Very interesting. How would memory (or previous chat context awareness) work in the case of hot swapping, when multiple users to hot swapping models like threads.