I guess it is like every time you ask/talk, for up to how long ago you want Chat GPT to refer to the history: every human has a limit, and for machine AI, what is the context (other than your latest question as input).
the longer you request, the higher the overhead, but probably the highest accuracy - but could be some drawback - as human sometimes want to forget the past and move forward.
Anyway, like the "context" + "new input" to your new programming command for the algorithm to recompute