) -> str:
+        return " ".join(value.strip().lower().split())
+
+    def _existing_entities_by_text(self, filters):
+        """Return existing entity rows keyed by normalized payload data."""
+        try:
+            listed = self.entity_store.list(filters=filters, top_k=10000)
+        except Exception as e:
+            logger.debug(f"Exact entity lookup failed, falling back to semantic dedup: {e}")
+            return {}
+
+        rows_by_text = {}
+        for row in _vector_store_list_rows(listed):
+            payload = getattr(row, "payload", None) or {}
+            text = payload.get("data")
+            if not isinstance(text, str):
+                continue
+            normalized = self._normalize_entity_text(text)
+            if normalized and normalized not in rows_by_text:
+                rows_by_text[normalized] = row
+        return rows_by_text
+
    def _upsert_entity(self, entity_text, entity_type, memory_id, filters):
        """Upsert an entity into the entity store, linking it to a memory."""
        try:
            entity_embedding = self.embedding_model.embed(entity_text, "add")
            search_filters = {k: v for k, v in filters.items() if k in ("user_id", "agent_id", "run_id") and v}
+            exact_match = self._existing_entities_by_text(search_filters).get(self._normalize_entity_text(entity_text))

-            existing = self.entity_store.search(
-                query=entity_text,
-                vectors=entity_embedding,
-                top_k=1,
-                filters=search_filters,
-            )
+            existing = []
+            if exact_match is None:
+                existing = self.entity_store.search(
+                    query=entity_text,
+                    vectors=entity_embedding,
+                    top_k=1,
+                    filters=search_filters,
+                )

-            if existing and existing[0].score >= 0.95:
+            semantic_match = existing[0] if existing and existing[0].score >= 0.95 else None
+            match = exact_match or semantic_match
+            if match:
                # Update existing entity's linked_memory_ids
-                match = existing[0]
                payload = match.payload or {}
                linked_ids = payload.get("linked_memory_ids", [])
                if memory_id not in linked_ids:
@@ -609,7 +644,7 @@ class Memory(MemoryBase):
                return
            seen = set()
            for entity_type, entity_text in entities:
-                key = entity_text.strip().lower()
+                key = self._normalize_entity_text(entity_text)
                if not key or key in seen:
                    continue
                seen.add(key)
@@ -971,7 +1006,7 @@ class Memory(MemoryBase):
            for idx, (memory_id, text, embedding, payload) in enumerate(records):
                entities = all_entities[idx] if idx < len(all_entities) else []
                for entity_type, entity_text in entities:
-                    key = entity_text.strip().lower()
+                    key = self._normalize_entity_text(entity_text)
                    if key in global_entities:
                        global_entities[key][2].add(memory_id)
                    else:
@@ -1009,6 +1044,7 @@ class Memory(MemoryBase):
                if valid:
                    valid_indices, valid_keys = zip(*valid)
                    valid_vectors = [entity_embeddings[i] for i in valid_indices]
+                    exact_matches = self._existing_entities_by_text(search_filters)

                    # 7c: Batch search for existing entities
                    valid_texts = [global_entities[k][1] for k in valid_keys]
@@ -1024,10 +1060,12 @@ class Memory(MemoryBase):
                    for j, key in enumerate(valid_keys):
                        entity_type, entity_text, memory_ids = global_entities[key]
                        matches = existing_matches[j] if j < len(existing_matches) else []
+                        exact_match = exact_matches.get(key)

-                        if matches and matches[0].score >= 0.95:
+                        semantic_match = matches[0] if matches and matches[0].score >= 0.95 else None
+                        match = exact_match or semantic_match
+                        if match:
                            # Update existing entity
-                            match = matches[0]
                            payload = match.payload or {}
                            linked = set(payload.get("linked_memory_ids", []))
                            linked |= memory_ids
@@ -1603,7 +1641,7 @@ class Memory(MemoryBase):
        seen = set()
        deduped = []
        for entity_type, entity_text in query_entities[:8]:
-            key = entity_text.strip().lower()
+            key = self._normalize_entity_text(entity_text)
            if key and key not in seen:
                seen.add(key)
                deduped.append((entity_type, entity_text))
@@ -2045,22 +2083,51 @@ class AsyncMemory(MemoryBase):
            )
        return self._entity_store

+    @staticmethod
+    def _normalize_entity_text(value: str) -> str:
+        return " ".join(value.strip().lower().split())
+
+    def _existing_entities_by_text(self, filters):
+        """Return existing entity rows keyed by normalized payload data."""
+        try:
+            listed = self.entity_store.list(filters=filters, top_k=10000)
+        except Exception as e:
+            logger.debug(f"Exact entity lookup failed, falling back to semantic dedup: {e}")
+            return {}
+
+        rows_by_text = {}
+        for row in _vector_store_list_rows(listed):
+            payload = getattr(row, "payload", None) or {}
+            text = payload.get("data")
+            if not isinstance(text, str):
+                continue
+            normalized = self._normalize_entity_text(text)
+            if normalized and normalized not in rows_by_text:
+                rows_by_text[normalized] = row
+        return rows_by_text
+
    async def _upsert_entity_async(self, entity_text, entity_type, memory_id, filters):
        """Async variant of `_upsert_entity` — per-entity search-then-update-or-insert."""
        try:
            entity_embedding = await asyncio.to_thread(self.embedding_model.embed, entity_text, "add")
            search_filters = {k: v for k, v in filters.items() if k in ("user_id", "agent_id", "run_id") and v}
+            exact_match = (
+                await asyncio.to_thread(self._existing_entities_by_text, search_filters)
+            ).get(self._normalize_entity_text(entity_text))

-            existing = await asyncio.to_thread(
-                self.entity_store.search,
-                query=entity_text,
-                vectors=entity_embedding,
-                top_k=1,
-                filters=search_filters,
-            )
+            existing = []
+            if exact_match is None:
+                existing = await asyncio.to_thread(
+                    self.entity_store.search,
+                    query=entity_text,
+                    vectors=entity_embedding,
+                    top_k=1,
+                    filters=search_filters,
+                )

-            if existing and existing[0].score >= 0.95:
-                match = existing[0]
+            semantic_match = existing[0] if existing and existing[0].score >= 0.95 else None
+            match = exact_match or semantic_match
+            if match:
                payload = match.payload or {}
                linked_ids = payload.get("linked_memory_ids", [])
                if memory_id not in linked_ids:
@@ -2163,7 +2230,7 @@ class AsyncMemory(MemoryBase):
                return
            seen = set()
            for entity_type, entity_text in entities:
-                key = entity_text.strip().lower()
+                key = self._normalize_entity_text(entity_text)
                if not key or key in seen:
                    continue
                seen.add(key)
@@ -2512,7 +2579,7 @@ class AsyncMemory(MemoryBase):
            for idx, (memory_id, text, embedding, payload) in enumerate(records):
                entities = all_entities[idx] if idx < len(all_entities) else []
                for entity_type, entity_text in entities:
-                    key = entity_text.strip().lower()
+                    key = self._normalize_entity_text(entity_text)
                    if key in global_entities:
                        global_entities[key][2].add(memory_id)
                    else:
@@ -2547,6 +2614,7 @@ class AsyncMemory(MemoryBase):
                if valid:
                    valid_indices, valid_keys = zip(*valid)
                    valid_vectors = [entity_embeddings[i] for i in valid_indices]
+                    exact_matches = await asyncio.to_thread(self._existing_entities_by_text, search_filters)

                    # 7c: Batch search for existing entities
                    valid_texts = [global_entities[k][1] for k in valid_keys]
@@ -2563,9 +2631,11 @@ class AsyncMemory(MemoryBase):
                    for j, key in enumerate(valid_keys):
                        entity_type, entity_text, memory_ids = global_entities[key]
                        matches = existing_matches[j] if j < len(existing_matches) else []
+                        exact_match = exact_matches.get(key)

-                        if matches and matches[0].score >= 0.95:
-                            match = matches[0]
+                        semantic_match = matches[0] if matches and matches[0].score >= 0.95 else None
+                        match = exact_match or semantic_match
+                        if match:
                            payload = match.payload or {}
                            linked = set(payload.get("linked_memory_ids", []))
                            linked |= memory_ids
@@ -3137,7 +3207,7 @@ class AsyncMemory(MemoryBase):
        seen = set()
        deduped = []
        for entity_type, entity_text in query_entities[:8]:
-            key = entity_text.strip().lower()
+            key = self._normalize_entity_text(entity_text)
            if key and key not in seen:
                seen.add(key)
                deduped.append((entity_type, entity_text))
@@ -206,7 +206,10 @@ def parse_vision_messages(messages, llm=None, vision_details="auto"):
        elif isinstance(content, dict) and content.get("type") == "image_url":
            if llm is None:
                continue
-            image_url = content["image_url"]["url"]
+            image_url_obj = content.get("image_url")
+            image_url = image_url_obj.get("url") if isinstance(image_url_obj, dict) else None
+            if not image_url:
+                raise ValueError("image_url content part is missing image_url.url")
            try:
                description = get_image_description(image_url, llm, vision_details)
                returned_messages.append({"role": role, "content": description})
@@ -1,3 +1,4 @@
+import logging
 import os
 from typing import List, Dict, Any

