Context windows and limits

What actually fills a model's context window, how to budget it, and why fitting inside it doesn't guarantee the model uses it well. Defines the context window and everything that shares it; works through the token budget across system instructions, history, attachments, retrieval, and tool calls; distinguishes the two ways providers measure maximum length; covers what happens on overflow and the four separate memory stores people conflate; goes deeper into why information can still be missed even when it fits — the attention-competition mechanism and the lost-in-the-middle result; refreshes the compute and cost consequences of context length; makes the case against sending everything; maps conversation-management strategies at a conceptual level; and ends with how to test effective context use and what should never enter the window at all.