How much text an AI can read and keep in view at once, counted in tokens.
An AI doesn't store your chat in a memory the way you do. Every time it answers, it rereads everything: your messages, any pasted text and its own replies. The most it can take in at once is its context window, measured in tokens.
When the window fills up, older parts get pushed out or cut off, depending on the app. That's why an AI can seem to forget the start of a long chat. Windows have grown fast: about 100,000 tokens in 2023 and around a million in 2024.
Bigger isn't automatically better. One study found models more often miss details buried in the middle of a long input.
Think of it like this
Picture the whiteboard in class. When it's full, the teacher erases the oldest notes to make room. Anything erased is gone for a student who only looks at the board right now.
For example
After a very long chat, "What was my name again?" might fail if that message has slid out of the window.
Common mix-ups
Mix-upAI stores the conversation in a memory and looks it up later.
ActuallyUnless a separate memory feature is switched on, the AI in a new chat window knows nothing of your earlier conversation. And the longer a chat gets, the more there is to reread for every reply, so answers can slow down or usage limits fill up sooner.
Mix-upA big window means it reads every part of a long text equally well.
ActuallyA 2023 study called "Lost in the Middle" found a U-shape: models found the needed information best when it was at the very start or end, and did worse when it was in the middle. When you give a long document, don't bury the key material in the middle, and put your question at the very end.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- When an AI seems to forget what was said earlier in a long chat, what could you try?
- Activity
- Act out the blackboard comparison. Say the board only fits five lines; as classmates add one sentence each, erase the oldest line whenever the board is full. After the tenth sentence, ask "What was the first one?" Then check the 2023 (E0224) and 2024 (E0328) records to see how much windows grew.
Records about this
Source-checked moments from AI history.
Context length jumped by an order of magnitude, sparking the RAG-is-dead debate
The first large-scale counterevidence to the assumption that you can simply put in a long context
The starting gun for the long-context race; made feeding in whole documents without RAG a real option