What is a context window?
How much text a model can hold in mind at once — prompt plus reply — measured in tokens; run past it and the earliest text falls away.
A context window is the amount of text an AI model can hold in mind at any one moment: your prompt, the conversation so far, and its own reply, all counted together. It’s measured in tokens, small chunks of text rather than whole words, and it works like a limited desk rather than a filing cabinet — once the desk is full, something has to give way for anything new to land on it.
What happens when you exceed it
When a conversation grows past the window’s limit, the earliest material falls out of view first. The model doesn’t flag this or apologise for it; it simply stops carrying that text forward, which is why a long chat can feel like it’s “forgotten” how it started. Ask it to recall an instruction from many messages back and it may answer confidently from what remains in view, with no sign that anything is missing. Researchers studying these systems, including at Stanford’s Institute for Human-Centered AI, are careful to describe this as a mechanical limit on what text is present, rather than forgetting in any humanlike sense.
Why it matters
The context window sets a hard ceiling on how much document, code or conversation history you can hand a model in a single exchange. A bigger window lets you paste in a longer report or keep a longer back-and-forth before older parts drop away. What it does not do is give a model memory between separate conversations: unless a tool is deliberately built to store and re-load information, closing a chat and opening a new one clears the window entirely, regardless of its size. NIST’s guidance on AI system evaluation treats context length as one input among several that shape reliability, not a stand-in for genuine memory. Tools such as ChatGPT and others, including Claude, Gemini and Copilot, all work within some form of this limit, though they manage it in different ways.