Pretty interesting take this. I wonder if there is a minimal state management we could evolve which would be sufficient for LLMs to use while still making it possible for a human to reason about the abstraction. It won't be as bloated as the existing ones we came up with organically however.