The context window

The context window is everything a model can see when it writes a reply. Text that does not fit is invisible to it.

At the start of a chat you tell an AI chatbot, "Answer in short bullet points." An hour and many messages later, after you pasted in a long reading, its replies come back as long paragraphs. You never withdrew the instruction.

Your first instructionSent at the start
Many later messagesYours and the chatbot's
The latest replyWritten from the text it receives
Think first. The chatbot followed the instruction at first and stopped later. What could have changed between the first reply and the latest one?
Students often say the chatbot got tired or chose to ignore the instruction. Steer them toward what the model can still see.

The system prompt, your pasted files, the conversation and the reply all share one window, measured in tokens. When it is full, the oldest text usually goes first and nothing warns you. Raise Files you pasted past half the window, then raise Conversation so far.

system prompt files you pasted conversation, still visible dropped out
Used
—
Dropped
—
More detail

Change the window size

What counts toward the window

The system prompt, every pasted file, every message you send and every reply all count toward the same total.

A model has no memory outside the window. When you open a new chat, the window starts empty, and the model knows nothing about the old chat unless the product copies text from it into the new window. Anything the model should know has to be in the window.

You can now say what a model can see and why a new chat starts blank.

More detail

Long sessions with an agent

An agent session is one long conversation. The files it reads, the commands it runs and your messages all use the window, so a long session drifts when early instructions drop out.