Google has detailed a persistent memory layer for Private AI Compute that is designed to carry an assistant’s context between devices without giving Google access to the stored information. The architecture keeps each user’s memory in dedicated encrypted storage while decryption keys remain exclusively on that person’s devices.
When an assistant needs context, the device opens an authenticated, end-to-end encrypted connection to a hardware-isolated cloud enclave. The enclave temporarily decrypts the data in protected memory, processes the request, saves updated context and encrypts it again. That changes Private AI Compute from a stateless service, which erased context after each task, into one that can resume work over time.
The intended uses include continuing on a laptop from instructions previously viewed through smart glasses or carrying a long conversation between mobile and web. Cloud processing provides capacity for frontier models that cannot run fully on a personal device, while device-held keys are meant to preserve the privacy properties of local processing.
Google is publishing an updated technical brief, verification protocols and a tamper-resistant public record of server software. Devices will check that the cloud code is authentic before sending data. The design and independent audit provide material for review, but the announcement describes an architecture rather than broad product availability or real-world proof against every implementation failure.