All work
Open source
Engram
Persistent memory for AI agents
Engram is an open-source memory layer for LLM agents. Agents that run for weeks need to remember what happened last Tuesday without re-reading every conversation. Engram stores raw turns as they come in, then finds the right evidence at question time with hybrid search (BM25 plus vector search), graph traversal and durable task ledgers.
01Highlights
- 90.8% on LongMemEval-S (500 questions), 94.1% on LoCoMo-10 (1,540) and 79.7% on BEAM 1M (700)
- Retrieval finds the gold session for all 470 answerable LongMemEval questions
- No LLM calls at ingest: embeddings run on-device with all-MiniLM-L6-v2
- Cuts LLM context cost by up to 15× while holding over 90% accuracy on LongMemEval and LoCoMo
- Hybrid search with BM25, vectors and reciprocal rank fusion
02Stack
- Python
- PostgreSQL
- pgvector
- BM25
- RRF
- Graph traversal