MemGPT treats the context window as scarce working memory and gives the model function-call tools to page information to and from a larger external store. The mechanism makes long-running conversations and documents possible by exposing retrieval and write-back as addressable operations. Coordination Systems can borrow that memory boundary for durable, reusable state, while Open Epistemic Intelligence must still solve ownership and propagation; tiered memory alone supplies neither.
The context window is treated as physical memory, with a larger external store behind it. The analogy is virtual memory in an operating system.
The model itself issues function calls to move information between the two tiers, deciding what to page in and what to write out. Nothing outside the model decides what is worth remembering.
That lets a conversation or a document run past the context limit. The pattern generalizes past the specific system: give the model an addressable store and the tools to manage it.