The analogy
Imagine a long but not infinite whiteboard, on which you and the AI write in turns. Before answering you, it rereads the whole board from the start: that's how it "holds the thread". But the board has an edge. When you reach the bottom and need space to keep writing, someone has to erase the topmost lines, the oldest ones. Those lines disappear: the AI no longer rereads them, because they're no longer there.
That's why in a long chat it seems forgetful precisely about the beginning. It didn't get distracted: the initial part was erased from the board to make room for the rest. Everything it "remembers" is only what's still written at this moment.
How it really works
At each turn the AI receives the entire text of the conversation as a block and rereads it in full: there's no warehouse where it keeps what you told it two minutes ago, there's only the text that passes before it every time. This text is measured in units called tokens (pieces of words), and the context window is the maximum number of tokens that fit.
As long as the conversation stays within the window, the AI sees everything and holds the thread. When it exceeds it, the oldest content is cut or compressed into a summary, and the precise details of the beginning are lost. It's not a malfunction: it's the edge of the board making itself felt.
What you can do in practice
- If a piece of information is important and the chat has gotten long, restate it: don't take for granted that the beginning is still "on the board".
- Put the instructions or constraints that matter toward the end of your message: they're the freshest lines, the least at risk of being erased.
- For a long job, every so often ask the AI for a summary of the key points, then open it in a new chat and paste it: you start fresh without losing the gist.
- Open a different chat for each different topic: keeping everything in a single endless conversation fills the board with stuff that doesn't belong.
A common misconception
People believe that the AI, by talking with you within the chat, is studying you and getting better and better at understanding you. It's not learning: it's rereading. Nothing from that conversation is stored as stable knowledge; it's all there on the temporary board. Close the chat and, aside from the long-term memory features that activate separately, nothing of that exchange remains. The next chat starts from a clean board.
Frequently asked questions
How long is the context window?
It depends on the model and changes often: some handle very long conversations, others less. The point that matters isn't the exact number, it's that a limit always exists. Even the most capacious window, in a chat that never ends, sooner or later fills up.
If I close the chat, does the AI remember next time?
By default no: each chat starts clean. The exception is long-term memory, a separate feature that some platforms offer and that keeps notes about you from one conversation to the next. It's a different thing from the board we're talking about here.
Why does it forget mid-chat a detail I'd given it earlier?
Two possibilities: the window filled up and that detail dropped out, or it's still written but buried under a lot of other text and the AI gives it less weight. In both cases the remedy is the same: repeat it near the question.
Does a longer chat make the AI more "on the ball" with me?
Beyond a certain point it's the opposite. The more you fill the board, the more the important details risk being erased or drowned in the noise, and the answers get worse. For a good result it's better to have a focused conversation and, when needed, to start fresh with a summary, not to pile on endlessly.