Skip to content
Matt Zimak Book an intro call (opens in a new tab)

Matt Zimak Automations A company brain

A company brain, updated by every meeting.

TRANSCRIPTFACTSSAFE EDITCORE CHANGEAPPROVEDCONTEXTMeetingTranscribedn8nExtracts factsCuratorClaude, nightlyThe brainOne GitHub repoApprovalCore changes onlyEvery agentClaude, Claude Code

A company's knowledge base in one GitHub repo that people and AI agents read alike: strategy, decisions, customers, suppliers, people. After every meeting the new facts go in with their source, and a change to strategy, decisions or finance waits for a person to approve it in Slack. Fed by every meeting of a family business group since July 2026; four more companies run on the same structure.

The problem.

What a company knows lives in meetings, chats and a few people's heads. A new hire, an adviser or an AI agent starts from zero, and a decision made in March gets argued again in June. A wiki helps until nobody updates it.

How it works.

One repo per company
Strategy, products, brand, goals, customers, suppliers, people and decisions, one file per thing, in plain text that people and agents read alike.
Fed by every meeting
Fireflies records the meeting. n8n files the transcript in the repo and pulls out the facts, each with the sentence it came from.
A curator every night
A Claude agent folds the new facts into the right files: dated timeline entries, meeting notes, new contacts and suppliers as drafts, each citing its meeting. Once a week it writes a short brief per company.
A person for what matters
A change to strategy, a decision, the goals or finance becomes a pull request and a card in Slack, in the team's language, with Approve and Reject. Approve merges it; nothing core changes without it.
Any agent
Ask it in a Claude chat, edit it with Claude Code, or point Cursor, Codex or any other agent at the repo. It is plain files in Git, so no tool owns it.

What you need.

A GitHub organisation
Free. One private repo per company, yours from day one, seeded from the open-source template.
Meetings with transcripts
Fireflies here; any recorder that writes a transcript works. n8n files it in the repo.
A Slack channel
Where the approval cards arrive, with Approve and Reject, and one person who clicks.
n8n
Your own account, set up in the build, for the transcript and fact pipeline.
A scheduled Claude agent
The nightly curator. Reading the brain works from a Claude chat, Claude Code or any other agent.

What it takes.

  • The repo first, the loop after.

    The brain is seeded from the template and your documents; the meeting loop joins once transcripts flow. Here: the repo in June 2026, the loop on 7 July.

  • One click per core change.

    Everyday facts file themselves. Strategy, decisions and finance wait for a person's Approve in Slack.

  • A fixed audit fee, credited to the build.

    Scoped from the audit, never priced blind; a fractional seat is a monthly retainer. You hear the numbers on the intro call.

Running today.

  • 149 meetings folded in.

    Since July 2026, each fact with the sentence it came from, plus a weekly brief per company.

  • Five brains, one structure.

    A family business group with four companies, a design studio group, and my own three companies, each on the same template.

  • Open source.

    The repo template and the meeting loop are public on GitHub: template-intelligence and brain-curation-loop.

The meeting transcripts also feed the CEO's AI assistant.

Questions owners ask.

What goes in on its own, and what waits for a person?

Timelines, meeting notes, new contacts and suppliers go in as drafts with their source. A change to strategy, a decision, the goals or finance becomes a pull request and a card in Slack, and nothing core changes without a click on Approve.

Can the agent invent a fact?

No source, no write. A missing number, date or name stays missing, an uncertain fact is flagged and goes to a person, and a decision, once merged, is never rewritten. A separate weekly pass looks for broken links, duplicates and contradictions.

What about sensitive topics in a meeting?

Pay, health and family topics are noted as a sensitive topic and routed to a person, never quoted.

Which tools can read it?

Any agent: a Claude chat, Claude Code, Cursor or Codex, and people in their editor, on GitHub or in Obsidian. It is plain files in Git, so no tool owns it and every change can be undone.

Start with a call.

Fifteen minutes on what your company knows, where it lives today, and which meetings should feed it first.