How Norma Remembers: Designing Memory for a Personal Assistant

Our memory design fits on a sticky note:

Keep the evidence. Recompute the understanding.

When you tell Norma something, it keeps your actual words, with when you said them. That is the evidence. On top of it, Norma builds its current understanding: Vera's birthday is 27 November. You're allergic to nuts. You switched from coffee to tea (again).

Why keep both? Because understanding changes and evidence doesn't. If you say "I drink coffee in the morning", then "actually I've switched to tea", then "mostly green tea now", Norma shouldn't have three competing opinions about your beverage life. It should know the current answer and still remember how it got there. Your words are the source of truth. Norma's summary of you is always rebuildable from them, which also means it can always be corrected.

Early versions of our memory tried to squeeze everything into neat little boxes: attribute: drink, value: tea. Tidy. Also wrong surprisingly often. Real life doesn't fit in boxes. "I'm allergic to nuts" is not the same as "I don't like nuts", and a dog called Bosse is not a species.

So Norma now stores what you said as a sentence, in your words, and only adds structure where it is obvious (a date is a date). The result: fewer weird misunderstandings, and no more of your pet being filed under biology.

You shouldn't have to say "remember this" for Norma to remember. But saving every sentence the moment you type it turned out to be a bit like taking notes during a conversation without listening to the end of it. People correct themselves. They say "Tuesday, no wait, Wednesday".

So Norma writes memory once a conversation settles, looking at the whole exchange, the way a good assistant would jot notes after a meeting rather than mid-sentence. And when you do say "remember this" or "forget that", it happens right away.

Ask "what should I cook tonight?" and a good assistant remembers your nut allergy, even though you never said the word "allergy". Norma looks things up the way you would expect a person to: by meaning, by names and nicknames ("my sister", "Vera", same person), and by time (what's true now, not what was true last spring).

And while your memory is still small, Norma simply keeps all of it in view. No clever filtering, no chance of the one important thing falling through the cracks. The fancy search only kicks in once there is a lot to search.

A few kinds of memory get VIP treatment: allergies, diets, budgets, "never call before 10", "don't book flights unless you have to". We call them constraints. They are always considered, whatever you are asking about. Forgetting your coffee order is awkward. Forgetting your allergy is not an option.

Tell Norma to forget something and it's gone. Not hidden, not "archived", not lingering in a summary it wrote last Tuesday. Forgetting reaches every place the information could have spread, and we test that on purpose: forget something, then go looking for it. That test has to come back empty, every time.

We also measure the opposite failure: Norma inventing or overgeneralising. If you say "I like sushi when we're in Spain", Norma should not decide you are a sushi person. If you say "maybe Japan next year", Norma should not book you a flight. A memory that is confidently wrong is worse than no memory at all.

Memory is one part of a bigger picture. Norma is built from five pieces, and we keep them deliberately separate:

Memory knows your world: people, preferences, dates, constraints.

Workstreams know what's going on: the dinner being planned, the renovation, the job search.

Tasks own a specific outcome: "find a date for the team dinner and book a table".

The activity log records everything Norma did, so it can explain itself ("why did you do that?") and pick up exactly where it left off, even if a server restarts mid-thought.

The gateway decides what Norma is allowed to do out in the world: send an email, add a calendar event, or nothing at all without your yes.

Put simply: memory describes your life, and the rest describes Norma's work in it. Keeping those apart is what lets Norma be helpful without being creepy, and powerful without being reckless.

Good memory is useless if it costs a fortune to run. So every step in Norma's pipeline uses the smallest tool that does the job well: plain database lookups where no AI is needed, a tiny fast model for quick yes/no decisions, and a strong model only where judgment matters. Every improvement we ship has to be at least as good and no more expensive than what it replaces. That is how we plan to make an assistant this capable available to everyone, not just the few who can afford it.