Long-term memory for AI agents — extracts, evolves, and retrieves facts across conversations.
- 🧠 Remembers what matters — extracts atomic facts from every conversation turn, not raw chat logs
- ⚡ Evolves over time — ADD, UPDATE, DELETE, NOOP decisions keep memory fresh and conflict-free
- 🎯 Retrieves the right context — semantic search + importance reranking surfaces relevant facts at query time
- 🗣️ Multi-party ready — per-speaker memory isolation prevents cross-speaker contamination in group conversations
- 🔌 Any LLM, any time — swap Claude, Gemini, Azure OpenAI, or Groq with a single env var
- 📊 Benchmark-validated — 56.3% QA accuracy on LoCoMo, +39.3 pp on temporal questions over RAG
Every conversation turn passes through a four-stage pipeline:
flowchart TD
A[User Message] --> B[ExtractionPipeline]
B -->|candidate facts| C[EvolutionEngine]
C -->|ADD / UPDATE / DELETE / NOOP| D[QdrantMemoryStore]
D -->|stored embeddings + payloads| E[(Qdrant Vector DB)]
F[Query] --> G[MemoryRetriever]
G -->|embed query| H[Vector Search\ntop-k × 3 candidates]
H --> E
E -->|candidates| I[Local Cross-Encoder\ncross-encoder/ms-marco-MiniLM-L-6-v2]
I -->|top-k reranked facts| J[Answer Generation LLM]
J --> K[Response]
style I fill:#8e44ad,color:#fff
style E fill:#e74c3c,color:#fff
style J fill:#3498db,color:#fff
Evaluated on the LoCoMo benchmark across 10 conversations (n=1540 QA pairs) with per-speaker memory isolation.
| Metric | Score | Details |
|---|---|---|
| Fact extraction recall | 95.0% | n=100, QA pairs with evidence |
| Fact extraction precision | 52% | Relevant facts / total extracted |
| Conflict resolution accuracy | 100% | 26 test cases (ADD/UPDATE/DELETE/NOOP) |
| K | Hit@K |
|---|---|
| 1 | 11% |
| 5 | 25% |
| 10 | 38% |
| 20 | 56% |
| Category | Accuracy |
|---|---|
| Temporal | 64.2% |
| Open-domain | 60.5% |
| Single-hop | 40.1% |
| Multi-hop | 40.6% |
| Overall | 56.3% |
| System | Overall | Temporal | Notes |
|---|---|---|---|
| RAG baseline (ours) | 44.4% | 24.9% | Direct retrieval over raw turns |
| Pipeline v2 (ours) | 46.6% | 57.3% | Per-speaker isolation + round-robin |
| Eidetic Memory (ours) | 56.3% | 64.2% | + local cross-encoder (ms-marco-MiniLM-L-6-v2) |
| Mem0 | ~66.9% | — | 3× more LLM calls per query |
| Memobase | 75.78% | 85.05% | — |
| Hindsight (OSS-20B) | 83.18% | 76.32% | — |
| Hindsight (OSS-120B) | 85.67% | 79.44% | — |
| Run | Score | Details |
|---|---|---|
| Baseline | 14.8% | Gemini, conv-30 only, top-k=10 |
| + Per-speaker isolation | 31.8% | Multi-namespace retrieval |
| + Fixed merge | 52.4% | Round-robin interleaving |
| + Named entities + two-pass | 53.2% | n=233 |
| + Isolation + RR, no rerank | 25.8% | n=233 — reranker is load-bearing |
| + Isolation + RR + local cross-encoder | 55.8% | n=233 |
| + Isolation + score-based + local cross-encoder | 55.8% | n=233 — merge strategy irrelevant |
| + NER + two-pass + local cross-encoder | 56.6% | n=233 |
| + Local cross-encoder (full run) | 56.3% | Full n=1540, all 10 convs, v2 judge |
flowchart LR
Q[Question] --> VA["Vector Search<br/>Speaker A namespace"]
Q --> VB["Vector Search<br/>Speaker B namespace"]
VA -->|top-k×3 facts| RR["Round-Robin Merge<br/>zip_longest interleave"]
VB -->|top-k×3 facts| RR
RR -->|combined candidates| JR["Local Cross-Encoder<br/>(ms-marco-MiniLM-L-6-v2)"]
JR -->|top-k reranked| LLM["Answer Generation<br/>1 LLM call"]
LLM --> ANS[Answer]
style JR fill:#8e44ad,color:#fff
style LLM fill:#3498db,color:#fff
# 1. Clone and install
git clone https://github.com/CodeNinjaSarthak/eidetic-memory.git
cd eidetic-memory
uv sync --all-packages
# 2. Configure — copy and fill in your API keys
cp .env.development.example .env.development
# 3. Run the benchmark (requires ingested memories)
# Uses ~/eval-venv (sentence-transformers + ONNX); cross-encoder model
# (~66MB) downloads to ~/.cache/huggingface/ on first run.
~/eval-venv/bin/python eval/eval_qa_accuracy.py \
--conv-ids conv-26 conv-30 conv-41 conv-42 conv-43 conv-44 \
conv-47 conv-48 conv-49 conv-50 \
--local-rerank \
--output eval/results/my_results.json \
--concurrency 1See eval/README.md for ingestion instructions and full evaluation documentation.
eidetic-memory/
├── backend/
│ ├── apps/api/ # FastAPI routes
│ ├── services/ # llm · memory · retrieval · storage
│ └── packages/config/ # Shared settings
├── frontend/ # Next.js chat UI
└── eval/ # LoCoMo evaluation harness
Dependency graph (no circular deps):
config → storage → llm → retrieval → memory → api
make check # lint + 152 tests
make run # start API with hot reload