Google outlines persistent memory for Private AI Compute across devices

Google’s planned shift to retained context

Google’s Private AI Compute Team announced plans for private server-side memory on September 23, 2026. The team says the addition will allow its cloud AI platform to retain information across requests and devices. For people using personal AI assistants, the intended consequence is continuity: information from an earlier interaction could remain available when they move to another device.

Google describes its previous Private AI Compute architecture as stateless, meaning it erased context after each task. The proposed memory layer changes that model by allowing context to persist. The company argues that retaining lists of personal facts or preferences alone does not provide the continuity it wants personal assistants to support.

Google illustrates the intended experience with assembly instructions first viewed through smart glasses and later retrieved on a laptop, or a conversation continued between mobile and web. These are scenarios the company says the architecture is designed to support; they do not establish a rollout to those devices or interfaces. The announcement presents the memory capability as forthcoming.

How Google says memory will be protected

Under Google’s proposed design, retained information sits in encrypted storage, with the keys needed to unlock it held exclusively on personal devices. A device connects to an isolated cloud environment through an authenticated, encrypted channel. Google says this secure enclave temporarily decrypts information to process a request, stores new context and immediately encrypts it. Its privacy claim therefore covers a system that processes decrypted information inside protected cloud memory.

Google also says it is publishing a public record of its server software alongside an updated technical whitepaper. According to the company, devices running Private AI Compute will be able to check that software is authentic and unchanged before transmitting personal information. It also reports releasing updated technical methods and results from an independent cybersecurity audit, while inviting the privacy community to examine the platform’s protections.