quick reflections on llm memory

i wrote a blog post recently suggesting that decoder-only models lack stable inner mental states (beliefs) at least in the way that humans have them. if llms have stable inner states it's in the form …