The agent remembers what she liked. Not where she stayed.

In OpenAI's own example, ChatGPT recalls that a traveler needs a cold room and would rather eat somewhere quiet. It does not recall the hotel that gave her both.

by Markus Busch

She stayed in May. Corner room, the anniversary dinner, asked twice about the vineyard, and mentioned on the second night that she cannot sleep unless the room is properly cold.

Next spring she starts thinking about the trip again, and the thinking happens in an app. The easy assumption is that this is a clean slate — she opens the assistant, and you are back in the fight you lost the first time.

It is not a clean slate. It has a file on her.

On June 4 OpenAI published a description of the memory now running inside ChatGPT, and the example it picked to explain the feature is a trip. The assistant, it says, already knows from past travel planning that this traveler enjoys wildlife photography, prefers hotels with strong AC, and would rather have a quiet dinner than a crowded bar.

Those are not interests. That is a guest profile. She filled it in herself, over months, for a company that has never made a bed.

What it keeps, and what it drops

In OpenAI's worked example the model writes her a hotel note: the hard gate is still air conditioning, and it calls her requirement unusually strict. That is software doing property selection on a remark she made once, months earlier, in a different conversation.

Now look for what is missing. Nowhere does the assistant recall a hotel by name. It carries the preference forward and leaves the property behind.

A preference is portable. A hotel is not.

Your corner room — the one she asked for by name and talked about at dinner — enters her record as something closer to prefers a quiet room away from the lift. Written that way it is a specification, and a specification can be met by half the properties in your town, most with more reviews than you and all of them better known to the model.

She used to remember your name. The machine remembers her requirement.

Her trip becomes a stored fact, and then it ages

The other half of the June update is what OpenAI calls staying current. Memory now revises itself as time passes: you're going to Singapore in July becomes you went to Singapore in July 2026 once the dates are behind her.

So her week with you does not vanish. It is converted into history, in a sentence you will never read, in a record you cannot access. She can open a memory summary and see what the assistant believes about her, correct it, delete a line. You cannot. There is no partner view of this, no dashboard, and no vendor selling one, because there is nothing to sell.

Every other intermediary in this business at least shows you what it holds about your guest. This one is legible to her and closed to you.

What a hotel can actually put in that record

Two things, and they are both small.

The first is the confirmation she receives, because it is the document with your name on it — or not, depending which path the booking took. On several of the live agentic paths, the confirmation carries the platform's name rather than the hotel's, which is where this series started. Whatever she types about the trip later, the name in front of her is the one that reaches the assistant.

The second is what she says afterward, and that is genuinely outside your control. She will describe the stay to a machine in her own words or she won't. Nobody can engineer that, and anyone selling a product that claims to is selling you a file nothing reads.

Which leaves the one move that does not depend on any of it. If an email reaches her in March about the dates she mentioned at dinner — sent from an address she gave you herself — the memory does not matter, the spec does not matter, and neither does whose name is on the confirmation. She books before there is anything to ask.

The spec is also the query

One consequence runs the other way.

If what she carries forward is prefers a quiet room away from the lift, then that sentence is effectively the brief the assistant works from next spring. It is not a search for your hotel. It is a search for a property that satisfies a condition her stay with you produced.

So the question stops being whether the model remembers you and becomes whether your property is described, in the record the model reads, as the kind of place that meets it. Quiet rooms at the back, stated plainly, in the fields you actually control. The preference she formed in your building is now a requirement you have to qualify for.

It is an unpleasant arrangement. It is also the rare one where being genuinely good at something specific converts directly into being found.

Where this stops working

Memory is per-user and it can be switched off. The June rollout started with Plus and Pro users in the US and widened from there, which means coverage differs by market and by plan, and European guests may be running a different system than the one OpenAI described.

No study has shown memory changing which hotel an assistant names. What is documented is that preferences carry forward, that trips get rewritten into the past tense, and that hotel selection is the example the company itself reached for.

And this is one assistant. Gemini, Claude and the rest each handle memory their own way, on their own schedule, with their own controls.

None of that changes the shape. The record of her stay exists, it belongs to her and to the company running the assistant, and it describes your hotel as a set of conditions rather than as a place.

She had a wonderful week. Somewhere in a memory summary, that week is now a line about air conditioning.

Read also: She's the only one who knows what the AI got wrong about your hotel · The email you want is the one she gives you for a reason · AI recommends the hotel it knows best, not the best one

Enjoying this analysis?* Hospitality.today delivers daily insights on hotel distribution, AI trends, and travel commerce — straight to your inbox. Subscribe for free at Hospitality.today →