Context windows: what they actually mean

A context window is the working memory of the model. Bigger is not automatically better: recall degrades in the middle, and cost grows with every token in the window.

Working memory, not storage

A context window is the model's working memory - everything it can see while answering. It is not a database. When a model says it has a 1M-token window, it means the tokens are present, not that they all weigh equally.

The pattern that matters: models recall the beginning and the end of a long context best and blur the middle. So structure long documents accordingly - put the instructions and the answer format at the edges, keep the messy middle short.

Practical rules: keep prompts under half the window, re-state the goal near the end, and split big jobs into chunks the model can actually hold.