Stable diffusion most likely don't do it, but even static model that as an input takes embedding of all your historic prompts + current prompt would get progressively give you better inputs as you use it.