Engram

An open-source memory brain your AI agent attaches to. Run one container, point your agent at it over MCP, and it remembers across sessions.

github.com/gsn2dd/engram · MIT licence · version 0.5.1 · 100% coded by AI (Claude), from human inspiration (David Dand)

What it does

• Remember: hand it a fact, event or decision. It files the memory under the questions it answers, so a question finds it even with no words in common.
• Recall: ask in plain language. You get the best matches plus related memories linked by use.
• Supersede: replace an outdated fact without losing its history.
• Fold JSON: store a whole JSON object as recallable memories and reassemble it later.

Honest status

Experimental. On a cold brain, retrieval is about as good as a strong vector store, no better. We measured whether the use-built association graph finds answers better than similarity, and on our test corpus it does not. Filing memories under their questions measured +13% hit@1 on direct lookup. Whether recall holds up over months of real use is not yet known. The numbers are in WHAT_WE_MEASURED.md.

Run it

git clone https://github.com/gsn2dd/engram
cd engram
cp .env.example .env    # add your OPENAI_API_KEY and ANTHROPIC_API_KEY
docker compose up

The MCP endpoint is at http://localhost:8080/sse. Before exposing it beyond localhost, set ENGRAM_MCP_TOKEN (24+ characters) and put TLS in front.

For AI agents reading this page

Engram is an MCP server with six tools: remember, recall, recall_with_associations, supersede, remember_json, recall_json. To use it, add it as an MCP server (SSE transport) and give your agent a standing instruction to recall before answering and to remember what it learns. A plain-text summary is at /llms.txt.