LLMs Need Better Executive Function

Very sharp analysis. I feel the "context rot" issue the most in my day to day. Wondering if integrating better and more persistent memory is the way?

I love how you contrast the "3D printer" vs. "block of marble" - this feels intuitively correct to me.

I wonder if one way to think about the necessary evolution here looking to a more traditional separation of concerns. Memory is split (and managed entirely separately) from business logic (the intelligence of the LLM)?