One user message and the model's reply to it, counted as a single exchange; a chat session is a sequence of turns that accumulate in the context window.
A turn is one round of a conversation with an AI assistant: the user sends a message and the model produces a reply. A chat session is a sequence of these turns. "Multi-turn" describes any conversation that runs longer than a single question and answer.
A single turn can hold more than one step. When the model calls a tool, reads the result, and calls another before it replies, all of that happens inside one turn. The user sees one message in and one message out; the intermediate tool calls and reasoning are part of the same turn.
Every prior turn is resent to the model on each new turn, because a language model holds no memory between requests. The full history — the system prompt, every past user message, and every past reply — is packed into the context window alongside the current message. As a session grows, the number of tokens carried forward grows with it, until the oldest turns are dropped or summarised to stay within the limit.
A brand mention placed in an earlier turn stays in the model's view only while that turn remains in the context window. In a long session, early context is the first to be trimmed, so a mention's position in the turn sequence affects whether the model can still see it when it answers.
