Stateful LLM agents accumulate long-term memory from web pages and tool outputs. The software that orchestrates the model—the agent harness—does not record who is allowed to act on that memory. This creates a vulnerability: an instruction hidden in a fetched page can potentially influence the agent's behavior without proper authorization.
The paper introduces ClawMMU, a protection-coherent virtual memory system for such agents. The name and framing borrow from operating-system memory management: just as an OS tracks which processes can access which pages, ClawMMU aims to track the provenance and permissions of memory objects in an agent's context.
Because the source is a single preprint with a truncated abstract, the article cannot verify implementation details or experimental results. The contribution, as described, is a conceptual and architectural step toward making agent memory security a first-class concern, rather than an afterthought.