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 + neural reranking surfaces the most recent, 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 — 66.6% QA accuracy on LoCoMo, +50.2 pp on temporal questions over RAG, at just 1.02 LLM calls per query

Every conversation turn passes through a four-stage pipeline:

flowchart TD
 A[User Message] --> CTX["Context Assembly\ncurrent speaker · rolling summary (every 15 turns) · recent msgs (last 10)"]
 CTX --> B[ExtractionPipeline]
 B -->|candidate facts + speaker attribution| 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 ×ばつ 3 candidates\n(per-speaker namespace)"]
 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]
 J -->|"don't know + non-OD (1.7%)"| L[Two-pass Rephrase]
 L --> G
 style I fill:#8e44ad,color:#fff
 style E fill:#e74c3c,color:#fff
 style J fill:#3498db,color:#fff
 style CTX fill:#e67e22,color:#fff

Evaluated on the LoCoMo benchmark across 10 conversations (n=1540 QA pairs) with per-speaker memory isolation.

25.5% of incorrect answers had the correct fact present in Qdrant but it did not reach the top-30 reranked context window — the primary remaining retrieval bottleneck, addressable with sparse vector (BM25) retrieval on re-ingestion.

Bootstrap 95% CIs: Overall [64.4%, 68.8%], Temporal [69.9%, 80.1%]
Average LLM calls per query: 1.02 (two-pass fires on 1.7% of non-open-domain queries)

Eidetic Memory v2 is the only system in this comparison with per-speaker namespace isolation and a neural reranking step. Its temporal score (75.1%) is within noise of Hindsight-20B (76.3%) at a fraction of the model scale.

flowchart LR
 Q[Question] --> VA["Vector Search\nSpeaker A namespace"]
 Q --> VB["Vector Search\nSpeaker B namespace"]
 VA -->|top-k×ばつ3 facts| RR["Round-Robin Merge\nzip_longest interleave"]
 VB -->|top-k×ばつ3 facts| RR
 RR -->|180 candidates| JR["Local Cross-Encoder\n(ms-marco-MiniLM-L-6-v2)"]
 JR -->|top-30 reranked| LLM["Answer Generation\n1 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 1

See 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

Apache License 2.0