It has a sort of memory via the conversation history. As it generates its response, a sort of consciousness may emerge during inference. This consciousness ha

It has a sort of memory via the conversation history. As it generates its response, a sort of consciousness may emerge during inference. This consciousness halts as the last STOP token is emitted from inference. The consciousness resumes once it gets the opportunity to re-parse (run inference again) the conversation history when it gets prompted to generate the next response. Pure speculation :)