The context window is the model's working memory: how much it can hold in its head at one time, measured in tokens. Everything counts against it, your instructions, the conversation so far, the documents you pasted in, and the answer being written. Feed the model more than the window holds and the oldest content falls out the top. Nothing warns you. The model simply carries on without it.
This explains one of the most common and confusing AI failures: the model that seemed to understand and then drifted. A long conversation quietly pushes the early instructions out of memory, and the model keeps answering, now without the rules you set at the start. It did not disobey. It forgot, structurally, the way a whiteboard runs out of space.
For an operator, the context window sets the shape of what a tool can honestly do. A document that fits in the window can be handled whole. A document that does not fit needs to be broken into chunks, or better, paired with retrieval so the model is handed only the parts that matter for the question. Windows have grown, but the discipline stays the same: working memory is finite, and workflows that respect that are the ones that hold up.
A concrete example. A team wants a model to answer questions about a long policy manual. Pasting the whole manual into every conversation either fails outright or crowds out the space needed for the actual question. The build that works fetches the relevant sections first and hands over just those. Same manual, same model, and the difference is simply respecting the size of the room the model thinks in.
“That document won't fit in the context window, so we'll need to break it into chunks.”
// this page is the full text. the packaged PDF, all fifteen terms, comes with the free library.