@@ -9,6 +10,8 @@ try:
 except ImportError:
    COHERE_AVAILABLE = False

+logger = logging.getLogger(__name__)
+

 class CohereReranker(BaseReranker):
    """Cohere-based reranker implementation."""
@@ -78,8 +81,9 @@ class CohereReranker(BaseReranker):
                
            return reranked_docs

-        except Exception:
+        except Exception as e:
            # Fallback to original order if reranking fails
+            logger.warning("Cohere reranking failed, falling back to original order: %s", e)
            for doc in documents:
                doc['rerank_score'] = 0.0
            final_top_k = top_k or self.config.top_k
@@ -1,3 +1,4 @@
+import logging
 from typing import List, Dict, Any, Union
 import numpy as np

@@ -12,6 +13,8 @@ try:
 except ImportError:
    TRANSFORMERS_AVAILABLE = False

+logger = logging.getLogger(__name__)
+

 class HuggingFaceReranker(BaseReranker):
    """HuggingFace Transformers based reranker implementation."""
@@ -139,8 +142,9 @@ class HuggingFaceReranker(BaseReranker):

            return reranked_docs

-        except Exception:
+        except Exception as e:
            # Fallback to original order if reranking fails
+            logger.warning("HuggingFace reranking failed, falling back to original order: %s", e)
            for doc in documents:
                doc['rerank_score'] = 0.0
            final_top_k = top_k or self.config.top_k
@@ -1,3 +1,4 @@
+import logging
 import re
 from typing import Any, Dict, List, Union

@@ -6,6 +7,8 @@ from mem0.configs.rerankers.llm import LLMRerankerConfig
 from mem0.reranker.base import BaseReranker
 from mem0.utils.factory import LlmFactory

+logger = logging.getLogger(__name__)
+

 class LLMReranker(BaseReranker):
    """LLM-based reranker implementation."""
@@ -151,8 +154,9 @@ class LLMReranker(BaseReranker):
                scored_doc['rerank_score'] = score
                scored_docs.append(scored_doc)

-            except Exception:
+            except Exception as e:
                # Fallback: assign neutral score if scoring fails
+                logger.warning("LLM reranking failed for a document, assigning neutral score: %s", e)
                scored_doc = doc.copy()
                scored_doc['rerank_score'] = 0.5
                scored_docs.append(scored_doc)
@@ -1,3 +1,4 @@
+import logging
 from typing import List, Dict, Any, Union
 import numpy as np

@@ -11,6 +12,8 @@ try:
 except ImportError:
    SENTENCE_TRANSFORMERS_AVAILABLE = False

+logger = logging.getLogger(__name__)
+

 class SentenceTransformerReranker(BaseReranker):
    """Sentence Transformer based reranker implementation."""
@@ -102,8 +105,9 @@ class SentenceTransformerReranker(BaseReranker):
                
            return reranked_docs

-        except Exception:
+        except Exception as e:
            # Fallback to original order if reranking fails
+            logger.warning("SentenceTransformer reranking failed, falling back to original order: %s", e)
            for doc in documents:
                doc['rerank_score'] = 0.0
            final_top_k = top_k or self.config.top_k
@@ -1,3 +1,4 @@
+import logging
 import os
 from typing import List, Dict, Any

@@ -9,6 +10,8 @@ try:
 except ImportError:
    ZERO_ENTROPY_AVAILABLE = False

+logger = logging.getLogger(__name__)
+

 class ZeroEntropyReranker(BaseReranker):
    """Zero Entropy-based reranker implementation."""
@@ -89,8 +92,9 @@ class ZeroEntropyReranker(BaseReranker):
                
            return reranked_docs

-        except Exception:
+        except Exception as e:
            # Fallback to original order if reranking fails
+            logger.warning("Zero Entropy reranking failed, falling back to original order: %s", e)
            for doc in documents:
                doc['rerank_score'] = 0.0
            final_top_k = top_k or self.config.top_k
@@ -1,82 +1,352 @@
 """
 Entity extraction from text using spaCy NLP.

-Extracts four types of entities from text:
+Extracts three types of entities from a spaCy-processed document:
 - **Proper nouns**: Capitalized multi-word sequences (person names, places, brands)
 - **Quoted text**: Text in single or double quotes (titles, specific terms)
 - **Noun compounds**: Multi-word noun phrases with specific modifiers (e.g., "machine learning")
- **Noun fallback**: Single nouns from circumstantial compound patterns

 Public API:
-    extract_entities(text: str) -> List[Tuple[str, str]]
+    ``extract_entities(text)`` accepts a string and owns spaCy model loading.
+    ``extract_entities_batch(texts)`` uses ``nlp.pipe`` for batched extraction.

-Internal:
-    _extract_entities_from_doc(doc) -> List[Tuple[str, str]]
+Returns:
+    List of ``(entity_type, entity_text)`` tuples where entity_type is one of
+    PROPER, QUOTED, TOPIC, or IDENTIFIER. Returns ``[]`` if spaCy is unavailable.
 """

 from __future__ import annotations

-import logging
+from dataclasses import dataclass
 import re
-from typing import List, Tuple

-logger = logging.getLogger(__name__)
+
+@dataclass(frozen=True)
+class _EntityCandidate:
+    entity_type: str
+    text: str
+    source: str
+    start: int
+    end: int
+    confidence: float
+    priority: int
+

 # Words that are too generic to be useful as entity heads
 _GENERIC_HEADS = {
-    "thing", "stuff", "way", "time", "experience", "situation", "case",
-    "fact", "matter", "issue", "idea", "thought", "feeling", "place",
-    "area", "part", "kind", "type", "sort", "lot", "bit", "day", "year",
-    "week", "month", "moment", "instance", "example", "technique",
-    "method", "approach", "process", "step", "tool", "result", "outcome",
-    "goal", "task", "item", "topic", "scale", "size", "level", "degree",
-    "amount", "number", "style", "look", "color", "colour", "shape",
-    "form", "piece", "section", "side", "end", "edge", "surface", "point",
+    "thing",
+    "stuff",
+    "way",
+    "time",
+    "experience",
+    "situation",
+    "case",
+    "fact",
+    "matter",
+    "issue",
+    "idea",
+    "thought",
+    "feeling",
+    "place",
+    "area",
+    "part",
+    "kind",
+    "type",
+    "sort",
+    "lot",
+    "bit",
+    "day",
+    "year",
+    "week",
+    "month",
+    "moment",
+    "instance",
+    "example",
+    "technique",
+    "method",
+    "approach",
+    "process",
+    "step",
+    "tool",
+    "result",
+    "outcome",
+    "goal",
+    "task",
+    "item",
+    "topic",
+    "scale",
+    "size",
+    "level",
+    "degree",
+    "amount",
+    "number",
+    "style",
+    "look",
+    "color",
+    "colour",
+    "shape",
+    "form",
+    "piece",
+    "section",
+    "side",
+    "end",
+    "edge",
+    "surface",
+    "point",
+}
+
+# Entity labels emitted by spaCy that are usually safe to treat as named
+# entities. Numeric and temporal labels are intentionally excluded.
+_ACCEPTED_NER_LABELS = {
+    "PERSON",
+    "ORG",
+    "GPE",
+    "LOC",
+    "FAC",
+    "PRODUCT",
+    "WORK_OF_ART",
+    "EVENT",
+    "NORP",
+    "LAW",
+    "LANGUAGE",
+}
+
+_REJECTED_NER_LABELS = {
+    "DATE",
+    "TIME",
+    "CARDINAL",
+    "ORDINAL",
+    "QUANTITY",
+    "MONEY",
+    "PERCENT",
+}
+
+# Generic role words and title-cased English words that should not become
+# single-token named entities just because spaCy tagged them as PROPN.
+_GENERIC_SINGLE_ENTITY_TERMS = {
+    "user",
+    "assistant",
+    "agent",
+    "customer",
+    "client",
+    "person",
+    "people",
+    "human",
+    "memory",
+    "message",
+    "conversation",
+    "chat",
+    "session",
+    "system",
+    "top",
 }

 # Modifiers that describe circumstance, not content
 _CIRCUMSTANTIAL_MODS = {
-    "solo", "individual", "team", "group", "joint", "collaborative",
-    "first", "last", "next", "previous", "final", "initial", "main", "side",
+    "solo",
+    "individual",
+    "team",
+    "group",
+    "joint",
+    "collaborative",
+    "first",
+    "last",
+    "next",
+    "previous",
+    "final",
+    "initial",
+    "main",
+    "side",
+    "top",
 }

