Every request to a language model includes instructions, any supplied documents, the conversation so far, and the model's response. The context window is the maximum combined size the model can handle. Window sizes have grown dramatically, but they are still finite, and larger requests cost more and can be slower.
Systems manage context deliberately by retrieving only relevant passages, summarizing older conversation, and storing durable facts in memory rather than repeating them every time. Good context management improves quality and keeps costs predictable.