Why ChatGPT forgets what I said earlier in the chat

What’s actually happening

Nothing is stored between your messages. OpenAI’s documentation puts it plainly:

A conversation feels continuous because the app resends the transcript every single time — your first message, the reply, your second message, and so on, all of it, on every turn.

That resent bundle has to fit in one fixed space. Anthropic’s documentation lists what counts towards it: “the system prompt, every message in messages (including tool results, images, and documents), and your tool definitions”. The reply counts too, because it is generated inside the same budget.

And the bundle only grows. The same page calls this progressive token accumulation: as the conversation advances, “each user message and assistant response accumulates within the context window, and previous turns are preserved completely”.

Preserved completely — until they cannot be. When the transcript passes the window, something has to go, and in a chat product the thing that goes is the beginning. Anthropic notes it in a footnote on that same page:

Run the demo above and the arithmetic is unremarkable. Four short messages, about 92 tokens for Claude Haiku 4.5. Two and a half thousand exchanges later the transcript is 230,000 tokens, the window holds 200,000 — and 1,305 messages are on the wrong side of the line. In the raw arithmetic nothing is summarised or half-remembered: the oldest messages are simply not in the request. Whether your app softens that cut is a product choice, and the next paragraphs come back to it.

This is why the failure feels so unlike human forgetting. A person who forgets your budget constraint still knows they had a conversation with you. A model that never received it answers as though the constraint was never mentioned, with complete confidence — because for that request it never was.

Products differ in how gracefully they handle the edge. The blunt version is the cut above. The gentler version has a name — compaction: before old turns fall off, the app condenses them into a short summary and keeps that summary inside the window. A summary is smaller, and it is also lossy — your exact words do not survive it, they get paraphrased. That is why a long chat sometimes half-remembers: the model is no longer reading what you wrote, it is reading an abridgement of it. Other products retrieve pieces back on demand, or keep a separate memory store. What none of them can do is exceed the window.

What to do about it

Restate the constraint next to the work. The rule you typed in message three has an expiry date. If it still matters at message eighty, type it again in message eighty. This costs you one line and removes the entire failure mode.

Start a new chat with a written brief. A fresh conversation that opens with a tight summary of the decisions so far beats a long conversation whose beginning has already fallen off. You choose what survives instead of letting the oldest-first rule choose for you.

Watch for the moment it starts contradicting the setup. Answers that quietly drop an earlier requirement — the budget, the tone, the format, the language — are the visible symptom of the line being crossed. Treat that as a signal to restart rather than to argue.

Do not paste the same long document twice. Every paste is charged against the same space, and two copies push the beginning out twice as fast. Paste once, then refer back to it.

Measure before you paste. Context fit counts a document in tokens and shows what share of each model’s window it takes, so you can see whether the thing you are about to add leaves room for the conversation around it.