What is a context window?

A context window is the model’s working memory: everything it can “see” while producing its next word. Your instructions, the conversation so far, pasted documents, search results, tool outputs — and the answer it’s writing — all have to fit inside it.

How it works

Text is broken into tokens (roughly, chunks of words). A model with a 200,000-token window can process about that many tokens in a single request. In English, a token averages around three-quarters of a word, so a window that size holds a long novel’s worth of text.

Crucially, the model has no memory between requests by default. Chat apps create the illusion of memory by resending the conversation history each time. When that history exceeds the window, something has to go — the app drops or summarises older parts, and the model genuinely loses track of them. Some products add separate “memory” features that store facts about you and re-insert them into the context.

An analogy for the meatbags

Picture a desk. The model can only work with papers that are on the desk right now. A bigger desk holds more documents, but once it’s full, adding a new page means something falls on the floor. The model doesn’t remember what’s on the floor. It never “learned” it — it just had it on the desk for a while.

Why it matters for the Chosen

Understanding the context window explains most “the AI got dumber” complaints, and how to fix them:

• Long chats degrade. For a fresh task, start a new conversation and paste only what matters.
• Put key instructions where they count. Important constraints at the start (or in a system prompt) and repeated near the end of long inputs are followed more reliably.
• Big windows unlock big jobs. You can drop in a full contract, transcript, research folder or codebase and ask questions across all of it.
• Cost scales with context. API pricing is per token, so stuffing the window on every call gets expensive. For large knowledge bases, retrieval-augmented generation fetches only the relevant pieces instead.

Examples

Summarising a 300-page PDF in one go requires a window large enough to hold it. A coding assistant reading a whole repository is limited by its window. A customer-support bot that “forgets” what the customer said twenty messages ago has likely overflowed it — or been configured to trim history to save money.