A context window is the amount of text — measured in tokens, not words or pages — that an AI model can take into account at one time when it generates a response; anything earlier than that window, in a long conversation or a long document, stops directly shaping what the model says next.
Vendors document this limit because it has practical consequences for anyone feeding a system real work. Anthropic’s own developer documentation frames the topic as needing readers to “understand how the context window works, how extended thinking and tool use count toward it, and how to manage context as conversations grow” — a plain acknowledgment that the window is a real constraint to manage, not just a specification on a marketing page.
The practical risk is not abstract for a Canadian firm handling real files. A long conversation that pastes in an entire underwriting file, then keeps adding messages back and forth, will eventually push the earliest pages out of the model’s view — and unless the tool is built to flag that, nothing tells the person using it that the model has quietly stopped accounting for the start of the file it is supposedly reviewing.
Picture a mortgage agent pasting a lengthy file into a chat tool, then asking follow-up questions for twenty minutes. Once the running conversation exceeds the model’s context window, the earliest pages of that file are no longer part of what the model is working from — so a question about an early page, asked late in the conversation, may get an answer based only on the model’s memory of the chat so far, not the actual page. Splitting that same file into shorter, separate conversations, or re-supplying the page when it matters, keeps the relevant material inside the window instead of trusting that the model still remembers it.
See also: what is an embedding, what is an AI model.
Working around a fixed context window — deciding what to keep in view and what to summarize or drop — is a build decision. custom-ai-solutions covers how that gets designed instead of left to chance.
