Concept guide · How AI works
A context window is working memory, not a memory.
Everything a model can read while it answers has to fit in one space, measured in units called tokens. It is not a filing cabinet the model keeps. It is the desk it works on for one job, cleared for the next one unless you put things back on it.
What it actually is
Anthropic's documentation defines the context window as "all the text a language model can reference when generating a response, including the response itself," and says this "represents a working memory for the model." A token is roughly a short chunk of text, a little under a word on average.
More space is not automatically better. The same page warns that "as token count grows, accuracy and recall degrade, a phenomenon known as context rot." So what you put on the desk matters as much as how big the desk is.
A normal example
You have a sixty-page staff handbook and you want to know the notice period for parental leave. There are three honest ways to ask, and they are not the same job for the model.
The proof: one question, three ways
Put the entire handbook in the message. The full text sits on the desk, so the exact clause is present. This works only while the document fits, and it costs the most.
Keep the handbook in a Project (a saved workspace in tools like ChatGPT or Claude) or an uploaded file, and let the tool fetch only the few relevant paragraphs before answering. The desk stays clear, and the answer still rests on the real clause.
Run all three and you get three answers you can lay side by side against the real handbook. That comparison is the whole lesson: same model, same question, different things on the desk.
Try it now
- Pick a long document you know well and one specific question with a definite answer.
- Ask it three ways: paste the whole document, ask it over the stored file, and summarise then ask.
- Line up the three answers against the actual clause in the source.
The decision this gives you
Paste when the whole thing is small and any word might matter. Retrieve when the source is large and only a slice is relevant. Summarise when you want the gist and can accept losing the specifics. When a number, a date or an exact clause is the point, do not trust a summary.
Assuming a long chat "remembers" everything you said at the start. Once the window fills, chat interfaces such as claude.ai "can also manage the context window on a rolling first in, first out basis," meaning the oldest turns drop off. The model has not changed its mind, it can no longer see the beginning.
The human bit
You decide what goes on the desk and what a good answer has to match. Naming the exact clause, number or source to check against is your job, not the model's.
Where this stops being exact
Providers differ. Some now summarise earlier turns automatically so a conversation can continue past the limit. A separate "memory" feature that stores facts across chats is a different thing from the window for one answer. Window sizes vary by model, and a cached prompt still occupies the window: caching "changes what you pay for those tokens, not whether they count."