Skip to content

For developers

Your agents forget. Give them memory you can actually trust.

Myco is an open-source, self-hosted memory layer for AI agents. It runs keyless on your own Postgres, the program writes the facts (not the LLM), and every fact carries a source, a confidence score, and an audit trail. Wire up one MCP server and stop rebuilding retrieval for every feature.

Get started in minutes →View on GitHub →

Stateless agents forget. Black-box memory can't be trusted.

They start from zero every session
You re-explain the same context on every run, and nothing your agent learns ever compounds.
You can't see where a fact came from
Hosted memory rewrites itself behind an API. No provenance, no way to know if a fact still holds, no way to roll back a bad one.
Bespoke RAG glue rots
A vector store here, chunking there, session logic everywhere. It drifts and breaks as you ship.
See it in action

What your agents can do with real memory.

Not a chat window that forgets. A shared, source-traceable brain every agent reads and writes, on your own Postgres.

Fresh session, zero chat history
What did we decide about auth, and why?
✓

Myco answers from the decision doc that recorded it, with the full source chain from brain_why. Your agent didn't reconstruct it from context, it queried durable memory.

decision-auth.mdbrain_why
Every agent shares one brain
Claude Code writes a fact; Cursor and your production agent read it the same second. One Postgres source of truth, not a per-thread scratchpad.
Memory that gets more reliable as it grows
Independent sources corroborate a fact and its confidence rises; a contradiction supersedes the stale one, visibly. Your agent stops acting on what is no longer true.
Trace any answer to its source
brain_why returns the evidence chain (supported by N mentions across N documents), so you can debug, audit, and answer trust questions without guessing.
Industries

Where Myco is indispensable.

Pick your world. Here is what becomes possible for developers, in the industries where memory you can audit is a requirement, not a nice to have.

FinTech

Regulated

The problem: Regulators now demand a replayable audit trail for every AI decision.

What you can build

Wire your fraud or ops agent to one memory layer with an RLS-enforced, replayable decision trail. Access control is deterministic, not LLM-trusted.

Why Myco: Every decision traces to its source and rule via brain_why, the program writes the facts (no silent hallucination into the record), and it self-hosts so financial data never leaves your VPC.

Run it. Seed it. Trust it.

01
Run it
Self-host with one docker compose up and npm i @mycobrain/mcp-server. Full-text search runs immediately, keyless. Add local Ollama (one model pull) and semantic search plus the knowledge graph run keyless too, no API keys required.
02
Seed it
Point any MCP-compatible agent at Myco, then import your ChatGPT and Claude history, folders, and GitHub repos. One deduped, provenance-tracked document per conversation.
03
Trust it
Agents read and write through 13 MCP tools. The program writes the facts. Corroboration raises confidence, contradiction supersedes the stale fact, and brain_why traces any answer to its source.

Benchmarked in the open

  • ✓73.6% end-to-end QA on the full 500-question LongMemEval oracle subset, no sampling.
  • ✓recall@5 of 89.2% keyless with local embeddings, 91.6% with the recency reranker, on the full set.
  • ✓Apache-2.0. Audit it, fork it, self-host it, run it forever.
  • ✓Every number reproduces from a benchmark that ships in the repo.
Keyless
runs with no API keys
13 tools
one MCP server, every agent
brain_why
provenance on every fact

Ship agents you can actually trust.

Self-host it free, point your agents at it, and watch a new fact supersede a stale one. Managed cloud is on the waitlist.

Get started in minutes →View on GitHub →
Myco for developers