Main idea

A ChatGPT conversation held over several days stays entirely displayed: you scroll through the messages, you find what was said three days earlier. The interface therefore gives the impression of a complete memory.

Three things blur together under that impression, and they do not coincide:

• the transcript kept and displayed on screen;
• what the model can actually use at the moment it writes its answer;
• its memory, which keeps certain elements judged useful and is not exhaustive storage of everything that was said.

A paragraph visible on screen can therefore be out of the model's reach, with no signal saying so.

Why it matters

The consequence is practical: you cannot rely on a conversation to retrieve, five days later, three precise paragraphs chosen today. What you believe is archived is in fact displayed.

That is what justifies persistence decided by hand, outside the tool. Not out of distrust, but because what you see and what the tool has are two different sets.

Nuances and limits

The gap depends on the tool and on the length of the exchange; it isn't a universal property of assistants.

And the transcript keeps its value: it remains readable by a human. The defect isn't that it preserves badly, it is that we credit it with a function it doesn't have.

Open questions

• How do you know, at the moment of asking a question, what the model still has at hand?