How I Built a Personal AI Brain
The site you’re reading is written mostly by a machine that reads my notes. Here’s how it works, why I built it, and what I got wrong the first three times.
I have spent fifteen years building software for an industry that had to invent itself in public. You learn a few things doing that. One of them is that the bottleneck was never ideas — it was remembering what I already knew. Every good thought I ever had went into a notes app, a voice memo, a napkin, a Slack DM to myself, and then into the void.
So I built a brain that doesn’t forget. Not a chatbot. A place — one my AI agents read from, write to, and keep tidy while I sleep.
The idea I stole
This isn’t original. Andrej Karpathy published a gist describing an “LLM wiki” — a folder of markdown files an AI maintains as long-term memory. Craig Dos built a beautiful public version of it. I looked at both and thought: that’s the shape of it. A brain is not a search index. It’s a set of pages that link to each other and get better every time you touch them.
The whole system is three folders and one rule.
raw/— everything I capture, untouched. Voice memos, clipped articles, PDFs, half-thoughts. I drop; I never organize.wiki/— pages the AI writes by readingraw/. Interlinked, deduplicated, revised. This is the actual memory.reports/— the polished, standalone pieces. Like this one.
The rule: nothing is public until I say so. Every page is born private. The AI can suggest something is worth sharing, but a human — me — flips the switch.
The loop
Once a week I run a single command. The AI sweeps my inbox, transcribes any audio, and for each new source it writes a summary page — then it stops and shows me the takeaways before it commits anything. It cross-links the new material into the existing wiki, updates the index, and flags anything it thinks the public would find useful.
Then I read. I flip a few pages to publish: true. It rebuilds the site, runs a safety pass, and deploys.
That safety pass matters more than any of the clever parts. Before anything goes live, a script greps the built output for a deny-list — client names, dollar figures, family details — and refuses to deploy if it finds a single one. The failure mode of a public brain isn’t that it’s boring. It’s that it leaks. So the leak-guard isn’t a feature; it’s the whole license to run the thing at all.
A public second brain is only as good as the wall between what it knows and what it says.
What I got wrong
I tried to make it one site. The brain and my actual homepage wanted to be the same thing. They don’t. A wiki that grows in public is the wrong front door for a person — it reads like a filing cabinet. So the polished, deliberate stuff lives on the front page, and the living, exploratory brain lives one step behind it. Two engines. One flows into the other.
I tried to organize the raw layer. Wasted weeks. The entire point of raw/ is that capturing has to be frictionless or you stop doing it. Organizing is the AI’s job, not mine. My job is to drop the thing and walk away.
I over-built before I had anything to remember. I designed workflows for content that didn’t exist yet. Empty rooms. Now the rule is: put one real thing in before you build the shelf for a hundred.
Why bother
Because compounding only works if nothing leaks out of the system. Every note I take now makes every future note more useful, because the AI connects them. The report you’re reading will get better — I’ll come back to it, the brain will have learned more, and the next version will be sharper than this one. That’s the point. It’s not a document. It’s a living draft that improves every time I think.
I don’t trust my memory anymore. I trust the folder.
This is a living document — first published during the site’s launch, and updated as the brain grows. If it reads differently next time you visit, that’s the system working.