 # Adjectives too vague to make a compound entity specific
 _NON_SPECIFIC_ADJ = {
-    "many", "few", "several", "some", "any", "all", "most", "more",
-    "less", "much", "little", "enough", "various", "numerous", "multiple",
-    "countless", "great", "good", "bad", "nice", "terrible", "awful",
-    "awesome", "amazing", "wonderful", "horrible", "excellent", "poor",
-    "best", "worst", "fine", "okay", "new", "old", "recent", "past",
-    "future", "current", "previous", "next", "last", "first", "latest",
-    "early", "late", "former", "modern", "ancient", "big", "small",
-    "large", "tiny", "huge", "enormous", "long", "short", "tall", "high",
-    "low", "wide", "narrow", "thick", "thin", "deep", "shallow",
-    "similar", "different", "same", "other", "another", "such", "certain",
-    "important", "main", "major", "minor", "key", "primary", "real",
-    "actual", "true", "whole", "entire", "full", "complete", "total",
-    "basic", "simple", "interesting", "boring", "exciting", "special",
-    "particular", "general", "common", "unique", "rare", "typical",
-    "usual", "normal", "regular", "possible", "likely", "potential",
-    "available", "necessary", "only", "solo", "individual", "team",
-    "group", "joint", "collaborative", "final", "initial", "side",
+    "many",
+    "few",
+    "several",
+    "some",
+    "any",
+    "all",
+    "most",
+    "more",
+    "less",
+    "much",
+    "little",
+    "enough",
+    "various",
+    "numerous",
+    "multiple",
+    "countless",
+    "great",
+    "good",
+    "bad",
+    "nice",
+    "terrible",
+    "awful",
+    "awesome",
+    "amazing",
+    "wonderful",
+    "horrible",
+    "excellent",
+    "poor",
+    "best",
+    "worst",
+    "fine",
+    "okay",
+    "new",
+    "old",
+    "recent",
+    "past",
+    "future",
+    "current",
+    "previous",
+    "next",
+    "last",
+    "first",
+    "latest",
+    "early",
+    "late",
+    "former",
+    "modern",
+    "ancient",
+    "big",
+    "small",
+    "large",
+    "tiny",
+    "huge",
+    "enormous",
+    "long",
+    "short",
+    "tall",
+    "high",
+    "low",
+    "wide",
+    "narrow",
+    "thick",
+    "thin",
+    "deep",
+    "shallow",
+    "similar",
+    "different",
+    "same",
+    "other",
+    "another",
+    "such",
+    "certain",
+    "important",
+    "main",
+    "major",
+    "minor",
+    "key",
+    "primary",
+    "real",
+    "actual",
+    "true",
+    "whole",
+    "entire",
+    "full",
+    "complete",
+    "total",
+    "basic",
+    "simple",
+    "interesting",
+    "boring",
+    "exciting",
+    "special",
+    "particular",
+    "general",
+    "common",
+    "unique",
+    "rare",
+    "typical",
+    "usual",
+    "normal",
+    "regular",
+    "possible",
+    "likely",
+    "potential",
+    "available",
+    "necessary",
+    "only",
+    "solo",
+    "individual",
+    "team",
+    "group",
+    "joint",
+    "collaborative",
+    "final",
+    "initial",
+    "side",
 }

 # Generic tail words to strip from compound entities
 _GENERIC_ENDINGS = {
-    "work", "works", "job", "jobs", "task", "tasks", "stuff", "things",
-    "thing", "info", "information", "details", "data", "content",
-    "material", "materials", "activities", "activity", "efforts", "effort",
-    "options", "option", "choices", "choice", "results", "result",
-    "output", "outputs", "products", "product", "items", "item",
+    "work",
+    "works",
+    "job",
+    "jobs",
+    "task",
+    "tasks",
+    "stuff",
+    "things",
+    "thing",
+    "info",
+    "information",
+    "details",
+    "data",
+    "content",
+    "material",
+    "materials",
+    "activities",
+    "activity",
+    "efforts",
+    "effort",
+    "options",
+    "option",
+    "choices",
+    "choice",
+    "results",
+    "result",
+    "output",
+    "outputs",
+    "products",
+    "product",
+    "items",
+    "item",
 }

 # Capitalized single words that are too generic to be proper nouns
 _GENERIC_CAPS = {
-    "works", "items", "things", "stuff", "resources", "options", "tips",
-    "ideas", "steps", "ways", "methods", "tools", "features", "benefits",
-    "examples", "details", "notes", "instructions", "guidelines",
-    "recommendations", "suggestions", "overview", "summary", "conclusion",
-    "introduction", "pros", "cons", "advantages", "disadvantages",
+    "works",
+    "items",
+    "things",
+    "stuff",
+    "resources",
+    "options",
+    "tips",
+    "ideas",
+    "steps",
+    "ways",
+    "methods",
+    "tools",
+    "features",
+    "benefits",
+    "examples",
+    "details",
+    "notes",
+    "instructions",
+    "guidelines",
+    "recommendations",
+    "suggestions",
+    "overview",
+    "summary",
+    "conclusion",
+    "introduction",
+    "pros",
+    "cons",
+    "advantages",
+    "disadvantages",
 }

 # Markdown/formatting markers to skip during extraction
@@ -120,121 +390,208 @@ def _has_artifacts(txt: str) -> bool:
    )


-def extract_entities(text: str) -> List[Tuple[str, str]]:
-    """Extract named entities, quoted text, and noun compounds from text.
-
-    This is the public API that accepts a string. It loads the spaCy model
-    internally and delegates to _extract_entities_from_doc().
-
-    Args:
-        text: Input text to extract entities from.
-
-    Returns:
-        Deduplicated list of (entity_type, entity_text) tuples.
-        Entity types: PROPER, QUOTED, COMPOUND, NOUN.
-        Returns empty list if spaCy is unavailable.
-    """
-    from mem0.utils.spacy_models import get_nlp_full
-
-    nlp = get_nlp_full()
-    if nlp is None:
-        return []
-
-    doc = nlp(text)
-    return _extract_entities_from_doc(doc)
+def _clean_text(txt: str) -> str:
+    txt = re.sub(r"^\*+\s*|\s*\*+$", "", txt.strip())
+    txt = re.sub(r"\s*:+$", "", txt)
+    txt = re.sub(r"^\d+\s*\.\s*", "", txt)
+    return " ".join(txt.split())


-def extract_entities_batch(texts: List[str], batch_size: int = 32) -> List[List[Tuple[str, str]]]:
-    """Extract entities from multiple texts using spaCy's nlp.pipe() for batched NER.
-
-    Uses spaCy's efficient batch processing pipeline instead of calling
-    nlp() individually per text. Significantly faster for multiple texts.
-
-    Args:
-        texts: List of input texts to extract entities from.
-        batch_size: Number of texts to process in each spaCy batch.
-
-    Returns:
-        List of entity lists, one per input text. Each entity list contains
-        (entity_type, entity_text) tuples. Returns list of empty lists if
-        spaCy is unavailable.
-    """
-    if not texts:
-        return []
-
-    from mem0.utils.spacy_models import get_nlp_full
-
-    nlp = get_nlp_full()
-    if nlp is None:
-        return [[] for _ in texts]
-
-    results = []
-    for doc in nlp.pipe(texts, batch_size=batch_size):
-        results.append(_extract_entities_from_doc(doc))
-    return results
+def _norm_text(txt: str) -> str:
+    return " ".join(txt.lower().split())


-def _extract_entities_from_doc(doc) -> List[Tuple[str, str]]:
-    """Extract entities from a spaCy Doc object.
+def _looks_like_technical_identifier(text: str) -> bool:
+    return bool(re.fullmatch(r"[A-Za-z_][\w-]*(?:\.[A-Za-z_][\w-]*)+", text))

