JDSS/mem0
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: reinstate graph memory terminology (native entity linking) (#5601)

Browse Source
This commit is contained in:
rudrajmehta-mem0
2026-06-16 21:22:55 -07:00
committed by GitHub
parent d772f9a961
commit 6aa9bffa55
10 changed files with 126 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/core-concepts/memory-evaluation.mdx
+4 -4
View File
@@ -17,7 +17,7 @@ Some benchmarks today — particularly smaller ones like LoCoMo and LongMemEval
## Architecture Overview
Mem0's memory system operates across two phases — **extraction** (writing) and **retrieval** (reading) — with an entity linking layer connecting them.
Mem0's memory system operates across two phases, **extraction** (writing) and **retrieval** (reading), with a graph memory layer (entity linking) connecting them.
### Memory Extraction (Distillation)
@@ -27,14 +27,14 @@ When new conversations arrive, the extraction pipeline processes them through fi
2. **Context Lookup** — Find related existing memories to avoid duplicates
3. **Distill Memories** — Single-pass LLM extraction produces ADD-only facts from input + context
4. **Deduplicate + Embed** — Hash-based deduplication, then vectorize new memories
5. **Entity Linking** — Identify entities (proper nouns, quoted text, compound noun phrases) and link them across memories
5. **Graph Memory (Entity Linking)**: Identify entities (proper nouns, quoted text, compound noun phrases) and link them across memories into a graph
Memories are distributed across three storage layers, each tuned for a specific retrieval pattern:
| Store | Contents | Purpose |
|---|---|---|
| **Vector Database** | Memory text, embeddings, metadata (timestamps, hash, categories, attributed_to) | Primary fact storage + semantic retrieval |
| **Entity Store** | Entities + embeddings + linked memory IDs | Entity-based retrieval boost |
| **Graph / Entity Store** | Entities + embeddings + linked memory IDs | Graph connections across memories + entity-based retrieval boost |
| **SQL Database** | History log (ADD events) + rolling message window | Audit trail + extraction dedup context |
<Info>
@@ -76,7 +76,7 @@ The combined score outperformed every individual signal across every category te
*Mean tokens: 6,956*
The two largest gains are **temporal queries (+29.6)** and **multi-hop reasoning (+23.1)**. Both categories directly test the ADD-only architecture (preserving temporal context) and entity linking (connecting facts across memories).
The two largest gains are **temporal queries (+29.6)** and **multi-hop reasoning (+23.1)**. Both categories directly test the ADD-only architecture (preserving temporal context) and graph memory / entity linking (connecting facts across memories).
### LongMemEval