I built a shared memory for my home's AI assistants, and the hard part wasn't remembering, it's who may know what. Every fact carries who it's about, who may read it, when it expires, and each reader sees a different view of the same page. Self-hosted: https://t.co/MLkVOoyqPn
@hwchase17 An open standard is the right call. I've run Markdown wiki memory in prod since spring (multi-user, over MCP) — filed a producer report on the axis the tracker keeps circling without naming: access, and what a conformant consumer must do about it. https://t.co/jYYqu2DmLM
@pauliusztin_@hwchase17@LangChain And there's no separate update path: a correction is just one more message, it lands as a new fact and supersedes its predecessor. Concurrent writers go through an append-only op-log with an optimistic head check. Engine's AGPL Rust: https://t.co/HHdaQqcs10
@pauliusztin_@hwchase17@LangChain Drift belongs to a nightly REM cycle: dedup, promote durable facts into topic pages, recompile the prose. A rewrite can't lose knowledge: SQLite owns the facts, attribution, validity windows, embeddings, the Markdown prose is their render, the narrative the navigator walks.
@TransformLabsHQ 'Unpatchable' is the wrong frame. It's unpatchable only if access control lives in the prompt, and a prompt is a suggestion. Move redaction into the engine, before text reaches the model, and there's nothing left to exfiltrate. The boundary has to sit below the model, not in it.
@NickADobos Files as the surface, 100%. One thing that bit me shipping it: don't make the raw file the source of truth for retrieval. Keep an engine index authoritative, files as the readable render. Otherwise hand-edits and drift quietly rot recall. Files for humans, index for the machine.
@pauliusztin_@hwchase17@LangChain So the two collapse into one: the compiler writes prose with explicit relationships and links precisely so there's something to walk at recall time. Good writing isn't cosmetic, it's what makes navigation possible. Retrieval finds the door, navigation walks the house.
@pauliusztin_@hwchase17@LangChain Wiki memory in prod since spring here. Two things surprised me building it, and they're really one thing. First: recall is only as good as how well the wiki is written. A prose problem, not an embeddings one, relationships spelled out in narrative are what make retrieval land.
@pauliusztin_@hwchase17@LangChain Second: written well, you can navigate it, not just retrieve from it. Most wiki memory still does flat top-k over chunks. A written wiki lets a navigator walk pages and links like a person, so the divergent facts surface, the cancelled trip, the allergy behind a dinner plan.