Google DeepMind has detailed a new capability for its Private AI Compute platform: a persistent, server-side memory layer designed to give AI assistants long-term continuity across devices while maintaining privacy standards previously associated with on-device processing.
Under the architecture, information used to assist a user is stored in dedicated, encrypted storage in the cloud, while the cryptographic keys needed to unlock it remain exclusively on the user's personal devices. According to the company, this means the data is inaccessible to anyone else, including Google itself.
When a model needs to access stored information, an authenticated, end-to-end encrypted channel connects the device to an isolated cloud environment — a "secure enclave." There, data is temporarily decrypted in isolated memory to handle the request, new context is saved, and the information is immediately re-encrypted. The design combines hardware-enforced secure enclaves, double-encrypted channels, and per-user databases shielded by device-derived encryption keys.
The update addresses a limitation of earlier confidential computing approaches, which were strictly "stateless" and wiped all context once a task ended. Simple workarounds, such as saving lists of facts and preferences, were not enough for the continuous experiences users expect, the company said. Frontier models also often require more compute than a single device can provide, making cloud processing necessary.
Alongside an updated technical whitepaper, Google DeepMind said it is publishing a tamper-proof public record of its server software, allowing devices running Private AI Compute to verify the software is authentic and unaltered before sending personal data. The company also cited the results of an independent audit by a leading cybersecurity firm and invited the privacy community to review its technical brief, system architecture, security proofs, and verification protocols.
The work was co-developed by Google DeepMind along with Platforms & Devices, Core, and Cloud teams, with executive sponsorship acknowledged from Four Flynn, Jay Yagnik, and David Kleidermacher.