I think they have some model for "stuff that sounds like context to be remembered" and then they are appending to the prompt