Memento's Problem Was Never the Notes

A friend who builds language models recently compared today's AI to Memento. The hero can't form new memories, so every few minutes his mind resets. To survive, he surrounds himself with notes, Polaroids and tattoos.

The comparison is sharp. A model's knowledge freezes the day its training ends. Everything after that lives outside it: retrieval, harnesses, carefully managed context. It works, and it works well enough that companies built on it are worth billions.

His point was that retrieval is not learning. Learning means compressing what you saw into concepts you can use somewhere new. Finding the right document is not the same thing. So the future, he argued, is continual learning: models that keep learning after they ship.

I agree with half of that. Retrieval is not learning. But I think the film is making a different point.

Rewatch the film

Leonard's tragedy isn't that he relies on notes. Notes work. He solves problems with them every day.

The tragedy is that his notes go stale. Some contradict each other. He can't tell which ones he wrote himself and which ones someone else got him to write. When a note turns out to be wrong, nothing tells him. He just keeps acting on it.

That is the problem I see in AI agents today, and it isn't solved by a better memory inside the model.

Two different jobs

There are really two jobs here.

The first is learning new skills: a new language, a new kind of task, the tacit "how we do things" that's hard to put into words. That belongs inside the model, and continual learning is the right research direction for it.

The second is knowing what is true in your world right now. Who owns the renewal this week. What you decided yesterday. Which plan replaced the old one. That is a different kind of knowledge, and I don't think it belongs inside any model, even one that learns on the fly.

Why your world shouldn't live in the weights

You switch models. People already use several agents and move to a better one every few months. Whatever one model learned about you stays with that model. Your understanding should move with you.

You need to see where it came from. A fact in a model's weights can't point to the email or meeting behind it. You can't check it, correct it once, or delete it when someone asks.

It changes every week. Who owns a deal is state, not a concept to generalize. Retraining a model every time a meeting moves is the wrong tool for the job.

Keeping the notes true

So the question becomes: how do you keep the notes honest? That's what we've been building at Team0.

• They build themselves from the work. Not from what you remember to tell an agent, but from the email, meetings, documents and agent conversations where the work already happens.
• They are compressed, not stored. Team0 doesn't keep piles of text to search later. It keeps a model of your world: people, companies, commitments, decisions, and who said what.
• The newer version wins. When a plan changes, the new one replaces the old one, and the old one stays in history with what changed it.
• Heard is not the same as confirmed. Something only mentioned in passing can be searched, but it isn't stated as fact.
• Every note shows its source. You can open anything, see where it came from, and fix it once.
• Every agent reads the same notes. Tell one agent something, and the next one already knows.

That is compression too. It just happens outside the model, where you can see it, correct it and take it with you.

Where this stops

This doesn't teach a model a new skill. If continual learning delivers, models will get better at learning how to do things. Team0 will still hold what is true in your world, so whichever model you use next starts from it.

Leonard didn't need a better memory first. He needed notes he could trust. So do our agents.