What a personal AI should remember, and what it should forget

The idea

Ask people what they want from AI memory and they say: remember everything. Build that and it is worse than you expect. A system that keeps every fact equally forever cannot tell you what matters, because it has no opinion about relevance and no way to know that something stopped being true.

Memory is the part of this we have rewritten most often, and the lesson each time was the same. The interesting question is not how much you keep. It is what kind of thing each memory is.

Four kinds, four lifespans

• Identity. Who is in your life and how they relate to you. Slow-changing, high value, worth being careful about. Get this wrong and every answer inherits the error.
• Commitments. What you said you would do, and by when. High value and sharply perishable. A commitment that has passed should stop being an active memory and become history.
• Preferences. How you like things: phrasing, timing, what you never want raised at 11pm. These accumulate quietly and are what makes an assistant feel like yours rather than like a product.
• Episodes. What happened. Last night's sleep, Tuesday's walk, the conversation about the school. Individually low value, collectively where every pattern comes from. Keep the shape, do not keep every reading forever.

Continued

Most disappointing memory features treat all four as one thing, which is why they end up feeling like a notes app you cannot search properly.

Supersession is the whole game

Facts about a life do not accumulate, they replace each other. You moved. You changed jobs. The plan from March is not the plan. If the system holds both versions with equal confidence, it will sometimes answer with the old one, and one confidently stale answer costs more trust than ten helpful ones earn.

So new information has to be able to retire old information, and the retired version has to still be reachable, because knowing what changed is often the actual answer. We ended up applying the same instinct to documents in chat: when Souvéa regenerates something, it updates in place and keeps the earlier versions in a history you can open. Memory wants the same treatment. One current answer, a trail behind it.

What should be forgotten

Some things should expire without being asked to. Details of a resolved logistical thread. Precise readings once the trend is what matters. Anything you deleted at the source, which should not survive in a derived layer somewhere.

And some things should never be kept in the first place. A personal AI does not need a permanent record of a bad week to be useful next week. Forgetting is not a gap in the product. It is part of what makes it liveable.

If you cannot look, it is not memory

Every memory system infers things, and some of them will be wrong. It will decide two people are the same person. It will hold a preference you no longer have.

The only workable answer is that you can look at what it thinks it knows, correct it, and delete it. We used to treat that as a trust feature. It is really quality control, because you are the only one who can tell the difference between a good inference and a plausible one. It is also the reason we keep pushing Souvéa's reasoning into view rather than hiding it: an assistant that shows its work can be argued with.

Continued

None of this is finished. Ask us again in six months and the shape will have changed, but the principle probably will not: keep less than you can, be explicit about what replaced what, and let the person whose life it is see the file.