Newest papers

ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents

Jinhao Hu, Bardia Mohammadi, Mofasshara Binte Rafique, Lars Klein, Akhil Arora, Ashvin Goel, Laurent Bindschaedler · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28093 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los agentes de IA conservan la memoria a largo plazo, la mayor parte de la cual proviene de páginas web y resultados de herramientas. Los agentes aprovechan, el software que ejecuta el modelo, no registran quién puede actuar en esta memoria, por lo que una instrucción oculta en una página recogida puede ser tratada como la propia instrucción del usuario. Proponemos ClawMMU, una capa de protección que etiqueta cada pieza de memoria con su propietario, alcance, orígenes y autoridad. ClawMMU comprueba estas etiquetas cada vez que el contenido entra en el contexto del modelo, se escribe al almacenamiento, regresa del almacenamiento o conduce una llamada a una herramienta. Su regla central es que la información puede subir en autoridad, pero la autoridad nunca se eleva. Debido a que los controles dependen

Machine translation (Meta M2M100); the English original follows.

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.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Stateful firewall, Software, Layer (electronics), Web page, Programming language

Is this paper real? This page was built from OpenAlex and arXiv records, not generated. Software can verify any reference in one call: GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.28093, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txt

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange