Google DeepMind has detailed an update to its Private AI Compute platform that adds persistent, server-side memory for AI assistants. The design stores personal context in an encrypted cloud vault, but the cryptographic keys needed to unlock it never leave the user's devices. When a model needs information, an authenticated, end-to-end encrypted channel connects the device to an isolated cloud enclave, where data is briefly decrypted for processing, updated, and immediately re-encrypted.
The work addresses a known limitation: earlier Private AI Compute was stateless, wiping all context after each task. Workarounds such as saved lists of personal facts were not enough for continuous, cross-device experiences like resuming a conversation on a laptop that began on a phone. The new memory layer is meant to enable that continuity without lowering the privacy bar associated with on-device processing.
To support trust, the company is publishing a tamper-proof public record of its server software, so devices can verify that the code is authentic before sending data. The post also mentions an independent audit by a cybersecurity firm and invites community review of the technical brief, security proofs, and verification protocols.
No independent sources are included, so the claims reflect only Google DeepMind's account of the architecture and its safeguards.