engram

a cognitive memory system that actually remembers things.

one sqlite file, hybrid retrieval that fuses five signals, memory layers that model how brains actually work, and a neural visualization that shows the whole thing firing in real time.

98.1% R@5 on LongMemEval (ICLR 2025) — highest published score, beating MemPalace (96.6%), Emergence AI (86%), and every other memory system benchmarked.

what it does

• hybrid retrieval — HNSW dense + BM25 + entity graph BFS + Hopfield associative pattern completion, fused with intent-weighted reciprocal rank fusion, cross-encoder reranking, deep MLP reranking
• memory layers — working, episodic, semantic, procedural, codebase. memories promote upward when useful, decay if unused
• entity graph — extracts people, tools, projects from every memory. multi-hop traversal via recursive SQL CTEs
• 63 MCP tools — plugs into Claude Code or any MCP client. 72 tests. docker-ready
• multi-backend embeddings — local (MLX/CPU), Voyage AI, OpenAI, Google Gemini. auto-detects from model name
• web dashboard — 17 panels including neural map, search, analytics, cognition, drift detection

install

pip install engram-memory-system

or from source:

git clone https://github.com/raya-ac/engram.git
cd engram && pip install -e .

quick links

• Installation — full setup guide
• Quick Start — first steps
• MCP Tools Reference — all 63 tools
• Architecture — how it works inside
• Benchmarks — LongMemEval results

supported by

system	what
Claude Code	native MCP integration
Any MCP client	63-tool JSON-RPC server
REST API	57 HTTP endpoints
CLI	15 commands
Docker	single container deployment
Python API	direct library usage