Research note · 2026-10-06

Memory with a source, an owner and an expiry

Memory is the feature that makes an assistant feel like it knows you. It is also the feature most likely to make it wrong in a way you cannot see.

The question people actually ask is how the assistant knows something. Most memory systems cannot answer, because the fact was distilled from many interactions and the origin was thrown away.

We propose three properties. Every memory keeps a pointer to its source. Every memory has an owner and a scope, so it can be limited to one task or one person. And every memory can expire or be removed, with a record that it was removed.

Provenance does not make a memory true. It makes it checkable, and it makes correction cheap: fix the source and everything derived from it can be re-derived.

The hard part is what to do with summaries, where many sources collapse into one sentence. That is the central open question for this direction.

Related project: Recall