-    Ported from platform's shared.core.utils.entity_extraction.extract_entities().
-    """
-    entities: List[Tuple[str, str]] = []
-    text = doc.text
+
+def _has_internal_cap_or_digit(text: str) -> bool:
+    return any(ch.isdigit() for ch in text) or any(ch.isupper() for ch in text[1:])
+
+
+def _looks_like_metric_count_token(tok) -> bool:
+    return tok.pos_ == "NUM" and bool(re.fullmatch(r"\d[\d,]*(?:\.\d+)?", tok.text))
+
+
+def _is_metric_list_context(tokens: list, idx: int) -> bool:
+    prev_text = tokens[idx - 1].text if idx > 0 else ""
+    next_text = tokens[idx + 1].text if idx + 1 < len(tokens) else ""
+    return prev_text in {":", ",", ";"} or next_text in {",", ";"}
+
+
+def _strip_trailing_metric_counts(span_tokens: list, all_tokens: list) -> list:
+    while len(span_tokens) > 1 and _looks_like_metric_count_token(span_tokens[-1]):
+        tok = span_tokens[-1]
+        if "," not in tok.text and not _is_metric_list_context(all_tokens, tok.i):
+            break
+        span_tokens = span_tokens[:-1]
+    return span_tokens
+
+
+def _is_list_item_name_token(tokens: list, idx: int) -> bool:
+    tok = tokens[idx]
+    if not tok.text or tok.text in _FORMATTING_MARKERS or not tok.text[0].isupper():
+        return False
+    if not any(ch.isalpha() for ch in tok.text) or _is_bad_single_name_token(tok):
+        return False
+    next_tok = tokens[idx + 1] if idx + 1 < len(tokens) else None
+    if not next_tok or not _looks_like_metric_count_token(next_tok):
+        return False
+    return _is_metric_list_context(tokens, idx) or _is_metric_list_context(tokens, idx + 1)
+
+
+def _is_name_like_token(tok, tokens: list | None = None, idx: int | None = None) -> bool:
+    if not tok.text or tok.text in _FORMATTING_MARKERS:
+        return False
+    if not tok.text[0].isupper():
+        return False
+    if not any(ch.isalpha() for ch in tok.text):
+        return False
+    if _is_bad_single_name_token(tok):
+        return False
+    if tok.pos_ == "PROPN" or tok.tag_ in {"NNP", "NNPS"}:
+        return True
+    if tokens is not None and idx is not None and _is_list_item_name_token(tokens, idx):
+        return True
+    if _has_internal_cap_or_digit(tok.text):
+        return True
+    return (
+        tokens is not None
+        and idx is not None
+        and tok.pos_ == "NOUN"
+        and tok.dep_ not in {"compound", "amod"}
+        and not _is_sentence_start(tokens, idx)
+    )
+
+
+def _is_bad_single_name_token(tok) -> bool:
+    lower = tok.text.lower()
+    return lower in _GENERIC_SINGLE_ENTITY_TERMS or lower in _GENERIC_CAPS or tok.is_stop
+
+
+def _add_candidate(
+    candidates: list[_EntityCandidate],
+    entity_type: str,
+    text: str,
+    source: str,
+    start: int,
+    end: int,
+    confidence: float,
+    priority: int,
+) -> None:
+    cleaned = _clean_text(text)
+    if not cleaned or len(cleaned) <= 2 or _has_artifacts(cleaned):
+        return
+    candidates.append(
+        _EntityCandidate(
+            entity_type=entity_type,
+            text=cleaned,
+            source=source,
+            start=start,
+            end=end,
+            confidence=confidence,
+            priority=priority,
+        )
+    )
+
+
+def _add_ner_candidates(doc, candidates: list[_EntityCandidate]) -> None:
    tokens = list(doc)
+    for ent in doc.ents:
+        if ent.label_ in _REJECTED_NER_LABELS or ent.label_ not in _ACCEPTED_NER_LABELS:
+            continue
+        ent_tokens = _strip_trailing_metric_counts(list(ent), tokens)
+        if not ent_tokens:
+            continue
+        if any(tok.pos_ == "CCONJ" and tok.text.lower() == "and" for tok in ent_tokens):
+            continue
+        if len(ent_tokens) == 1 and _is_bad_single_name_token(ent_tokens[0]):
+            continue
+        if (
+            len(ent_tokens) == 1
+            and ent_tokens[0].dep_ in {"compound", "amod"}
+            and ent_tokens[0].head.pos_ in {"NOUN", "PROPN"}
+        ):
+            continue
+        _add_candidate(
+            candidates,
+            "PROPER",
+            "".join(tok.text_with_ws for tok in ent_tokens).strip(),
+            "spacy_ner",
+            ent_tokens[0].i,
+            ent_tokens[-1].i + 1,
+            0.95,
+            0,
+        )

-    # === PROPER NOUN SEQUENCES ===
+
+def _add_technical_identifier_candidates(tokens: list, candidates: list[_EntityCandidate]) -> None:
+    for tok in tokens:
+        if _looks_like_technical_identifier(tok.text):
+            _add_candidate(
+                candidates,
+                "IDENTIFIER",
+                tok.text,
+                "technical_identifier",
+                tok.i,
+                tok.i + 1,
+                0.9,
+                1,
+            )
+
+
+def _add_proper_name_candidates(tokens: list, candidates: list[_EntityCandidate]) -> None:
+    allowed_inner_connectors = {"of", "the", "for", "at", "in"}
    i = 0
    while i < len(tokens):
        tok = tokens[i]
-        if tok.text in _FORMATTING_MARKERS:
+        if not _is_name_like_token(tok, tokens, i):
            i += 1
            continue
-        is_cap = tok.text and tok.text[0].isupper()
-        is_label = i + 1 < len(tokens) and tokens[i + 1].text == ":"

-        if is_cap and not is_label and tok.pos_ in {"PROPN", "NOUN", "ADJ"}:
-            seq = [(tok, i)]
-            j = i + 1
-            while j < len(tokens):
-                t = tokens[j]
-                if (t.text and t.text[0].isupper()) or t.text.lower() in {
-                    "'s", "of", "the", "in", "and", "for", "at", "is",
-                }:
-                    seq.append((t, j))
-                    j += 1
-                else:
-                    break
-            # Strip trailing function words
-            while seq and seq[-1][0].text.lower() in {"of", "the", "in", "and", "for", "at", "is", "'s"}:
-                seq.pop()
-            if seq:
-                has_mid_cap = any(
-                    not _is_sentence_start(tokens, idx)
-                    for (t, idx) in seq
-                    if t.text[0].isupper() and t.text.lower() not in {"'s", "of", "the", "in", "and", "for", "at", "is"}
-                )
-                if has_mid_cap:
-                    phrase = "".join(t.text_with_ws for (t, idx) in seq).strip()
-                    if len(phrase) > 2:
-                        entities.append(("PROPER", phrase))
-            i = j
-        else:
-            i += 1
+        span_tokens = [tok]
+        j = i + 1
+        while j < len(tokens):
+            current = tokens[j]
+            if _is_name_like_token(current, tokens, j):
+                span_tokens.append(current)
+                j += 1
+                continue
+            if (
+                current.text.lower() in allowed_inner_connectors
+                and j + 1 < len(tokens)
+                and _is_name_like_token(tokens[j + 1], tokens, j + 1)
+            ):
+                span_tokens.extend([current, tokens[j + 1]])
+                j += 2
+                continue
+            break

-    # === QUOTED TEXT ===
+        name_tokens = [
+            t
+            for t in span_tokens
+            if _is_name_like_token(t, tokens, t.i) or (0 <= t.i < len(tokens) and _is_list_item_name_token(tokens, t.i))
+        ]
+        if len(name_tokens) > 1 or not _is_bad_single_name_token(name_tokens[0]):
+            text = "".join(t.text_with_ws for t in span_tokens).strip()
+            _add_candidate(candidates, "PROPER", text, "proper_name_span", i, j, 0.8, 2)
+        i = max(j, i + 1)
+
+
+def _add_quoted_candidates(text: str, candidates: list[_EntityCandidate]) -> None:
    for m in re.finditer(r'"([^"]+)"', text):
        if len(m.group(1).strip()) > 2:
-            entities.append(("QUOTED", m.group(1).strip()))
+            _add_candidate(candidates, "QUOTED", m.group(1).strip(), "quoted", -1, -1, 0.75, 3)
    for m in re.finditer(r"(?:^|[\s\(\[{,;])'([^']+)'(?=[\s\.,;:!?\)\]]|$)", text):
        if len(m.group(1).strip()) > 2:
-            entities.append(("QUOTED", m.group(1).strip()))
+            _add_candidate(candidates, "QUOTED", m.group(1).strip(), "quoted", -1, -1, 0.75, 3)

