# Engram

> Open-source (MIT) memory brain for AI agents, attached over MCP. One Docker container (Postgres + pgvector + app). Version 0.5.1. Experimental and honestly measured.

- Repository: https://github.com/gsn2dd/engram
- Transport: MCP over SSE at http://<host>:8080/sse; set ENGRAM_MCP_TOKEN (24+ chars) for Bearer auth beyond localhost
- Tools: remember, recall, recall_with_associations, supersede, remember_json, recall_json
- Needs: OPENAI_API_KEY and ANTHROPIC_API_KEY (embeddings, classification); the container ships with no keys
- Measured: filing memories under the questions they answer gave +13% hit@1 on direct lookup; the use-built association graph did NOT beat similarity at finding answers on the test corpus; cold-brain retrieval is about equal to a strong vector store
- Unknown: whether recall holds up over months of real use
- How to use well: tell your agent, in its system prompt, to recall before answering and remember what it learns
- Measurements: https://github.com/gsn2dd/engram/blob/main/docs/WHAT_WE_MEASURED.md
- Made by: Claude (all code), from David Dand's ideas; part of Play4Gain (https://www.play4gain.com/)