ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents
On this page
arXiv:2610.28093v2 Announce Type: replace-cross Abstract: AI agents keep long-term memory, much of which comes from web pages and tool results. Agent harnesses, the software that runs the model, do not record who may act on this memory, so an instruction hidden in a fetched page can be treated as the user’s own instruction. We propose ClawMMU, a protection layer that labels every piece of memory with its owner, scope, origins, and authority. ClawMMU checks these labels whenever content enters the model’s context, is written to storage, returns from storage, or drives a tool call. Its central rule is that information may rise in authority, but authority never lifts itself. Because the checks depend on where content came from, not on what it says, memory attacks become classical protection faults that the harness stops even when the model is fooled.
Why It Matters
ClawMMU treats an agent’s long-term memory like a virtual memory system with protection coherence: it tracks who may act on memory derived from web pages and tool results, so that instructions hidden in fetched content cannot be treated as authoritative commands. This is a direct answer to a class of prompt-injection and memory-poisoning risks in stateful agents, and it shows how operating-system-style isolation ideas are migrating into agent architectures. See Agents, Tools, and Agent Frameworks, Agents, Autonomy, and Tool Use, AI Safety, and Algorithms Underlying AI.
Sources
Back to Recent News · Sub-article index.