-    # === NOUN-NOUN COMPOUNDS ===
+
+def _add_topic_phrase_candidates(doc, candidates: list[_EntityCandidate]) -> None:
    for chunk in doc.noun_chunks:
        chunk_tokens = list(chunk)
-        split_indices: list = []
-        poss_splits: list = []
+        split_indices: list[int] = []
+        poss_splits: list[int] = []
        for idx, tok in enumerate(chunk_tokens):
            if tok.dep_ == "case" and tok.text in {"'s", "\u2019s", "'"}:
                split_indices.append(idx)
@@ -243,14 +600,14 @@ def _extract_entities_from_doc(doc) -> List[Tuple[str, str]]:
                split_indices.append(idx)

        if split_indices:
-            groups: list = []
+            groups: list[list] = []
            prev = 0
            for split_idx in split_indices:
                if split_idx > prev:
                    groups.append(chunk_tokens[prev:split_idx])
                if split_idx in poss_splits:
                    next_split = next((s for s in split_indices if s > split_idx), None)
-                    owned = chunk_tokens[split_idx + 1: next_split if next_split else len(chunk_tokens)]
+                    owned = chunk_tokens[split_idx + 1 : next_split if next_split else len(chunk_tokens)]
                    if owned:
                        first_content = next((t for t in owned if t.pos_ not in {"PUNCT", "PART"}), None)
                        if not (first_content and first_content.text and first_content.text[0].isupper()):
@@ -272,7 +629,8 @@ def _extract_entities_from_doc(doc) -> List[Tuple[str, str]]:
            content = [
                t
                for t in group
-                if t.pos_ not in {"DET", "PRON", "PUNCT", "PART", "ADP", "SCONJ", "NUM"} and (t.pos_ == "ADJ" or not t.is_stop)
+                if t.pos_ not in {"DET", "PRON", "PUNCT", "PART", "ADP", "SCONJ", "NUM"}
+                and (t.pos_ == "ADJ" or not t.is_stop)
            ]
            if not content:
                continue
@@ -286,78 +644,129 @@ def _extract_entities_from_doc(doc) -> List[Tuple[str, str]]:
            if compound_toks:
                is_circ = any(t.lemma_.lower() in _CIRCUMSTANTIAL_MODS for t in compound_toks)
                if is_circ:
-                    val = head.lemma_ if head.pos_ == "NOUN" else head.text
+                    val = head.text
                    if len(val) > 2:
-                        entities.append(("NOUN", val))
+                        _add_candidate(
+                            candidates,
+                            "TOPIC",
+                            val,
+                            "topic_phrase",
+                            head.i,
+                            head.i + 1,
+                            0.45,
+                            4,
+                        )
                else:
                    filtered = _strip_generic_ending(
                        [t for t in content if not (t.pos_ == "ADJ" and t.lemma_.lower() in _NON_SPECIFIC_ADJ)]
                    )
                    if filtered:
-                        phrase = _lemmatize_compound(filtered)
+                        phrase = " ".join(t.text for t in filtered)
                        if len(phrase) > 3 and " " in phrase:
-                            entities.append(("COMPOUND", phrase))
+                            _add_candidate(
+                                candidates,
+                                "TOPIC",
+                                phrase,
+                                "topic_phrase",
+                                filtered[0].i,
+                                filtered[-1].i + 1,
+                                0.45,
+                                4,
+                            )
            elif len(content) > 1 and has_spec_adj:
                filtered = _strip_generic_ending(
-                    [t for t in content if not ((t.pos_ == "ADJ" or t.dep_ == "amod") and t.lemma_.lower() in _NON_SPECIFIC_ADJ)]
+                    [
+                        t
+                        for t in content
+                        if not ((t.pos_ == "ADJ" or t.dep_ == "amod") and t.lemma_.lower() in _NON_SPECIFIC_ADJ)
+                    ]
                )
                if filtered:
-                    phrase = _lemmatize_compound(filtered)
+                    phrase = " ".join(t.text for t in filtered)
                    if len(phrase) > 3 and " " in phrase:
-                        entities.append(("COMPOUND", phrase))
+                        _add_candidate(
+                            candidates,
+                            "TOPIC",
+                            phrase,
+                            "topic_phrase",
+                            filtered[0].i,
+                            filtered[-1].i + 1,
+                            0.45,
+                            4,
+                        )

-    # === FALLBACK: Mis-tagged VERB heads ===
-    processed = {e[1].lower() for e in entities if e[0] == "COMPOUND"}
-    generic_verb_heads = _GENERIC_HEADS | {"find", "buy", "purchase", "sale", "deal", "trip", "visit"}

-    def collect_compounds(head):
-        return [t for t in doc if t.head == head and t.dep_ == "compound"]
+def _spans_overlap(a: _EntityCandidate, b: _EntityCandidate) -> bool:
+    if a.start < 0 or b.start < 0:
+        return False
+    return a.start < b.end and b.start < a.end

-    for tok in doc:
-        if tok.pos_ == "VERB" and tok.dep_ in {"pobj", "dobj", "nsubj"}:
-            comps = sorted(collect_compounds(tok), key=lambda t: t.i)
-            if comps:
-                phrase_toks = comps if tok.lemma_.lower() in generic_verb_heads else comps + [tok]
-                phrase = " ".join(t.text for t in phrase_toks)
-                if phrase.lower() not in processed and len(phrase) > 3 and " " in phrase:
-                    entities.append(("COMPOUND", phrase))
-                    processed.add(phrase.lower())

-    # === DEDUPLICATION & CLEANUP ===
-    seen: set = set()
-    deduped = []
-    for t, e in entities:
-        k = e.lower().strip()
-        if k not in seen and len(k) > 2:
-            seen.add(k)
-            deduped.append((t, e))
+def _resolve_candidates(candidates: list[_EntityCandidate]) -> list[tuple[str, str]]:
+    deduped_by_text: dict[str, _EntityCandidate] = {}
+    for candidate in candidates:
+        key = _norm_text(candidate.text)
+        current = deduped_by_text.get(key)
+        if current is None or (candidate.priority, -candidate.confidence) < (current.priority, -current.confidence):
+            deduped_by_text[key] = candidate

-    cleaned: List[Tuple[str, str]] = []
-    for etype, etext in deduped:
-        txt = re.sub(r"^\*+\s*|\s*\*+$", "", etext.strip())
-        txt = re.sub(r"\s*:+$", "", txt)
-        txt = re.sub(r"^\d+\s*\.\s*", "", txt)
-        if not txt or len(txt) <= 2 or _has_artifacts(txt):
+    ordered = sorted(
+        deduped_by_text.values(),
+        key=lambda c: (c.priority, -c.confidence, -(c.end - c.start), c.start),
+    )
+    accepted: list[_EntityCandidate] = []
+    for candidate in ordered:
+        if any(
+            _spans_overlap(candidate, existing)
+            and not (candidate.entity_type == "TOPIC" and " " in candidate.text and existing.entity_type == "PROPER")
+            for existing in accepted
+        ):
            continue
-        if etype == "PROPER" and " " not in txt and txt.lower() in _GENERIC_CAPS:
-            continue
-        cleaned.append((etype, txt))
+        accepted.append(candidate)

-    # Keep best type per entity (PROPER > COMPOUND > QUOTED > NOUN)
-    type_pri = {"PROPER": 0, "COMPOUND": 1, "QUOTED": 2, "NOUN": 3, "VERB": 4}
-    best: dict = {}
-    for t, e in cleaned:
-        k = e.lower()
-        if k not in best or type_pri.get(t, 99) < type_pri.get(best[k][0], 99):
-            best[k] = (t, e)
-    deduped = list(best.values())
+    accepted.sort(key=lambda c: (c.start if c.start >= 0 else 10**9, c.end, c.priority))
+    return [(candidate.entity_type, candidate.text) for candidate in accepted]

-    # Remove entities that are whole-word substrings of longer entities.
-    # Word-boundary anchoring avoids dropping distinct entities that only share a
-    # leading substring (e.g. "Sam" must survive alongside "Samsung").
-    all_lower = [e[1].lower() for e in deduped]
-    return [
-        (t, e)
-        for t, e in deduped
-        if not any(e.lower() != o and re.search(rf"\b{re.escape(e.lower())}\b", o) for o in all_lower)
-    ]
+
+def _extract_entities_from_doc(doc) -> list[tuple[str, str]]:
+    """Extract typed entity candidates from a spaCy Doc.
+
+    Args:
+        doc: A spaCy ``Doc`` object (from ``nlp(text)``).
+
+    Returns:
+        Deduplicated list of ``(entity_type, entity_text)`` tuples.
+        Entity types include PROPER, QUOTED, TOPIC, and IDENTIFIER.
+    """
+    tokens = list(doc)
+    candidates: list[_EntityCandidate] = []
+    _add_ner_candidates(doc, candidates)
+    _add_technical_identifier_candidates(tokens, candidates)
+    _add_proper_name_candidates(tokens, candidates)
+    _add_quoted_candidates(doc.text, candidates)
+    _add_topic_phrase_candidates(doc, candidates)
+    return _resolve_candidates(candidates)
+
+
+def extract_entities(text: str) -> list[tuple[str, str]]:
+    """Extract typed entity candidates from text."""
+    from mem0.utils.spacy_models import get_nlp_full
+
+    nlp = get_nlp_full()
+    if nlp is None:
+        return []
+    return _extract_entities_from_doc(nlp(text))
+
+
+def extract_entities_batch(texts: list[str], batch_size: int = 32) -> list[list[tuple[str, str]]]:
+    """Extract typed entity candidates from multiple texts."""
+    if not texts:
+        return []
+
+    from mem0.utils.spacy_models import get_nlp_full
+
+    nlp = get_nlp_full()
+    if nlp is None:
+        return [[] for _ in texts]
+
+    return [_extract_entities_from_doc(doc) for doc in nlp.pipe(texts, batch_size=batch_size)]
@@ -298,9 +298,9 @@ class AzureAISearch(VectorStoreBase):
            payload (Dict, optional): Updated payload.
        """
        document = {"id": vector_id}
-        if vector:
+        if vector is not None:
            document["vector"] = vector
-        if payload:
+        if payload is not None:
            json_payload = json.dumps(payload)
            document["payload"] = json_payload
            for field in ["user_id", "run_id", "agent_id"]:
@@ -169,7 +169,7 @@ class ChromaDB(VectorStoreBase):
        Args:
            vector_id (str): ID of the vector to delete.
        """
-        self.collection.delete(ids=vector_id)
+        self.collection.delete(ids=[vector_id])

