I like that this separates memory and state. That distinction is critical.
State is about where the workflow is right now.
Memory is about what should influence future reasoning.
Those need different rules.
For example, "email_sent=true" is workflow state or runtime evidence. A user preference is memory. An approved policy is organizational knowledge. A model-generated assumption is only a candidate, not truth.
If Agents treat all of these the same way, reliability suffers.
I think production Agents need memory systems that understand authority, provenance, lifecycle, and scope — not just persistence.