Why it forgets

The model has no memory at all. Everything it knows about your conversation, it was just told.

This is the single most misunderstood thing about how these systems work, and understanding it fixes a whole category of confusing behaviour.

What is actually happening

A model has no memory. Between one message and the next, it retains nothing.

So how does it follow a conversation? Because the entire conversation is sent again, from the beginning, every single time you type. Your first message, its reply, your correction, its next reply — all of it goes back in, as one long block of text, and the model produces the next piece.

This is called the context window: the maximum amount of text the model can consider at once.

Once you know this, several things stop being mysterious.

Why long conversations get worse

When a conversation grows past the context window, the earliest parts fall out of view. The model is not forgetting in the human sense — the old text is simply no longer being sent.

That is why, forty messages in, it contradicts something you said at the start. As far as it is concerned, you never said it.

Why long conversations get expensive

If the whole history is re-sent every turn, then every turn costs more than the last. A conversation of twenty exchanges does not cost twenty times one exchange — it costs considerably more, because each turn re-processes everything before it.

This is also why tools that manage context carefully exist. It is not tidiness; it is the bill.

Why it repeats itself

Models often restate things you already agreed, or answer a question you already answered. They are not being obtuse. The earlier agreement is somewhere in a very long block of text, and a model's attention is not perfectly even across it.

What actually helps

Start a new conversation for a new task. The cheapest and most effective habit there is. Long-running threads are where quality quietly degrades.

Put what matters at the beginning or the end. Attention is strongest at the edges of the context, weakest in the middle. A long document with one crucial paragraph buried in the centre is a real risk.

Summarise and restart. Ask it to write a short summary of where you have got to, then paste that summary into a fresh conversation. You have just compressed the history and cut the cost.

Give it the material it needs, not everything you have. A relevant two pages beats an irrelevant hundred.

Ask it to restate the task. "Before answering, summarise what I am asking for." This catches a dropped instruction before you get a wrong answer.

The mental model worth keeping

The model is not a colleague who remembers your project. It is a brilliant stranger who has been handed the transcript, reads it fresh each time, and has a limit on how much they can read at once.

Everything it appears to remember, it was told — by you, in this conversation, moments ago.