PROJECTS / SECOND BRAIN

An AI that remembers everything I know

Claude Code, Codex, Barnabas, Basic Memory, Syncthing · running daily

HOW A NOTE BECOMES KNOWLEDGE

captureraw inboxBarnabas → wikithis siteData Nerve

WHAT IT DOES

Think of it as a research assistant that never sleeps. It watches for new AI tools and news, teaches itself the ones worth knowing, and files everything as a wiki it keeps current. Anything I hand it, or anything it decides is worth chasing, it researches deep and writes down.

Then when my agents go to build, they read the wiki first. So they are instantly an expert on whatever it has studied. That is the whole trick, and it is why I can build things I had no business building.

PROBLEM

I read a lot and remembered almost none of it. Everything went into folders I never opened again. Note apps just gave me a nicer place to lose things.

BUILD

Everything I capture lands in a raw inbox, sorted by where it came from. I never clean it up. Once a day a librarian agent named Barnabas reads the new mess and compiles it into a clean, linked wiki. He is the only one allowed to write there. One librarian, one pen, zero sync conflicts.

Scraped feeds stay quarantined as dated quotes, never promoted to fact on their own. A scrape isn't something I know just because it landed in my inbox.

WATCH IT RUN

This is the vault with nobody touching the keyboard. Research reports, session handoffs, and scraped feeds land in the raw inbox, get classified, and get tied into the wiki by Barnabas as it happens. Every dot the globe gains starts as one of these files.

screen recording of the live vault · automated research, handoffs, and wiki compilation

HOW IT RUNS

Capture. Everything lands in one raw inbox, sorted by where it came from: session handoffs, outside sources, my own notes, screenshots, and scraped feeds. I never organize it by hand.

Ingest. Scheduled jobs pull in the outside world on their own: YouTube transcripts, AI news, and a daily radar of new tools. They run on schedules, not on me remembering.

Compile. Once a day Barnabas reads the new mess, classifies it, and ties it into the wiki with real links. Still the only one with the pen.

Trust. Scraped material stays quarantined as dated quotes. It only becomes fact if I capture it or research it on purpose.

Handoffs. Every working session ends with a short handoff note. Barnabas folds it into that project's living state page, so nothing gets lost between sessions.

Prove it. A 49-case retrieval eval keeps the whole thing honest, and the vault publishes its stats nightly. The globe on the home page reads from that feed.

WHAT BROKE

The scoreboard. Input side: 1103 raw files, 25 scheduled jobs, a 49-case retrieval eval scoring 79% recall (how often it pulls back the right note when I ask). Output side: the newsletter all of it was supposed to feed had two subscribers. Elite intake, two people watching.

And that 79% has a real 21% behind it. Most brains never get tested, so they never find out.

RECEIPTS

Raw files indexed1103
Retrieval eval49 cases
Retrieval recall79%
Scheduled jobs25
Newsletter subscribers, at the time2

Numbers pulled from the live vault. Including the one with two people in it.

LIVE RIGHT NOW

0raw files
0wiki pages
0scheduled jobs
0%recall
  • 2026-07-16 Daily Thoughts pipeline repair session2026-07-18 · sessions
  • 2026-07-17 Fanbase Weather collector T22 land T23 decision session2026-07-18 · sessions
  • 2026-07-17 Fanbase Weather collector T22 forums cron session2026-07-18 · sessions
  • 2026-07-17 Public Data Project Selection for dustincoledata Research2026-07-18 · sources
  • 2026-07-17 DCD Round-3 Variants Build Log2026-07-18 · captures

synced from the vault, nightly