The AI's working memory. Everything it can 'see' and 'think about' at once.
Imagine the model is reading a book out loud to answer your question, but it can only hold a certain number of pages in its hands at one time. The context window is how many pages those are. It includes your question, the conversation so far, any documents you pasted in, the AI's instructions, and the answer it's generating. When the window is full, the oldest pages get put down and forgotten. The AI does not remember anything that falls off the edge of the window unless you tell it again.
Bigger window = you can feed it more without it losing track. Claude Sonnet holds 1 million tokens (roughly 750,000 words, about 10 full novels). Older models held 4,000 tokens (about 3,000 words, 10 pages). When an agent starts acting weird after a long session, the usual reason is that the important context has fallen off the window. Fix it by summarizing, chunking, or starting fresh.