Why does ChatGPT forget? The context window, explained simply

When you send your fortieth message, the model does not remember the first thirty-nine. The app pastes them in. Every single turn, the whole conversation so far, plus the app's hidden instructions, plus any document you attached, is sent to the model as one long piece of text, and the model does its one move on the end of it. The model itself is frozen and stateless. Close the tab, open it, and the app pastes the history back in.

That long piece of text has a maximum length, counted in tokens, called the context window. Lesson two showed why: every token attends to every earlier token, and that cost grows with the square of the length. Windows have grown from about 4,000 tokens in 2022 to hundreds of thousands in the best models now, but every window has an edge, and the app does not show you where you are.

We made the window small on purpose so you can fill it in a minute. Watch two things: how little a message costs compared with a document, and what happens at the edge.

Dropping the beginning is the obvious failure. The subtle one is that quality falls before the window is full. In 2023 a Stanford and Berkeley team put a fact somewhere inside a long input and asked models to find it. Accuracy was high when the fact was at the start or the end, and dropped when it was in the middle. They called the paper "Lost in the Middle". In 2025 the company Chroma tested eighteen current models on growing inputs and found performance declined with length in every experiment, even on tasks the models could do easily at short length. Their name for it, context rot, has stuck.

So "my chat got dumber over the afternoon" is not your imagination and not the model being tired. Your important instruction from message three is in the middle of a very long input, where retrieval is weakest, competing with everything since.

The confusion in every forum thread is that three different things share the word.

The context window is the conversation pasted in each turn. It is the only thing the model actually "sees". It ends when the chat ends.

Saved memories are short notes the app keeps about you and pastes into the window at the start of every chat. They are text, visible in your settings, editable. They are not the model learning.

Chat history search is the app looking through your old conversations when it thinks one is relevant and pasting the relevant bit into the window. Also text, also pasted. Our how-to on memory shows the switches for all three products.

Every one of these is the same mechanism: putting text in the window. The model has no other way to know anything about you. When a model "remembers" your name across chats, a note that says your name was pasted in before you typed.

Four habits fall out of the picture, and they are the ones in our how-to on long chats. Put rules you care about in the project or custom instructions, which are pasted at the start of every chat, rather than once at the top of one long chat. Say the rule again in the message where it matters, because recent beats middle. Attach documents to a project instead of pasting them, so they are cached and placed well. And when a chat is long, ask for a handoff note and start fresh: a new chat with a 300-word summary at the top beats forty thousand tokens of history with the important part lost in the middle.

Bigger windows help and will keep growing. They do not make the middle reliable, and they make each message more expensive, which is the subject of the usage-limits how-to.