source: Google DeepMind: Advancing Private AI Compute with secure, server-side memory
level: technical
google deepmind announced a technical update to its private ai compute platform that adds secure, server-side memory. this new layer lets an ai assistant remember context across devices and sessions while keeping data private. the system stores information in encrypted cloud storage, but the decryption keys stay only on the user's own devices. even google cannot access the data. this resolves a long-standing tradeoff between cloud-scale ai power and the strict privacy of on-device processing.
the architecture uses hardware-enforced secure enclaves, end-to-end encrypted channels, and per-user databases protected by device-derived keys. when an ai model needs data, an authenticated channel connects the user's device to an isolated cloud enclave. the enclave temporarily decrypts the data in memory, processes the request, saves new context, and immediately re-encrypts it. previously, private ai compute was stateless and wiped all context after each task. the new design supports continuous assistance without exposing personal information to the cloud provider.
to build trust, google published a tamper-proof public record of its server software and an updated technical whitepaper. devices can verify the software is authentic before sending personal data. an independent cybersecurity firm audited the methods. this work was co-developed by google deepmind, platforms and devices, core, and cloud teams. the update enables use cases like resuming a conversation between a phone and a laptop or pulling up instructions seen earlier through smart glasses, all while keeping the underlying data private.
why it matters: this lets ai assistants remember user context across devices without exposing personal data to the cloud provider, a key step for trustworthy personal ai.
source: Google DeepMind: Advancing Private AI Compute with secure, server-side memory