Every assistant now promises to remember you. Almost none will show you the notebook. Memory you can't read is surveillance with better branding.
The pitch for AI memory writes itself: stop re-explaining your context, your people, your projects to a tool that forgets you between sessions. It is a real problem, and remembering is genuinely the fix. But look closely at how most assistants implement it and you find a one-way mirror. The model accumulates impressions of you somewhere in a vector store, and you get no list to read, no row to delete, no way to know what it thinks it knows.
Memory you can’t read is surveillance with better branding.
That opacity is not a small UX gap. Memory is the highest-stakes feature an assistant has, because it is the only one that outlives the session. A bad draft costs you ten seconds. A wrong or stale memory quietly shapes every draft after it. And a memory you never agreed to keeps a record you never agreed to keep.
Propose, don’t hoard
We think the fix is to treat memory like an inbox, not a warehouse. When Cue notices something that looks worth keeping, a preference, a person, a promise you made, it proposes it: a card that says what it noticed and where it noticed it, with a Remember button and a No thanks button. Nothing enters memory by default. The model suggests; you curate.
- Every fact is proposed with its source, and kept only when you approve it
- Every kept fact is plain text you can open, edit, or delete
- Excluded apps are excluded from learning entirely, not just from display
- Delete means gone: the fact is removed, not soft-hidden and still embedded
Local-first is the enforcement
A policy page can promise anything. Architecture is the promise you cannot break. Everything Cue learns about you, voice profiles, people, memory, the promises it tracks, lives in files on your Mac. Our servers meter usage by count, not content; there is no server-side profile of you to subpoena, breach, or train on, because one never exists.
The industry will spend the next few years adding memory to everything, because it is the feature users actually want. The dividing line will not be who remembers more. It will be who lets you read the notebook. An assistant that remembers you should be able to show you, in plain words, exactly what it remembers, and take a fact back the moment you tell it to. Anything less is not memory. It is a file being kept on you.