    def update(
        self,
@@ -185,7 +185,11 @@ class ChromaDB(VectorStoreBase):
            vector (Optional[List[float]], optional): Updated vector. Defaults to None.
            payload (Optional[Dict], optional): Updated payload. Defaults to None.
        """
-        self.collection.update(ids=vector_id, embeddings=vector, metadatas=payload)
+        self.collection.update(
+            ids=[vector_id],
+            embeddings=[vector] if vector is not None else None,
+            metadatas=[payload] if payload is not None else None,
+        )

    def get(self, vector_id: str) -> Optional[OutputData]:
        """
@@ -39,6 +39,8 @@ class OpenSearchDB(VectorStoreBase):

        self.collection_name = config.collection_name
        self.embedding_model_dims = config.embedding_model_dims
+        self.auto_refresh = config.auto_refresh
+
        self.create_col(self.collection_name, self.embedding_model_dims)

    def create_index(self) -> None:
@@ -148,8 +150,6 @@ class OpenSearchDB(VectorStoreBase):
            }
            try:
                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(
@@ -162,6 +162,14 @@ class OpenSearchDB(VectorStoreBase):
                logger.error(f"Error inserting vector {id_}: {e}", exc_info=True)
                raise

+        # Refresh once after the full batch (not per document) if explicitly enabled.
+        # Disabled by default for Serverless compatibility: OpenSearch Serverless does not
+        # support the indices.refresh() API, and refreshing per document would cause a
+        # cluster-level I/O stall on every insert.
+        # See: https://docs.aws.amazon.com/opensearch-service/latest/developerguide/serverless-genref.html
+        if self.auto_refresh:
+            self.client.indices.refresh(index=self.collection_name)
+
        return results

    def search(
@@ -98,7 +98,10 @@ class Qdrant(VectorStoreBase):
                self._bm25_encoder = SparseTextEmbedding(model_name="Qdrant/bm25")
                logger.info("BM25 encoder loaded (fastembed Qdrant/bm25)")
            except ImportError:
-                logger.warning("fastembed not installed — BM25 keyword search disabled. Install with: pip install fastembed")
+                logger.warning(
+                    "fastembed not installed - BM25 keyword search disabled. "
+                    'Install it with: pip install "mem0ai[extras]"'
+                )
                self._bm25_encoder = False  # sentinel: tried and failed
            except Exception as e:
                logger.warning(f"Failed to load BM25 encoder: {e}")
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "mem0ai"
-version = "2.0.7"
+version = "2.0.9"
 description = "Long-term memory for AI Agents"
 authors = [
    { name = "Mem0", email = "support@mem0.ai" }
@@ -27,6 +27,8 @@ import { useApiQuery } from "@/hooks/use-api-query";
 import { Memory } from "@/types/api";

 const PAGE_SIZE = 20;
+// Keep in sync with ALL_MEMORIES_LIMIT in server/main.py.
+const MEMORY_FETCH_LIMIT = 1000;

 export default function MemoriesPage() {
  const [userId, setUserId] = useState("");
@@ -41,7 +43,9 @@ export default function MemoriesPage() {
    refetch,
  } = useApiQuery<Memory[]>(
    async () => {
-      const params = userId.trim() ? { user_id: userId.trim() } : undefined;
+      const params = userId.trim()
+        ? { user_id: userId.trim(), top_k: MEMORY_FETCH_LIMIT }
+        : { top_k: MEMORY_FETCH_LIMIT };
      const res = await api.get(MEMORY_ENDPOINTS.BASE, { params });
      const raw = res.data?.results ?? res.data ?? [];
      return Array.isArray(raw) ? raw : [];
@@ -96,7 +100,7 @@ export default function MemoriesPage() {
    <div className="space-y-4">
      <h1 className="text-xl font-semibold font-fustat">Memories</h1>

-      {memories.length >= 1000 && (
+      {memories.length >= MEMORY_FETCH_LIMIT && (
        <UpgradeBanner
          id="memories-1k"
          message="1,000+ memories stored. Categories can help organize them."
@@ -16,7 +16,7 @@ from errors import (
    upstream_error,
    upstream_error_handler,
 )
-from fastapi import Depends, FastAPI, HTTPException, Request
+from fastapi import Depends, FastAPI, HTTPException, Query, Request
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import JSONResponse, RedirectResponse
 from mem0.exceptions import ValidationError as Mem0ValidationError
@@ -409,6 +409,7 @@ def get_all_memories(
    user_id: Optional[str] = None,
    run_id: Optional[str] = None,
    agent_id: Optional[str] = None,
+    top_k: Optional[int] = Query(None, ge=0, le=ALL_MEMORIES_LIMIT),
    _auth=Depends(verify_auth),
 ):
    """Retrieve stored memories. Lists all memories when no identifier is provided (admin only)."""
@@ -417,11 +418,14 @@ def get_all_memories(
            auth_type = getattr(request.state, "auth_type", "none")
            if _auth is not None and _auth.role != "admin" and auth_type not in {"admin_api_key", "disabled"}:
                raise HTTPException(status_code=403, detail="Admin role required to list all memories.")
-            return _list_all_memories()
+            return _list_all_memories(limit=top_k if top_k is not None else ALL_MEMORIES_LIMIT)
        filters = {
            k: v for k, v in {"user_id": user_id, "run_id": run_id, "agent_id": agent_id}.items() if v is not None
        }
-        return get_memory_instance().get_all(filters=filters)
+        params = {"filters": filters}
+        if top_k is not None:
+            params["top_k"] = top_k
+        return get_memory_instance().get_all(**params)
    except HTTPException:
        raise
    except Exception:
@@ -84,3 +84,101 @@ def test_generate_response_with_tools(mock_groq_client):
    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."}
+
+
+@pytest.mark.parametrize("model", ["groq/compound", "groq/compound-mini"])
+def test_generate_response_skips_json_mode_for_compound_models(mock_groq_client, model):
+    config = BaseLlmConfig(model=model, temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GroqLLM(config)
+    messages = [{"role": "user", "content": "Hi, I'm Alice and I love hiking."}]
+
+    # Compound models answer JSON-mode requests with empty or non-JSON content;
+    # the mock mirrors that plain-text reply. These tests pin request
+    # construction (response_format omitted), not end-to-end extraction.
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="Alice introduced herself and mentioned she loves hiking."))]
+    mock_groq_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages, response_format={"type": "json_object"})
+
+    _, kwargs = mock_groq_client.chat.completions.create.call_args
+    assert "response_format" not in kwargs
+
+
+def test_generate_response_keeps_json_mode_for_standard_model(mock_groq_client):
+    config = BaseLlmConfig(model="llama-3.3-70b-versatile", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GroqLLM(config)
+    messages = [{"role": "user", "content": "Hi, I'm Alice and I love hiking."}]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content='{"memory": ["Name is Alice", "Loves hiking"]}'))]
+    mock_groq_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages, response_format={"type": "json_object"})
+
+    _, kwargs = mock_groq_client.chat.completions.create.call_args
+    assert kwargs["response_format"] == {"type": "json_object"}
+
+
+def test_generate_response_keeps_non_json_response_format_for_compound_model(mock_groq_client):
+    config = BaseLlmConfig(model="groq/compound", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GroqLLM(config)
+    messages = [{"role": "user", "content": "Hi, I'm Alice and I love hiking."}]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content="Alice loves hiking."))]
+    mock_groq_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages, response_format={"type": "text"})
+
+    _, kwargs = mock_groq_client.chat.completions.create.call_args
+    assert kwargs["response_format"] == {"type": "text"}
+
+
+def test_generate_response_keeps_tools_when_skipping_json_mode(mock_groq_client):
+    config = BaseLlmConfig(model="groq/compound", temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GroqLLM(config)
+    messages = [{"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 = "Done."
+    mock_message.tool_calls = None
+    mock_response.choices = [Mock(message=mock_message)]
+    mock_groq_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages, response_format={"type": "json_object"}, tools=tools)
+
+    _, kwargs = mock_groq_client.chat.completions.create.call_args
+    assert "response_format" not in kwargs
+    assert kwargs["tools"] == tools
+    assert kwargs["tool_choice"] == "auto"
+
+
+def test_generate_response_handles_non_string_model(mock_groq_client):
+    config = BaseLlmConfig(model={"name": "custom-model"}, temperature=0.7, max_tokens=100, top_p=1.0)
+    llm = GroqLLM(config)
+    messages = [{"role": "user", "content": "Hi, I'm Alice and I love hiking."}]
+
+    mock_response = Mock()
+    mock_response.choices = [Mock(message=Mock(content='{"memory": []}'))]
+    mock_groq_client.chat.completions.create.return_value = mock_response
+
+    llm.generate_response(messages, response_format={"type": "json_object"})
+
+    _, kwargs = mock_groq_client.chat.completions.create.call_args
+    assert kwargs["response_format"] == {"type": "json_object"}
@@ -96,6 +96,24 @@ class TestParseVisionMessages:
        result = parse_vision_messages(messages, llm=None)
        assert result == messages

