Personal assistant on your own memory.
No cloud. No surveillance.

Cloud assistants know everything about you, but forget you with every new chat. Local assistants remember, but don't understand the context of your life. I'm building a third way — memory as a separate layer, owned by the human, not by the model.

What's wrong with current assistants

They wipe the session. They send your input to someone else's data center. They hallucinate, because they can't separate «what was told to me» from «what I made up myself». They equally forget your child's birthday and a random three-year-old question.

What I'm proposing

Memory with provenance: every fact tagged by source (observation / conversation / model's own conclusion). At decode-time, confabulation is blocked — the model cannot pass its own conclusion off as a fact. Decay differs by memory type, emotional weight shifts priority.

Storage on your hardware or your VPS. AGPL. Ready for FZ-152 on the territory of Russia.

Method

Not a patent, but a defensive publication. arXiv preprint in 6 months. In parallel — a series of articles on Habr and dev.to (Russian and English versions of each). I'm learning English by translating my own texts — you can see this publicly in the «English» section.

What already works

Memory Core on VDSina, brain-mcp, ingestion from Telegram via the Heckfy listener, provenance tags in the graph. The second-brain bot with company-profile assistants (alfa-sk, mech-stroy, intero). Iteration 8 of 27 in the L3 multi-tenant refactor.