DeepMind plans device-keyed cloud memory for assistants

Google DeepMind said it will add a persistent, server-side memory layer to Private AI Compute, the company's confidential-computing platform for cloud AI processing.

Under the described design, a user's stored context sits in dedicated encrypted storage in the cloud, while the cryptographic keys needed to unlock it are held exclusively on the user's personal devices — a configuration DeepMind says makes the data inaccessible to anyone else, "even Google." When a model needs that context, an authenticated, end-to-end encrypted channel links the device to an isolated cloud environment. Inside that "secure enclave," the data is temporarily decrypted in isolated memory to serve the request, any new context is saved, and the contents are immediately re-encrypted.

The stated goal is to resolve a standing trade-off: assistants that remember across sessions and devices have generally required cloud storage under provider control, while on-device processing offers stronger privacy but no cross-device continuity. DeepMind frames the memory layer as a "secure digital vault" that extends on-device privacy properties to cloud-scale storage.

Several things the post does not provide: a launch date, which products or Gemini surfaces will use the memory layer, measured security claims, details of the enclave hardware or remote attestation procedure, or any third-party audit. The announcement is architectural and forward-looking — phrased throughout as what the system "will" do — and describes no change to model capability. It builds on Private AI Compute, which Google announced previously.