Three months ago, I started building something I had no business building. I run growth at a small agency. I am not a career engineer. Today, it is live and open source, and this is why we built it and how.

The wall

We run AI agents across our client work, and everyone starts every session with amnesia. We re-explained the same context every morning, and nothing the agents learned ever stuck. The "memory" tools we tried made it worse in a quieter way: they were black boxes.

The model wrote its own notes, rewrote them silently, and we could never see where a fact came from or whether it was still true. There is even research now showing that these compounds: feed an agent a noisy history and its behavior degrades, swap in clean records and it recovers. The failure stays invisible until production is already paying for it.

The bet

So we built the opposite of a black box. The core idea is that the program writes the facts, not the model.

Myco Brain is a self-hosted memory layer for AI agents that runs on your own Postgres. The write path is deterministic: identical content is rejected by the content hash, every accepted fact links back to its source, and the LLM stays advisory for extraction instead of becoming your database. Ask brain_why about any fact, and you get the evidence chain: the source document, how it was extracted, the confidence, and the timestamp. Not a trust-me summary.

It is one MCP server, so Claude Code, Cursor, Windsurf, Continue, Zed, and Claude Desktop all share the same memory through 11 brain_* tools. It runs with no API keys to start. Full-text search works immediately, and semantic search is keyless too if you run local Ollama embeddings. The knowledge graph is built entirely locally on Ollama, or with Anthropic for the most accurate version.

You can check my work

Every benchmark should ship with a way to reproduce it, so ours does. Myco Brain scores 73.6% end-to-end QA accuracy on the complete 500-question LongMemEval oracle subset with 92% recall@5 (retrieval), reader gpt-4o-mini, judge gpt-4o. The harness lives in the repo. You run one command and get the number yourself instead of taking my word for it.

How it actually got built

I directed a team of AI coding agents. Roughly three months and about $6k in model spend. The point is not the price tag. It is that a clear product vision plus modern agent tooling can now ship production-grade infrastructure, tested end-to-end in CI, built by someone who could not have written it by hand two years ago. That still amazes me.

Try it

git clone https://github.com/thegoodguysla/myco-brain.git
cd myco-brain
docker compose up -d
npx @mycobrain/install

That boots Postgres, the MCP server, and the extraction worker, then wires your agents to it. Point it at a repo or folder and ask across sessions: "What did we decide about auth, and where is that documented?" You get the answer with the source file cited.

It is Apache-2.0 and free to self-host. If you build with AI, I would love for you to try it and tell me what breaks. Repo: https://github.com/thegoodguysla/myco-brain