The Context Window

Every model has a working memory, called the context window. Everything you say, and everything it replies, must fit inside. In this lesson, you will learn what the window holds, what happens when it fills up, and how to work with it smartly.

• A model's working memory
• Everything must fit inside
• What happens when it fills up

Context windows are growing

The window is measured in tokens. Early models could hold around two thousand tokens, about three pages of text. Later models reached thirty two thousand, then a hundred thousand, and today, some can hold a million tokens, or more, which is several long books at once. Bigger windows let you paste whole reports and codebases. But they are slower, and more expensive, and models can still overlook details buried in the middle.

What fills the window

Remember that the window holds everything. The instructions you give at the start. The whole conversation so far. Any documents you paste in. And the reply the model is writing right now. When the total is too long, the oldest parts must be dropped, or summarised. That is why very long chats can seem to forget the beginning. The model does not remember between chats. It only sees what is in the window.

1. System instructionshow the assistant should behave
2. Conversation historyall earlier messages
3. Pasted documentsreports, code, notes
4. The answer being writtentokens still to come

Working smartly with limits

A few habits help. Put the most important instructions first, and repeat key facts near the end. Do not paste more than you need. For long documents, ask for a summary in stages. And when a chat gets very long, start a fresh one, with a short summary of what matters. Think of the window as a desk. A tidy desk makes better work.

Summary