Context Windows Explained: Why Long AI Chats Forget Things

If you have ever had an AI assistant forget an instruction you gave twenty minutes earlier, you have run into the limits of its context window. Understanding how that window works makes it much easier to get consistent answers from ChatGPT, Claude, Gemini and similar tools.

What a context window is

A language model does not remember a conversation the way a person does. Each time you send a message, the application sends the model a block of text: system instructions, as much of the earlier conversation as fits, any files you attached, and your new message. The maximum size of that block is the context window. Anything that does not fit is simply not seen when the model writes its answer.

Context windows are measured in tokens rather than words. A token is a chunk of text — often a short word or part of a longer word. As a rough rule for English, one token is about four characters, so a thousand tokens is in the region of 750 words. Code, non-English languages and unusual formatting can use noticeably more tokens for the same amount of text.

Why long conversations degrade

Three separate effects make long chats less reliable:

• Truncation: when the conversation exceeds what the application sends, the oldest messages are dropped or replaced with a summary, so early instructions can disappear entirely.
• Attention dilution: even inside the window, models tend to use information at the beginning and end more reliably than details buried in the middle of a very long input.
• Contradictions: long chats accumulate outdated statements. If you changed your mind halfway through, both versions are still in the history, and the model may follow the wrong one.

Window size is not the whole story

Providers publish the maximum window for each model, and those limits have grown quickly. A bigger window helps, but it does not remove the problems above, and the size available in a consumer chat app can differ from the API limit. Window sizes also change frequently, so check the provider's current documentation rather than relying on numbers quoted in older articles.

Cost matters too. When you use models through an API, every token in the window is billed on every request. A conversation carrying a long history pays for that history again with each new message, which is why trimming context can noticeably reduce API spend.

Practical ways to keep important information in view

• Restate key constraints: when a requirement really matters, repeat it in the message where it applies instead of trusting the model to remember it.
• Start fresh at natural milestones: when a phase of work is done, ask for a summary and begin a new chat with that summary.
• Use project or custom instructions: most assistants let you set standing instructions that are sent with every message, outside the conversation history.
• Remove noise before sharing: paste the relevant function, not the whole repository; the latest draft, not all of them.
• Compress long histories: a compact, structured summary of decisions and state preserves most of the value at a fraction of the token count.

How to tell you are near the limit

Warning signs include the assistant asking for information you already provided, ignoring formatting rules it followed earlier, mixing up names or versions, or giving noticeably shorter and vaguer answers. When you see these, do not keep pushing in the same thread — summarize and restart.

The ContextSwitch extension shows an estimated token count for the visible conversation with a progress bar against a budget you choose for each assistant, so you can see when a chat is getting heavy before quality drops. The estimate uses the same four-characters-per-token approximation described above, so treat it as a guide rather than an exact figure.

Key takeaways

• The context window is everything the model can see at once; anything outside it is invisible.
• Tokens, not words, are the unit — roughly four English characters each.
• Long chats fail through truncation, diluted attention and contradictions.
• Summarizing and restarting at milestones is the most reliable fix.