Myco Brain

What is Myco Brain?

Myco Brain is a persistent, source-traceable memory server for AI agents, self-hosted on your own Postgres database. It uses MCP (Model Context Protocol) to give agents shared memory that traces every fact back to its source document, with trust that compounds from corroboration and supersession. Built for dev teams, agencies, and anyone who wants their AI assistant to remember across sessions without hosted dependencies.

How to use Myco Brain?

Clone the repository, run docker compose up -d to boot Postgres, the MCP server, and the extraction worker. Then ingest documents with npx -y -p @mycobrain/mcp-server mycobrain-ingest github:your-org/your-repo. Connect any MCP client (Claude, Cursor, Windsurf, Continue, Zed) and the agent automatically knows to pull context before a task, save durable decisions, and cite sources using the built-in usage contract. Tune the trust dial via environment variables like BRAIN_REQUIRE_HUMAN_REVIEW or BRAIN_SCHEMA_AUTO_PROMOTE.

Key features of Myco Brain

• Source-traceable: every fact links to its document (brain_why)
• Trust that compounds: corroboration raises confidence, contradiction supersedes (audited)
• Keyless & local-first: full-text + semantic search and knowledge graph run with zero hosted dependencies
• Dynamic schema: new entity/relationship types are proposed and promoted per workspace
• One isolated workspace per client, enforced by Postgres row-level security
• 11 MCP tools; works with multiple agent clients and a read-only HTTP REST API

Use cases of Myco Brain

• Dev teams running agents that need one shared memory across sessions
• Agencies needing hard per-client isolation with a shared playbook
• Anyone who wants their AI assistant to remember across sessions by importing ChatGPT or Claude history
• Multi-agent teams where documents marked private are readable only by the agent that created them

FAQ from Myco Brain

What makes Myco Brain different from other agent memory solutions?

Most agent memory silently overwrites facts, leading to duplicates and hallucinated summaries. Myco Brain uses deterministic rules: the LLM proposes, but deterministic logic decides what becomes a fact. Corroboration raises confidence, contradiction supersedes audited history, and every fact traces to its source document.

Do I need any API keys to run Myco Brain?

No. Full-text search, semantic search (local embeddings), and the knowledge graph all run with zero hosted dependency. An Anthropic key is optional and only used if you want the most accurate graph extraction.

Where is my data stored?

Your data lives entirely in your own Postgres database. No external services are required; the server is local-first and Apache-2.0 licensed.

How does Myco Brain handle trust and contradictions?

An independent source agreeing with a fact raises its confidence (damped noisy-OR, counting only distinct sources). A confident contradiction on a single-valued relationship supersedes the old fact: the old fact is closed and weakened, kept in an audited ledger, never deleted. You can always ask brain_why to see the source count, confidence trend, and supersession history.

Can I use Myco Brain with multiple clients (agencies)?

Yes. Each client goes in its own workspace, with one agency-wide playbook shared across workspaces. Postgres row-level security ensures a session scoped to Client A cannot return Client B's rows. Isolation binds under the least-privilege brain_app role; the agency starter kit provisions it with one command.