ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents

Agent memory is shifting from raw context toward virtual memory, where each page holds one fact and its typed, durable, derived representations. Paging gives structure, provenance, identity, and addressing but no protection, so untrusted pages act as policy. Protection is the missing abstraction. ClawMMU keeps owner, scope, and authority coherent through context admission, durable derivation, and effect dispatch, making memory attacks classical protection faults rather than content judgments. The design rests on one rule: information may rise in authority, but authority never lifts itself. A forged instruction admitted as evidence stays evidence and cannot act.