What is a context window?

In plain English

Every language model has a limit on how much it can consider in one go, measured in tokens (chunks of text roughly the size of a short word). That limit is the context window, and it covers everything in play: the hidden system instructions, documents you have pasted, the whole back-and-forth of the conversation, and the answer being written.

A useful image is a desk of fixed size. You can spread out as many papers as fit; once it is full, something must come off before anything new goes on.

Why it matters

The context window decides which tasks are realistic. A large window means you can drop in a contract, a transcript or a codebase and ask questions across the whole thing. A small one means you must summarise or feed material in chunks. It also explains why long chats start to drift: earlier turns get trimmed away to make room.

An example

You spend an hour with an assistant refining a marketing plan. Near the end it forgets the brand tone you specified at the start — not because it is careless, but because those early messages have scrolled out of the window. Restating the key constraints brings them back into view.

What to watch out for

A big context window is not the same as memory. Unless the product deliberately stores notes between sessions, starting a new chat wipes the slate. Accuracy can also sag in the middle of very long inputs, where models tend to attend more to the beginning and end. And filling the window costs tokens, so pasting an entire handbook when three relevant pages would do is slower and pricier.