The model itself remembers nothing
A language model call is stateless. Every time you invoke it, it sees only the text you send in that request and produces a response, then forgets the interaction ever happened. There is no persistent internal state carried between calls, no hidden memory bank inside the weights that updates as it talks to you. What feels like a continuous conversation with an agent is really an illusion built by re-sending the entire relevant history back into the model on every single call.
This matters enormously for agents, because an agent doesn't make one call, it makes dozens or hundreds across a long task: planning steps, tool calls, tool results, intermediate reasoning, error messages, retries. Each of those pieces of text has to live somewhere, and that somewhere is the context window sent with the next call. If it isn't in that context, as far as the model is concerned, it never happened.