+    def test_malformed_image_dict_raises_value_error(self):
+        # A malformed image part (missing the nested url) used to raise an
+        # uncaught KeyError that aborted add(); it should raise a clear ValueError.
+        mock_llm = Mock()
+        messages = [{"role": "user", "content": {"type": "image_url", "image_url": {}}}]
+        with pytest.raises(ValueError, match=r"missing image_url\.url"):
+            parse_vision_messages(messages, llm=mock_llm)
+        mock_llm.generate_response.assert_not_called()
+
+    def test_none_image_url_raises_value_error(self):
+        # image_url present but None (or any non-dict) must also raise the clear
+        # ValueError, not an AttributeError from calling .get() on None.
+        mock_llm = Mock()
+        messages = [{"role": "user", "content": {"type": "image_url", "image_url": None}}]
+        with pytest.raises(ValueError, match=r"missing image_url\.url"):
+            parse_vision_messages(messages, llm=mock_llm)
+        mock_llm.generate_response.assert_not_called()
+

 class TestRemoveSpacesFromEntities:
    """
@@ -0,0 +1,32 @@
+"""Reranker failures must be logged, not silently swallowed.
+
+Uses the LLMReranker because it is constructible without heavy ML deps (the
+``mock_llm`` fixture stubs the LLM factory). The fix under test is shared by all
+reranker providers: the ``except`` fallback now emits a ``logger.warning`` before
+degrading to the original order / a neutral score.
+"""
+
+import logging
+
+from mem0.reranker.llm_reranker import LLMReranker
+
+
+class TestRerankerFailureLogging:
+    def test_llm_failure_is_logged_and_falls_back(self, mock_llm, caplog):
+        _factory, llm_instance = mock_llm
+        llm_instance.generate_response.side_effect = RuntimeError("upstream 500")
+
+        reranker = LLMReranker({"provider": "openai"})
+        docs = [{"memory": "alpha"}, {"memory": "beta"}]
+
+        with caplog.at_level(logging.WARNING, logger="mem0.reranker.llm_reranker"):
+            result = reranker.rerank("q", docs)
+
+        # Graceful degradation preserved: every doc still comes back, scored neutral.
+        assert len(result) == 2
+        assert all(d["rerank_score"] == 0.5 for d in result)
+
+        # The failure is no longer silent.
+        warnings = [r for r in caplog.records if r.levelno == logging.WARNING]
+        assert warnings, "expected a warning to be logged on reranking failure"
+        assert "upstream 500" in caplog.text
@@ -618,6 +618,31 @@ class TestGetMemories:
        # 3. Verify the core logic: the param was mapped to the filters dict!
        _, kwargs = mock_memory.get_all.call_args
        assert kwargs["filters"] == {"user_id": "test_routing_user"}
+        assert "top_k" not in kwargs
+
+    def test_get_memories_entity_filters_forward_top_k(self, client, mock_memory):
+        response = client.get("/memories?user_id=test_routing_user&top_k=1000")
+
+        assert response.status_code == 200
+
+        _, kwargs = mock_memory.get_all.call_args
+        assert kwargs["filters"] == {"user_id": "test_routing_user"}
+        assert kwargs["top_k"] == 1000
+
+    def test_get_memories_admin_top_k_zero_not_defaulted(self, client, mock_memory):
+        mock_memory.vector_store.list.return_value = []
+
+        response = client.get("/memories?top_k=0")
+
+        assert response.status_code == 200
+        _, kwargs = mock_memory.vector_store.list.call_args
+        assert kwargs["top_k"] == 0
+
+    def test_get_memories_rejects_top_k_above_limit(self, client, mock_memory):
+        response = client.get("/memories?user_id=test_routing_user&top_k=1001")
+
+        assert response.status_code == 422
+        mock_memory.get_all.assert_not_called()


 # ===========================================================================
@@ -6,6 +6,7 @@ def _ensure_spacy():
    """Skip tests if spaCy model is not available."""
    try:
        import spacy
+
        spacy.load("en_core_web_sm")
    except Exception:
        pytest.skip("spaCy en_core_web_sm model not available")
@@ -33,8 +34,9 @@ class TestExtractEntities:

        entities = extract_entities("The machine learning engineer built a neural network")
        entity_texts = [e[1].lower() for e in entities]
-        has_compound = any("machine" in t and "learning" in t for t in entity_texts) or \
-                       any("neural" in t and "network" in t for t in entity_texts)
+        has_compound = any("machine" in t and "learning" in t for t in entity_texts) or any(
+            "neural" in t and "network" in t for t in entity_texts
+        )
        assert has_compound, f"Expected compound nouns, got {entities}"

    def test_empty_string(self):
@@ -76,9 +78,37 @@ class TestExtractEntities:
        for entity in entities:
            assert isinstance(entity, tuple)
            assert len(entity) == 2
-            assert entity[0] in ("PROPER", "QUOTED", "COMPOUND", "NOUN")
+            assert entity[0] in ("PROPER", "QUOTED", "TOPIC", "IDENTIFIER")
            assert isinstance(entity[1], str)

+    def test_handles_names_lists_and_identifiers(self):
+        from mem0.utils.entity_extraction import extract_entities
+
+        text = (
+            "User reported top inbound integration pages: OpenClaw 25,443, "
+            "Claude Code 8,916, Codex 2,573, Dify 656. "
+            "User compared Cartesia and Deepgram. "
+            "The email field for Mem0 lives at person.properties.email. "
+            "The qwen endpoint uses person.properties.email. "
+            "Johnson & Johnson was mentioned. "
+            "Glasses around my window. "
+            "On 2026-05-27 there were 90 days of stats."
+        )
+
+        entities = extract_entities(text)
+        entity_texts = {entity_text for _, entity_text in entities}
+        normalized = {entity_text.lower() for entity_text in entity_texts}
+
+        assert {"OpenClaw", "Claude Code", "Codex", "Dify", "Cartesia", "Deepgram", "Mem0"}.issubset(entity_texts)
+        assert "person.properties.email" in entity_texts
+        assert "qwen endpoint" in entity_texts
+        assert "Johnson & Johnson" in entity_texts
+        assert "top" not in normalized
+        assert "glasses" not in normalized
+        assert "Cartesia and Deepgram" not in entity_texts
+        assert "Claude Code 8,916" not in entity_texts
+        assert not {"8,916", "2,573", "656", "2026-05-27", "90"}.intersection(entity_texts)
+

 class TestExtractEntitiesBatch:
    def test_batch_processing(self):
@@ -511,6 +511,30 @@ def test_insert_with_http_error(azure_ai_search_instance):
    assert "Azure service error" in str(exc_info.value)


+def test_update_allows_empty_payload(azure_ai_search_instance):
+    """Test updating with an empty payload explicitly clears stored payload data."""
+    instance, mock_search_client, _ = azure_ai_search_instance
+
+    mock_search_client.merge_or_upload_documents.return_value = [{"status": True, "id": "doc1", "status_code": 200}]
+
+    instance.update("doc1", payload={})
+
+    mock_search_client.merge_or_upload_documents.assert_called_once_with(
+        documents=[{"id": "doc1", "payload": "{}", "user_id": None, "run_id": None, "agent_id": None}]
+    )
+
+
+def test_update_allows_empty_vector(azure_ai_search_instance):
+    """Test updating with an empty vector still sends the vector field to Azure."""
+    instance, mock_search_client, _ = azure_ai_search_instance
+
+    mock_search_client.merge_or_upload_documents.return_value = [{"status": True, "id": "doc1", "status_code": 200}]
+
+    instance.update("doc1", vector=[])
+
+    mock_search_client.merge_or_upload_documents.assert_called_once_with(documents=[{"id": "doc1", "vector": []}])
+
+
 # --- Tests for search method ---


