From stateless private compute to encrypted persistent memory
Google DeepMind and partner teams described a persistent server-side memory architecture for Private AI Compute on September 23. Per-user memories are stored in dedicated encrypted databases, while the keys needed to unlock them remain on personal devices. An authenticated end-to-end encrypted channel sends data into a hardware-isolated secure enclave, where it is briefly decrypted for processing and encrypted again.
Private AI Compute previously operated without retaining context after a request. The proposed layer targets continuity across sessions and devices while combining device-derived keys and per-user databases so Google says it cannot access plaintext memories.
Architecture claims are not the full privacy outcome
This is a technical architecture announcement, not proof that every product uses it by default or that an independent security audit is complete. Product policies still need to explain key recovery, lost devices, correction and deletion of memories, device revocation and handling of legal demands. Secure enclaves can also have implementation and supply-chain vulnerabilities.
At launch, people should be able to opt into memory, inspect, edit and delete individual records, reset everything and control retention. Independent verification and continuing transparency reports remain important.