Current LLMs are stateless as far as we know, their state when computing a new token is only the preceding text tokens, they don't store any metadata or save state from the previous calculations.