@@ -122,7 +122,7 @@ def test_delete_vector(chromadb_instance):

    chromadb_instance.delete(vector_id=vector_id)

-    chromadb_instance.collection.delete.assert_called_once_with(ids=vector_id)
+    chromadb_instance.collection.delete.assert_called_once_with(ids=[vector_id])


 def test_update_vector(chromadb_instance):
@@ -133,7 +133,31 @@ def test_update_vector(chromadb_instance):
    chromadb_instance.update(vector_id=vector_id, vector=new_vector, payload=new_payload)

    chromadb_instance.collection.update.assert_called_once_with(
-        ids=vector_id, embeddings=new_vector, metadatas=new_payload
+        ids=[vector_id], embeddings=[new_vector], metadatas=[new_payload]
+    )
+
+
+def test_update_vector_metadata_only(chromadb_instance):
+    # Metadata-only update (vector=None) must not wrap None in a list.
+    vector_id = "id1"
+    new_payload = {"name": "updated_vector"}
+
+    chromadb_instance.update(vector_id=vector_id, vector=None, payload=new_payload)
+
+    chromadb_instance.collection.update.assert_called_once_with(
+        ids=[vector_id], embeddings=None, metadatas=[new_payload]
+    )
+
+
+def test_update_vector_embedding_only(chromadb_instance):
+    # Vector-only update (payload=None) must not wrap None in a list.
+    vector_id = "id1"
+    new_vector = [0.7, 0.8, 0.9]
+
+    chromadb_instance.update(vector_id=vector_id, vector=new_vector, payload=None)
+
+    chromadb_instance.collection.update.assert_called_once_with(
+        ids=[vector_id], embeddings=[new_vector], metadatas=None
    )


@@ -151,6 +151,55 @@ class TestOpenSearchDB(unittest.TestCase):
        self.os_db.create_index()
        self.client_mock.indices.create.assert_not_called()

+    def test_auto_refresh_disabled_by_default(self):
+        """Test that auto_refresh is disabled by default (Issue #3739).
+
+        This ensures OpenSearch Serverless compatibility out-of-the-box since
+        the indices.refresh() API is not supported in serverless mode.
+        """
+        # Default instance should have auto_refresh=False
+        self.assertFalse(self.os_db.auto_refresh)
+        self.client_mock.reset_mock()
+
+        vectors = [[0.1] * 1536]
+        payloads = [{"key1": "value1"}]
+        ids = ["id1"]
+
+        self.os_db.insert(vectors=vectors, payloads=payloads, ids=ids)
+
+        # Verify index was called but refresh was NOT called (default behavior)
+        self.assertEqual(self.client_mock.index.call_count, 1)
+        self.client_mock.indices.refresh.assert_not_called()
+
+    def test_auto_refresh_enabled(self):
+        """Test that refresh is called once per batch (not per document) when auto_refresh=True."""
+        with patch("mem0.vector_stores.opensearch.OpenSearch", return_value=self.client_mock):
+            auto_refresh_db = OpenSearchDB(
+                host="localhost",
+                port=9200,
+                collection_name="test_auto_refresh",
+                embedding_model_dims=1536,
+                auto_refresh=True,  # Enable auto-refresh
+            )
+
+        self.assertTrue(auto_refresh_db.auto_refresh)
+        # auto_refresh_db reuses self.client_mock (patched above), so reset to drop
+        # the index calls made during construction before asserting on insert().
+        self.client_mock.reset_mock()
+
+        # Insert a batch of 3 vectors to verify the refresh is hoisted out of the
+        # per-document loop: index() is called once per document, but refresh()
+        # must fire exactly once for the whole batch.
+        vectors = [[0.1] * 1536, [0.2] * 1536, [0.3] * 1536]
+        payloads = [{"key1": "value1"}, {"key2": "value2"}, {"key3": "value3"}]
+        ids = ["id1", "id2", "id3"]
+
+        auto_refresh_db.insert(vectors=vectors, payloads=payloads, ids=ids)
+
+        # index() once per document, but refresh() only once for the batch
+        self.assertEqual(self.client_mock.index.call_count, 3)
+        self.assertEqual(self.client_mock.indices.refresh.call_count, 1)
+
    def test_insert(self):
        vectors = [[0.1] * 1536, [0.2] * 1536]
        payloads = [{"key1": "value1"}, {"key2": "value2"}]
@@ -1025,3 +1025,19 @@ class TestQdrantDatetimeRangeFilters(unittest.TestCase):
        types = {type(c.range) for c in result.must}
        self.assertIn(DatetimeRange, types)
        self.assertIn(Range, types)
+
+    def test_missing_fastembed_warns_with_extras_install_hint(self):
+        """When fastembed is missing, the BM25 warning must point at the mem0 optional group
+        that actually provides it (`mem0ai[extras]`), not the bare `fastembed` package -
+        fastembed is declared under [extras], so that is the discoverable install path."""
+        self.qdrant._bm25_encoder = None
+        # Setting the module to None in sys.modules makes `from fastembed import ...` raise ImportError.
+        with patch.dict("sys.modules", {"fastembed": None}):
+            with self.assertLogs("mem0.vector_stores.qdrant", level="WARNING") as cm:
+                result = self.qdrant._get_bm25_encoder()
+        self.assertIsNone(result)  # encoder unavailable -> None
+        self.assertIs(self.qdrant._bm25_encoder, False)  # sentinel: tried and failed, do not retry
+        joined = "\n".join(cm.output)
+        self.assertIn("mem0ai[extras]", joined)  # points at the right install
+        self.assertNotIn("pip install fastembed", joined)  # not the bare package
+        self.assertNotIn("\u2014", joined)  # no em-dash
Author	SHA1	Message	Date
kartik-mem0	368cc5fb1e	docs(changelog): correct score_details fields, graph-store removal note, and dead migration link	2026-06-25 11:59:57 +05:30
Taranjeet SinghandGitHub	ac296f7534	docs: make example code fences copy-safe (#5833 )	2026-06-24 22:19:40 -07:00
KartikandGitHub	ac8f862ff7	fix(mem0-plugin): store files_touched as a list to stop double JSON-encoding (#5806 )	2026-06-25 09:06:11 +05:30
soumil-rathi GitHub Soumil Rathi	b33fa5427c	fix(memory): align entity extraction precision (#5829 ) Making the entity extraction function cleaner Co-authored-by: Soumil Rathi <soumilrathi@gmail.com>	2026-06-24 15:48:37 -07:00
Ashutosh KasudhanandGitHub	5d573dd2ae	docs: added typescript support to deepseek provider (#5723 )	2026-06-24 20:55:14 +05:30
KartikandGitHub	98dbf90864	chore: update changelog, bump SDK versions to Python 2.0.8 and TypeScript 3.0.10 (#5825 )	2026-06-24 20:21:06 +05:30
David ShraderandGitHub	6ddf1669f4	fix(vector_stores): point fastembed-missing warning at mem0ai[extras] (#5622 )	2026-06-24 14:47:02 +05:30
David ShraderandGitHub	d3d2e89fd5	fix(llms): skip JSON response_format for Groq compound models (#5513 )	2026-06-24 14:42:22 +05:30
XiaojuandGitHub	43175d85f2	fix: preserve empty Azure AI Search update values (#5524 ) Signed-off-by: Xiaoju <xiaojuchh@gmail.com>	2026-06-24 14:41:55 +05:30
KartikandGitHub	661ecb9f0f	docs(hermes): update for v3 API, OSS self-hosted mode, and new tools (#5807 )	2026-06-24 11:59:13 +05:30
冯基魁andGitHub	09f181c577	fix(server): fetch filtered dashboard memories beyond default page (#5753 )	2026-06-24 11:49:06 +05:30
Zaid GitHub Zaid Malhis	3497f26a00	fix: add auto_refresh option for OpenSearch Serverless compatibility (#3893 ) Co-authored-by: Zaid Malhis <Malhis@users.noreply.github.com>	2026-06-24 11:48:52 +05:30
BartokandGitHub	e9c0547423	fix(ts-sdk): preserve customCategories names through key conversion (#5741 )	2026-06-24 11:15:19 +05:30
Yash SinghandGitHub	25bc1b7426	fix(memory): guard against malformed image_url in parse_vision_messages (#5631 )	2026-06-24 11:10:36 +05:30
Hrushikesh YadavandGitHub	fee344db85	fix(chroma): wrap scalar vector_id in list for delete() (#5703 )	2026-06-24 10:54:36 +05:30
BartokandGitHub	b611f69381	fix(chroma): wrap update() ids/embeddings/metadatas in lists (#5757 )	2026-06-24 10:54:15 +05:30
Muhammad FurqanandGitHub	c2862831db	fix(reranker): log reranking failures instead of swallowing them silently (#5717 )	2026-06-24 10:44:37 +05:30