What Is a Context Window? (And Why Your AI Forgets)
Updated July 23, 2026 · Data as of 2026-07-09
A context window is the total amount of text a model can consider in a single request, measured in tokens. It covers your instructions, the files you attach, the conversation so far, and the response the model is about to write.
Every model has a fixed ceiling. When a conversation grows past it, something has to go. Some tools drop the oldest messages, some summarize them into a shorter note, and some silently trim attached files. All three feel the same from the outside: the assistant that understood your project ten minutes ago now contradicts itself, re-suggests a change you already rejected, or invents a function that you deleted together. That is not the model getting worse. That is the earlier part of the conversation falling out of the window.
Context is shared, not reserved. If you attach a large file, that file competes for the same space as your instructions and the model's answer. This is why a long paste can make a model appear to ignore the question at the top of the message. It is also why a coding tool that quietly indexes your whole repository can outperform one that dumps every file into the request: the first one spends the window on the parts that matter.
Size is advertised in tokens, and the numbers move fast. What matters more than the headline figure is how the tool spends it. A harness with good retrieval, one that pulls in only the relevant files, gets more out of a small window than a naive tool gets out of a large one. Attention quality also drops toward the middle of very long inputs, so a model that technically accepts a whole book may still answer best when you give it the ten pages that matter.
Why it matters when you choose a tool
Context window is the single specification most closely tied to the complaint people actually have, which is that the assistant loses the thread. If your work is long documents, large codebases, or multi-hour sessions, favor tools with both a large window and visible context management, meaning you can see what is loaded and remove what is not needed. If your work is short and self-contained, window size is close to irrelevant and you should weigh price and speed instead. Larger windows also cost more per request on usage-based plans, because you are paying for every token you send.
Frequently asked questions
Does a bigger context window make the model smarter?
No. It makes the model better informed, which often looks like intelligence but is not the same thing. A larger window lets you supply more relevant material in one request. It does not improve reasoning, and filling a large window with irrelevant text usually makes answers worse, not better.
Why does my assistant forget things mid conversation?
The conversation exceeded the window and the tool discarded or summarized the earliest turns. Starting a fresh session and re-stating the essentials is usually faster than fighting it. Some tools let you pin key instructions so they survive trimming.
Does the context window include the model's answer?
Yes, in almost every case. Input and output share the same budget. If you fill the window with input, the model has less room to write, and some tools will refuse the request or truncate the response.
How do I know how much of the window I am using?
Better tools show a token counter or a context meter in the interface. If yours does not, treat roughly 750 words per 1,000 tokens as a working estimate and watch for the moment answers start drifting.
Back to the glossary · Take the quiz · How we score tools