How to decide what your AI agent should remember about a user

An AI agent should remember two kinds of things about a user: the ways they're different from your average user, and facts that would change what the agent recommends next time. Everything else is noise. That's the rule we used to build the traveler profile at Pangea, a consumer social travel platform I advise on AI, and it's the reason the AI there doesn't do the thing most assistants do with memory.

You've probably seen that thing. "Since tacos are your favorite food, here are five taco spots for your trip to Bangkok." The assistant found a relevant-looking memory and over-indexed on it. Simon Willison calls this the memory dossier problem: the model keeps a pile of everything you've ever said, then pulls from the pile at the wrong moments. Tidier storage doesn't fix it. Tidier hoarding is still hoarding.

Rule 1: store the deviations, not the averages. An algorithm compares each traveler's behavior and interests against the average traveler and keeps the differences. Nearly everyone lists "food" as an interest, so it tells us almost nothing and we don't lean on it. But if you're one of the few who lists "natural wine," or your trips show you leaving every six weeks like clockwork while most people take two trips a year, that's a real signal, and it shapes how the AI talks to you.

Rule 2: before saving anything from chat, ask one question. When someone tells Pangea's AI something, we check: does this fact change what we'd recommend in the future? If yes, it goes into the profile. If no, it doesn't. A general assistant's memory can resurface anything you've said on any topic. This way, we know which memories will actually inform an answer.

Rule 3: one profile, every surface. The same profile drives the trip-prep emails and the in-app chat. Tell the chat something important (a preference, a goal, your pace) and the next email reflects it. Users don't have to repeat themselves, and the product doesn't contradict itself.

This applies well beyond travel. If you're building a GTM agent (an account research agent, an outbound writer, a persona builder), the same filter works: keep what makes this buyer different from your typical buyer, and keep facts that would change the next message. "Uses Salesforce" is food. "Just lost their only RevOps hire" is natural wine.