The story
01The idea
One place to put things and one place to ask. A thought becomes a capture you can review before it is filed; a question becomes an answer that cites the notes it came from. Neither path hides the records underneath.
02Capture you can review
When you capture something, the system proposes what to do with it: skip it, ask you a question, or create or update a note. New types, folders and people stay as proposals until you accept them. Deciding whether a name refers to an existing person is your call, not the model’s.
03Answers with receipts
Every note is chunked and indexed by meaning as well as by keyword, and the two kinds of result are merged before the answering model sees them. Every source in an answer keeps a link back to the exact note, conversation or thread it came from.
04Postgres is the truth
Collaborative editing runs through Yjs and Hocuspocus, but the database is canonical: an empty live session can never overwrite a populated stored note. Every change is checkpointed and attributed to a person, an AI agent or a system job, and AI edits keep the instruction that produced them.
Export rebuilds the whole archive as a Markdown vault with folders, frontmatter, tags and wiki links, readable in Obsidian or anything else.
05The hard part
Generating an answer is easy. Keeping capture reversible, retrieval attributable, live edits recoverable and the archive useful outside the product that created it is what shaped the database, the review screens, the prompts, the version history and the export format together.