How much of a person fits in their conversations? I had two years of mine saved, so I ran the experiment. An AI read every one of them and wrote the manual on how I think. Every claim in that manual points back to the exact conversation it came from.
The answer was: more than I expected. Enough to write down who I am, how I decide, and when I need a push. My agents load that manual so they can act on my behalf and get it right. It was built in ten days by a team of AI agents doing the reading.
The index
The raw material is every AI conversation I had for two years: 1,601 conversations and 21,513 messages, April 2024 through June 2026. Each one was cut into searchable pieces, 19,810 of them, and embedded, which means turned into coordinates that let a computer measure how close two pieces of text are in meaning. It all lives in one small database, and a full search finishes in under a tenth of a second, which is exactly why there is no heavy search infrastructure underneath. It would have been a dependency without a job. Every message also gets sorted against 69 hand-written concept anchors across 14 areas of life, so themes can be tracked month by month.
The mind model
On top of the index, a team of AI agents read everything and wrote roughly 110 dated dossiers, where every claim cites the exact conversation it came from, and the citation reopens the source. The workflow ran in stages. Scouts found the material, diggers went deep, a synthesizer wrote it up, and an audit between stages caught a real gap and sent the team back for it. The dossiers distill into one model per area of life, plus a living layer small enough to load whole. One page for who I am, one for how to work with me, one for the decisions I have already made, and one for what I am working on right now. When my agents act on my behalf, that is what they are reading. They open a deeper domain model when they need it, and fall back to raw search last.
Questions about this system, or the problem it could solve for you?
Discuss your project