What an AI assistant actually remembers about your customers

A model reads the present conversation; whatever it knows beyond that was a deliberate act of storage, and you may ask who decided it.

Customers ask a fair question about any assistant they talk to: does this thing remember me? The marketing around AI makes the answer sound mystical, as if the model holds every conversation forever inside its weights. The truth is more mundane and, handled honestly, more reassuring.

Remembering is two different layers, and confusing them is where trust gets lost. The first layer is the conversation itself: the text the model is reading right now. It exists for the length of the exchange and it ends when the conversation ends. Nothing from a chat you closed yesterday is silently riding along in a chat you open today.

The second layer is the one that matters: records. Anything an assistant knows about you beyond the current conversation is there because an operator deliberately chose to store it, tied to your account, with a purpose and a retention rule. That is a policy decision made by a company, not an automatic behavior of the technology, and it deserves to be judged as one.

The difference matters because continuity and surveillance wear the same clothes from the inside. An assistant that recalls your last order, your preference, your open question, saves you from repeating yourself, and that continuity is useful. The same assistant accumulating your history silently, with no stated purpose and no end date, is a betrayal dressed as a feature.

So ask the operator, not the chat box. What do you keep about me, and where. For how long. Who on your side can read it. And how do I make you forget. A company that cannot answer those four plainly does not have an AI policy; it has an improvisation.

Our own line is simple. Continuity with the customer is the reason we keep records at all, so the retention is deliberate and bounded: stored under a purpose, kept for a stated time, readable by the people who serve you, and removable when you ask. And the assistant never implies it knows something it was not given. We would rather confess ignorance in a conversation than counterfeit memory in a file.