The whole rig — every input, where it goes, and what's still loose wire

Three flows. The big one is the knowledge pipeline: anything you capture funnels through one front door, lands in the vault, gets distilled, indexed, and becomes recallable. Drag any box to rearrange it; the wires follow. Click Fix / align to snap back to the clean layout.

snapshot 2026-06-05 · read live from disk · scroll the canvas sideways if cut off →

automatic & live works, but you trigger it by hand today not wired yet (build list)
drag boxes anywhere — Align tidies them onto a grid in the order you placed them snapshot layout: clean
1 · enter
2 · fetch
3 · front door
4 · vault (store)
5 · distill
6 · index
7 · recall

Side automation — weekly call commitments runs on its own in the cloud, separate from the brain

ZOOM
weekly call
team check-out chat
DRV
Google Drive folder
newest call chat file
GAS
Apps Script
parses commitments · Thu 8am + 10am PT
TG
Telegram group
your team group · your bot

Local fallbacks are parked on purpose: a local script + a launchd job — superseded by the cloud Apps Script, kept as backup.

Follow the wire — what happens, hop by hop

1
It enters. A YouTube link, a doc/PDF/web page, an Instagram reel, a voice memo, a quick brain-dump — or the one-time legacy imports already sitting in the vault.
2
Fetch the words. Links → yt-dlp · reels → instagram-station · audio → Whisper. Docs and web pages skip this and go straight to the front door.
3
One front door. markitdown flattens everything into clean markdown so each downstream step sees the same shape.
4
Store the raw. It lands in Obsidian/Inbox with a dated title — the truth, never rewritten. Brain-dumps and legacy notes drop straight into the vault.
5
Distill. graphify reads the raw note on local Ollama, chunks by meaning, and writes tidy concept notes into /Distilled. $0, offline.
6
Index — distilled only. gbrain embeds /Distilled (not the raw) so it's searchable by meaning.
7
Recall. You ask — phone (Hermes) or desktop (Claude Code). It searches gbrain first, answers with the note cited. Loop closed.

What's automatic vs a manual hop right now

Runs by itself

  • Every entry point + the front door — yt-dlp, instagram-station, Whisper, markitdown all work on command.
  • Recall — Hermes and Claude Code both query gbrain live.
  • The weekly call job — fully automatic in the cloud every Thursday.

You nudge it by hand

  • front door → vault — the wire half-lands: capture currently writes into fpm-ai's folder, not the vault yet.
  • vault → graphify — you run the distiller manually; nothing watches /Inbox.

Not wired yet (build list)

  • /Distilled → gbrain auto-sync — point gbrain at the vault + a scheduled re-index so notes appear in search with no command.
  • Green health lamp — one boot check that says the whole pipe is live.

Two sibling systems they sit beside the brain — don't confuse them with it

purpose.ai — the product

~/purpose.ai/ · Supabase + Claude SDK

The live product, shipping daily. Its database is product data (player profiles, dashboards) — a totally different store from your personal brain. The checker / auditor from fpm-ai graduate here, where verification protects real users. Everything in the rig exists to ship this.

fpm-ai — the control layer

~/fpm-ai/ · memory backend = gbrain

Your multi-agent framework: orchestrator + checker (OK/WARN/BLOCK on every output) + auditor. For the personal brain it's mostly held OFF on purpose (lean over machinery). It owns the intake organs (ingest / transcribe) that feed the pipeline above.