The idea in one paragraph.
Borrow virtual memory from operating systems: give the model a small in-context working set plus a large external store, and let it page between them with its own function calls.
THE MACHINERY
Virtual context management
Main context and external context, with paging between them under model control.
Self-directed paging
The model issues function calls to read and write its own memory hierarchy.
Interrupt-driven control
Yields and interrupts modelled on OS syscalls to manage long-lived sessions.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Proposes virtual context management, drawing on hierarchical memory in operating systems, which provide the appearance of large memory through movement between fast and slow memory.
Inspect a source excerpt
we propose virtual context management, a technique drawing inspiration from hierarchical memory systems in traditional operating systemsRead the passage in context on arXiv ↗
Manages memory tiers to extend context within a fixed window, and uses interrupts to manage control flow between itself and the user.
Analyses documents that far exceed the underlying model's context window.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Paging quality depends entirely on the model's judgement about what to evict.
- Read-only instructions bound some edits, but agent-written memories can still be mistaken or corrupted; the analogy does not establish general memory correctness.
Why it belongs in this atlas.
The hinge of the whole talk. Before MemGPT the context was append-only; here the agent gets create, read, update, delete over its own context. Every later rung is CRUD over something bigger.