The model itself doesn't change. The agent writes notes and loads the relevant ones into later conversations.
Good memory resembles the notebook a thoughtful colleague keeps about your shared work: the client's preferences, where each project stands, how you like updates written, the thing you asked them never to do again. It's a selection of what's useful, and it's much shorter than a transcript.
How does AI memory work?
The agent saves notes, usually to files or a database, and reads them back when they're relevant. Anthropic's memory tool documentation describes the pattern: the agent writes to files that persist between sessions and consults them later, instead of trying to hold everything in its context window. Some products also let the agent search through past chats.
Where those notes are kept, and who can read them, varies by product. Our guide to how AI agent memory works covers the practical side.
Why does memory make such a difference?
Without it, every conversation begins with you re-explaining your role, your business and your preferences, and the output is generic because the input was. With it, the gains build up: drafts sound more like you over time, you stop repeating background, and you can come back after a fortnight and pick up a thread where you left it.
Tell your agent once that the finance director at one client wants a five-line written summary and dislikes unplanned calls. Months later, when you ask for an update for that client, it comes back as five lines with no suggestion to "jump on a call". For everyday work, that kind of recall often matters more than a slightly cleverer model, a point we argue in memory over intelligence.
What can go wrong with AI memory?
Stored information can be wrong, stale or something you'd rather not have kept.
In June you mention that the refurbishment of your Harrow office finishes at the end of July. In August it's extended to October and you forget to say so. In September the agent drafts a note to the landlord as if the work were done, because the note it holds is out of date. The fix is a quick correction ("the Harrow work now runs to October") and a skim through what the agent remembers every month or so.
Memory is personal data, so it's worth knowing what happens to your data before you rely on it.
Common questions
Does memory make the AI model smarter?
No. The model is unchanged; memory changes what the model is given to work with, and better input produces better output. That improvement is easy to mistake for the AI getting cleverer.
Can I delete what an agent remembers?
That depends on the product, so check before you rely on it. You should be able to see what's stored, correct it and delete it. If you can't do any of those, you have no real control over that data.
Is AI memory the same as chat history?
Not usually. Chat history is the full record of what was said, while memory is a short set of facts the agent judged worth keeping. Some products can also search old chats, which blurs the line between the two.