What happens when ChatGPT's context window is full

The context window is how much text the model holds while it composes a reply. It covers the whole conversation rather than one message: everything you have typed, everything it has said back, your standing instructions, and whatever was read out of any file you attached. Every turn, that running total is handed to the model again from the start.

When the total stops fitting, something has to go before the model sees it. That trimming happens outside the model, which explains almost everything else about how this failure behaves.

What gets dropped, and in what order

OpenAI does not publish the trimming rule, so treat the following as observed behaviour rather than documented policy. What people see consistently is this.

In one line: your recent turns and your standing instructions survive, and the middle of the conversation is what you lose. That is the opposite of what most people assume, because the middle is usually where the reasoning was done.

Why the model never mentions it

By the time the conversation reaches the model, the missing part is already missing, and it does not arrive with a note saying content was removed. From the model's position, what it received is the conversation that happened.

So it does not hedge, apologise or flag a gap. It answers from what arrived, in the same confident register it used when it still had everything.

Degradation starts before anything is dropped

Recall across a long context is not even. Long-context models generally retrieve best from the beginning and the end of what they hold, and worst from the middle. This is a widely reported characteristic of the technology, not something OpenAI states about ChatGPT specifically.

So a long chat gets vague about its own middle well before the window is actually full. The boundary you care about is not a line where things vanish. It is a slope you are already on.

How it shows up

It asks you for something you already gave it. The clearest signal there is. The information did not become unclear, it left.

A rule you set stops holding. You said no bullet points forty messages ago and bullet points came back. The instruction did not weaken over time. It fell out of view.

Early specifics come back slightly wrong. A filename, a figure, a person's name. What you are seeing is reconstruction from later mentions rather than a read of the original.

It can no longer quote something it quoted happily an hour ago. The single most reliable tell, and the basis of the test below.

The test

Pick something distinctive from early in the chat that has not been repeated since, and ask for it word for word.

Quote, exactly, the first sentence of the text I pasted at the start of this conversation.

A paraphrase proves nothing, because it can be reassembled from later references to the same material. A verbatim quote either survives or it does not.

Files sit outside this

Attached files are not loaded into the conversation whole. ChatGPT reads them with a tool and pulls in the parts it needs, so a long attachment does not consume the window the way the same text pasted into a message would.

Two things follow. Uploading beats pasting for anything long, and an attached document can still be consulted in a chat that has forgotten what you said about it.

The file's own caps are a separate matter and they apply first. Text and document files are limited to 2 million tokens, and past that the file is truncated at upload with no warning, before any of this comes into play.

What to do when a chat is full

Start a new one. Usually the right answer, and the one people resist. Open with a six line brief of the conclusions so far and you get a clean window carrying only what matters.

Restate your constraints periodically. A short message repeating the rules moves them back to the recent end of the conversation, where they are safe.

Move stable material somewhere it is re-supplied. Custom instructions or a Project. Those are provided fresh each turn rather than scrolling out of reach. A custom GPT holding reference files, up to 10 files per GPT, does the same job, but only inside a Business, Enterprise or Edu workspace, since GPT creation has left personal accounts.

Keep one task per chat. Long chats that wander across three subjects burn the window on material none of the three needs.

The decision

Treat a long conversation as something with a working life rather than a permanent record. Repeated questions, dropped rules and shaky recall of early detail are the chat telling you it has run out of room, in the only way it can.

Run the verbatim quote test before you act on anything important from a long thread. It is the difference between an answer based on the discussion you had and one based on the last third of it.