The Librarian Who Never Forgets
Part 1 of 2. Written in June 2026, cleaned up in November. Part 2 tests what I believed here.
I built a second brain.
It’s a folder of plain text files. An AI reads them, writes them, links them and keeps them tidy. I feed it what I read, watch and think, and it gives me back a version of my own head that doesn’t forget.
Two leaks
I was consuming a lot and keeping almost nothing. The talk that changed my mind last year: gone. The thought from a long walk: figured out again months later, because I couldn’t find where I’d written it down.
So in early 2025 I built a tiny tool. Paste a YouTube or web link, get back a clean note. Written for me, filtered through a profile of what I care about.
The second leak was older. Since 2019 I’ve written a review every week. What I learned, what I got wrong, what’s shifting.
That’s a different memory: what I lived.
For years those sat in 2 separate places. The breakthrough was putting them in one. One system that remembers what I take in and what I work out, and links the two.
The tool kept growing. It learned to transcribe podcasts, find new sources, learn how I think. At some point I had an app I had to operate. I wanted a system that operates for me.
So I threw the app away and rebuilt it as files.
It went from an app I operated to a system that operates for me.
The picture
A private library. The shelves are mine, and I can walk in and read anything myself.
In it lives a librarian who never gets tired. Hand them a video, an article, a messy thought. They read it, file it, write a summary card and cross-reference it to everything else, including shelves I’d never have connected.
Ask a question and they answer from my own collection. Come back tomorrow and they remember the conversation.
The librarian is the AI. And the library is arranged for the librarian: about 80% of it is built for the machine, 20% for me. I get a better library because I stopped arranging it for browsing.
(Yes. A Zettelkasten with an AI bolted on. Every German nerd ends up here.)
What it does
- Ingests anything. Article, video, podcast, PDF, book. Out comes a note shaped to my profile, plus the verbatim source so nothing dies when a link rots.
- Compiles its own wiki. 10 notes on one topic get braided into 1 article, linked to the rest, updated when new sources arrive. My AI engineering article went from 49 to 64 sources in one week because my own work sessions feed it.
- Answers from my head. Questions get answered from what I’ve collected, in my words. Months after Chiang Mai I can ask “what was that café I loved?” and get the name, because the trip got compiled into a guide.
- Researches places. It searches in the local language, uses the words specialty places use about themselves, then checks the place is still open. Blogs keep dead cafés alive for months.
- Finds new things. It watches sources I trust, matches them to my profile and mixes in a few picks that should surprise me, so it stays a window.
- Remembers my work. Every session with the AI gets summarized and filed back. My work compounds the same way my reading does.
- Digests my past. 8 years of consulting, about 1,600 files, distilled into craft I can query.
- Writes in my voice. It knows I have a raw voice for my notes and a cleaner one for essays like this.
- Is a mirror. Years of weekly reviews in one place means it can show me patterns I can’t see: a worry that keeps coming back across years, a value that’s slowly moving. None of that is in here and it won’t be.
It compounds. A notes app is the same on day 1,000 as on day 1. This one gets better with every input.
How it works
3 ideas carry almost everything.
Judgment and machinery. The AI is great at judgment: reading, connecting, writing. Ask it the same factual question twice and you can get 2 answers. Fine for thinking, bad for facts.
So anything that has to be exact (is this café open, what’s in this transcript) is a small program. Same input, same answer, near zero cost. “Which place feels like me” stays with the AI.
Shelves and memory. The shelves are plain markdown files. They’re the truth and I own them. The memory is a search index on a server (gbrain) that gives every session fast recall across the whole library.
The memory can be rebuilt from the shelves any time. If it vanished tomorrow, nothing would be lost.
Plain text outlives any AI. Files and small programs, welded to nothing. It runs on Claude and Codex today. It’ll run on whatever comes next.
What I believed in June
I’m keeping this as I wrote it, because part 2 checks it.
My consulting craft, 8 years of it, becomes something the AI can draw on, so I never start from a blank page.
And the newest piece: I drop a task into an inbox, and the system does it or drafts it for my ok. The librarian starts doing. I thought that was the direction everything pointed.
It started as an app I operated. I believed it would become the engine my work runs on.
I wrote this so the people close to me could see what I disappear into.
Appendix: the machine room
- Vault: plain markdown, the only place real content lives. Obsidian is just the viewer.
- Skills: packaged procedures (clip, compile, research, review) loaded only when needed. A thin always-on layer, fat skills.
- Scripts: small Python programs for the exact work: place discovery and liveness, transcripts, session analysis, sync.
- gbrain: Garry Tan’s open-source brain, on a Hetzner ARM box. Postgres with pgvector for meaning search, keyword search and a link graph. Re-indexes the vault every 30 minutes, served to every AI session over MCP. Derived and rebuildable.
- Sync: Mac, phone and server, versioned. On the phone I open it in Obsidian a few times a week.
- Lineage: generation 1 was a Next.js app (May 2025 to February 2026) with a vector store, a discovery scorer and transcription. The patterns survived, the app didn’t.