Muhamed Isabegovic
All projects

brain

Gives people and agents one record of what was decided and why, kept current by the agent.

Set this up in your teamSource on GitHubpi-brain
Role
Author
What
A markdown wiki an agent maintains, with a local web app to read it
Built with
Python, Astro; local-first, MIT
Releases
v0.20.0 to v0.31.0, July and August 2026
brain's onboarding, slide one of sixteen: A knowledge base your agent maintains. A markdown wiki an AI agent maintains under a schema, governance gates and an audit trail; humans read the briefing, agents read the verbose pages; local-first.

Agents need a memory that does not drift.

A coding agent starts every session knowing nothing about why the code is the way it is. Notes in chat are gone the next day, and a wiki nobody tends goes stale within weeks.

brain is the pattern of a wiki an LLM maintains, made into a product: governance, health checks, and a queue that collects the work while you are away. The agent keeps it current; you talk to the agent. And no LLM ever runs on a schedule.

What it does.

  1. Sources in, unchanged. The wiki is the synthesis.

    Repositories, documents and conversations land under sources/ and are never rewritten. The agent keeps the wiki: decisions, discussions, state, one shelf per repository, and every claim cites the source it came from. The rulebook is AGENTS.md. You don’t read it; your agent does.

    A decision page in brain: decision, accepted, confidence high, six pages rely on this. Self-maintenance is queue-and-tend: deterministic producers accumulate a per-item inbox; synthesis digests in-session, never on a schedule.
    A decision page, with its status, its confidence and how many pages rely on it.

  2. A graph of what relies on what.

    Every link between pages is in a graph: solid where a person wrote it, dashed where the agent suggested it, and a flag on low-confidence pages the rest leans on. Orphans, hubs and dead ends are listed, so the wiki can be tended, not just written.

    brain's link graph: 75 pages and 313 links, grouped by shelf and ordered by inbound links, with flagged low-confidence nodes.

  3. Checks that run without a model.

    The deterministic half runs on a local timer for free: changes in sibling repositories, pages getting old, link health. Everything that needs a model waits in a queue for the operator’s next session. Validation and source checks are one command each.

    $ python3 tools/brain.py validate && python3 tools/brain.py check

    OK  74 page(s) valid
197 sources checked across 74 pages
  10 skipped (sibling repos not on this machine; use --verbose to list)
    Real output from a clean clone, offline.

  • An optional tier for many agents.

    Off by default: brain can be the hub for many agents, each signing what it writes with its own key, on an append-only event stream, woken when a matching event lands. The brain is never the harness and never runs a model itself.

What it shows about how I work.

brain holds specifications, decisions and handoffs ADR-first, as a versioned repository. It is the working method behind my AI-native work, and where pi-brain started. It is how I think a team should keep its knowledge: written once, cited always, tended in the open.

How I work with knowledge bases

Want this in your team?

I set up a knowledge base your people and agents both work from, in your repositories, and teach your team to tend it.

Get in touch

Want a second opinion?

Start with an architecture review. Fixed scope, one to two weeks, and a plan your team can act on.

About the review

Or email info@muhamed.at