I've been thinking about LLM memory since GPT3 came out. Back then, my LLM side project was story generation (i.e. fiction). Context windows for LLMs were tiny back then: 4K tokens, input + output. So just a few pages of text. How to write a novella if your entire knowledge of the text is only a few pages, as if you were an amnesiac author? Suppose you're writing a scene. Where does the scene take place? Your description must match any previous description of the same place and be consistent wit...