brain-curation-loop

Meetings in, a curated company brain out, with a human approving anything sensitive.

A company brain is a git repo of Markdown that AI agents read and update: one file per customer, competitor, supplier, person, decision and meeting, in the template-intelligence layout. This repo is the loop that keeps such a brain current without a person typing into it, and without an agent ever changing what the company is on its own.

Running in production since July 2026 on one client brain and three company brains.

The loop

meeting transcript (Granola, Fireflies, an upload form, ...)
        |
        v
 1. Knowledge-extract tap  (n8n)         structured facts with a verbatim quote each,
        |                                 staged as inbox/extracted/<date>-<meeting>.md
        v                                 with curated: false. Always-safe tier.
 2. Curator                (nightly cloud agent)
        |                                 folds staged facts into canonical pages.
        |                                 SAFE tier: commits to main with a citation.
        |                                 SENSITIVE tier: branch bot/<slug> + pull request.
        v
 3. PR approval in Slack   (n8n)         GitHub webhook -> Approve / Reject card ->
        |                                 squash-merge or close, from the chat.
        v
 canonical brain on main, every fact citing the meeting it came from

The rules that decide SAFE versus SENSITIVE live in the brain repo itself (docs/CURATION.md in the template), not in the agent's prompt, so the same curator can serve any brain and the company controls its own tiers.

What is in this repo

Path	Piece	Runs on
n8n/knowledge-extract-tap.md	Step 1. The agent node's system prompt, the structured-output schema, and the drop-in Code node that writes the staging file through the GitHub API	n8n, any LLM node with structured output
routines/curator-prompt.md	Step 2. The curator's prompt: rules files first, fold, tier, mark provenance, weekly note	a scheduled cloud agent with git and gh
n8n/pr-approval-in-slack.json	Step 3. Importable workflow (9 working nodes plus 3 notes): pull_request webhook, bot/* filter, Block Kit card, Slack action webhook, merge or close	n8n
docs/setup.md	Wiring the three pieces to a brain repo, a Slack app and a GitHub token
docs/design.md	Why the staging tier, why tiers live in the repo, why a quote per fact, what we got wrong first

Install, short version

1. Give the brain repo the curation contract: copy docs/CURATION.md from template-intelligence (or start the brain from the template).
2. n8n: add the extract tap to your transcript pipeline after the summary step (n8n/knowledge-extract-tap.md); import n8n/pr-approval-in-slack.json, replace the REPLACE_WITH_* placeholders, register the two webhook URLs.
3. Schedule the curator with routines/curator-prompt.md on a nightly cron against the brain repo.
4. Run one meeting through. Expect a file in inbox/extracted/, then a commit or a Slack card the next morning.

Full steps with the exact GitHub and Slack settings: docs/setup.md.

Related

• template-intelligence: the brain layout this loop writes to, including the curation contract.
• gbrain-company-brain: a gbrain skillpack that loads the same brain into gbrain with its types and relationships intact, so agents can query it.

License

MIT.