How a conversation actually works

Turns, tokens, context windows, and what gets forgotten.

A chat is not a mind. It is a window of tokens that grows until it is full, then older turns fall out.

Each message is broken into tokens — short chunks of text. The model reads the tokens it is allowed to see and predicts the next ones. That window includes hidden system instructions, your messages, attachments, tool results, and prior replies.

What it can see

Whatever still fits in the context window: the brief, the files you attached, recent turns, and any tool output from this session.

What it cannot see

Your other chats (unless a memory feature is on), files you did not attach, and anything already pushed out of the window.

Lab

Watch a context window fill

A teaching sketch, not a real tokenizer. As the chat grows, older turns leave so the new reply can fit.

Prior turns 8Attached files 1

System brief
Files
Prior turns
Room for a reply

About 84% of the window is in use. There is still room.

This is why a long thread gets sloppy, why “remember when I said…” fails, and why projects or a fresh chat with a tight brief often beat a month-old conversation. Memory features, when a product offers them, are a separate store. They are not the window.

Check

A week-old chat starts ignoring the original instructions. What is the most likely cause?

Choose an answer to see why.

Previous Next