Google’s Private AI Compute is the arrangement that lets a phone hand a task to a data centre without the data centre being able to read it. Until now it was strictly stateless: whatever context a request carried was discarded the moment the task ended. On 23 September the Private AI Compute team said that will change, and described a persistent memory layer that keeps context across sessions and devices.
The design, as Google describes it: information sits in dedicated encrypted cloud storage, the keys to unlock it are held only on the user’s own devices, and an end-to-end encrypted channel connects the device to an isolated secure enclave that decrypts the data, does the work, writes back new context and re-encrypts it. Alongside the description Google published an updated technical whitepaper, a tamper-proof public record of its server software, and the results of what it calls an independent audit by a leading cybersecurity firm.
Two things are worth holding onto. The audit firm is not named, so the audit cannot be looked up. And the whole post is in the: no product, no device, no date. This is an architecture commitment, not a feature you can use.