because it's not easy to identify exactly when to r/w memory accordingly, especially when you'd need to have an LLM decide when and if to do that

and to scale it in a way where you don't need a whole custom model loaded for 1 user (financially unviable)

just my immediate thoughts, could be wrong though.