I present Semantic Decontextualization, a privacy architecture for autonomous AI agent systems that eliminates the reward of data theft rather than raising its cost. Instead of protecting sensitive data through increasingly complex encryption, this architecture makes stored data meaningless without the user's physical presence: sensitive information is replaced with semantically opaque tokens on the server, while a re-contextualization map exists exclusively on the user's biometric-protected device. A complete server breach yields nothing of exploitable value. I formalize the Messy Desk Principle: security is achieved not by making the lock harder to pick, but by ensuring the intruder finds only fragments of a story they cannot interpret without the owner's context. I introduce the Information Window, a transient processing architecture that handles sensitive data exclusively in volatile memory without persistence. I identify five statistical inference attack vectors with proposed mitigations, and discuss the architecture's structural resistance to harvest-now-decrypt-later attacks. The architecture is implemented in the Nova autonomous agent system. Twelve provisional patent applications covering this and related methods were filed with the USPTO between February and April 2026 (primary application: 64/022,549).