Google outlines encrypted cloud memory for assistants across devices

Private AI Compute has so far discarded context when a task ends. Google now plans to add encrypted, per-user storage so an assistant can carry information from one request or device to another.

In its technical update, the Google Private AI Compute Team describes a system in which personal devices hold the keys to that storage. Google claims the design would keep the information inaccessible to anyone else, including Google itself.

Memory inside an isolated environment

When an assistant needs stored context, an authenticated, end-to-end encrypted connection would link the device to a secure cloud enclave. The enclave would temporarily decrypt the information in isolated memory, process the request, then encrypt any updated context for storage.

Google gives examples such as retrieving assembly instructions previously viewed through smart glasses or continuing a conversation between mobile and web. These are intended uses of the architecture, not an announcement that those experiences are available.

Alongside an updated technical paper, Google is publishing what it calls a tamper-proof public record of server software. Devices would check that software before sending personal data. The company also says it is sharing independent audit results, opening its privacy claims to outside scrutiny.