I saved two years of my own AI conversations. Then I had an AI read all of them and write the manual on how I think. Every claim in the manual points back to the exact talk it came from.
My agents load that manual so they know who they work with. It was built in ten days with a team of AI agents doing the reading.
The index
1,601 conversations, 21,513 messages spanning April 2024 to June 2026, parsed into 19,810 chunks and embedded at 4,096 dimensions. Storage is SQLite with brute-force cosine similarity in TypeScript. A full scan at this scale completes in under 100ms, so a vector database would have been a dependency without a job. Embedding runs in checkpointed batches and resumes from wherever it stopped, and the parser is adapter-shaped so another assistant's export drops into the same schema. Monthly theme classification runs every human turn against 69 hand-authored concept anchors across 14 life domains.
The mind model
On top of the index, a multi-agent workflow (scout, dig, synthesize, with a coverage audit between stages that caught and backfilled a real gap) wrote roughly 110 dated dossiers where every claim cites a conversation id that reopens the source transcript. Those distill into one model per domain and a small always-loaded wiki. Three tiers of retrieval: load the wiki whole, zoom into a domain model on demand, fall back to vector search last.
Questions about this system, or the problem it could solve for you?
Discuss your project