Your agent shouldn't have a memory. It should have a model.
The difference between recording what someone said and actually knowing them.
Think about someone you know well. Now try to remember the last conversation you had with them — word for word.
You can’t. Nobody can. That’s not how it works. You don’t keep a transcript of every conversation. What you keep is a picture — a living sense of who they are: what they care about, what they’re working on, what they’d never do. Each conversation quietly updates the picture, and then the exact words fade. When you need them, you don’t replay a recording. You consult the picture.
Most AI memory does the opposite.
The standard way to give an AI a memory is to save everything it’s told, and when a question comes up, search that pile for the passages that look most similar and hand them back. It’s a transcript with a search box. It works about as well as a friend who, instead of knowing you, frantically flips through recordings of everything you’ve ever said, looking for the relevant bit. Sometimes they find it. Often they surface something outdated, or miss the point, or read back a thing you told them months ago is no longer true.
The problem isn’t the search. It’s the shape of what’s being searched. A pile of old sentences has no idea that a fact from March was overturned in June. It doesn’t know which person “he” refers to. It can’t tell that two things it was told are about the same project. All of that meaning lives *between* the sentences — and a transcript threw it away the moment it stored them.
A model keeps the meaning. Instead of sentences, you store the things themselves — the people, the projects, the commitments — and what you know about each, how sure you are, and when you learned it. New information doesn’t pile up next to the old; it updates the picture, the way it does in your own head. Ask a question and you’re not searching text, you’re reading a structure that already knows how its pieces fit together.
This is the shift everything else is built on: stop recording what a person said, and start keeping a picture of who they are.
Get it wrong, and no amount of clever search saves you. Get it right, and the AI stops sounding like it’s reading from notes — and starts sounding like it actually knows you.