Privacy · 21 July 2026 · 5 min read
AI Memory Is Useful. It's Also a Privacy Question.
Persistent memory makes assistants dramatically better and quietly builds the most detailed profile anyone holds of you.
Memory is the feature that makes an assistant feel like it knows you. It remembers your work, your projects, your preferences, the way you like things written. It is genuinely one of the most useful things to happen to consumer AI.
What memory actually is
It is a profile. Built continuously, from the most candid text you produce, tied to an account with your name and payment details on it. Search history was already a revealing dataset. Chat history is more revealing, because people explain themselves to an assistant in a way they never did to a search box.
The three questions to ask any product with memory
- Where is it stored, and who can read it?
- Is it linked to your identity, or to an anonymous session?
- Can it be handed over: to an advertiser, an acquirer, or a court?
It does not have to be a trade
The assumption that useful memory requires a company holding your history is an architectural choice, not a law of nature. History can live encrypted on your own device. Identity can be separated from content so that the party who knows who you are cannot read what you asked, and the party who reads what you asked does not know who you are.
That is the design we build to. Not because memory is bad, but because it is too useful to give up and too sensitive to hand over.