Google DeepMind has announced an extension to Private AI Compute, its confidential computing infrastructure designed to support personal AI features across Google devices. The update introduces encrypted, secure server-side memory that allows AI assistants to retain context across sessions without exposing that information in clear form or granting access to third parties, including Google's own internal teams.
Private AI Compute is built on a confidential computing architecture centered on Google's TPUs, combined with isolated execution environments and hardware-based attestation mechanisms. The stated goal is to replicate the privacy guarantees of on-device processing while still leveraging cloud-scale compute for models too large to run directly on a smartphone. Until now, this infrastructure handled requests in a largely ephemeral manner, without persistent data retention between sessions.
The addition of server-side memory changes that model by introducing a form of contextual continuity needed for more personalized assistant features, such as tracking preferences or information mentioned in earlier conversations. Google states that this memory remains end-to-end encrypted and isolated within the same secure enclaves used for the rest of the processing pipeline, with encryption keys tied to the user's device rather than to centralized server-side identifiers.
The announcement reflects a broader industry shift as consumer AI assistants — ChatGPT, Gemini, Siri — progressively add persistent memory capabilities. The underlying technical challenge is reconciling this deeper personalization with verifiable privacy guarantees, an issue several players, from Apple's Private Cloud Compute to now Google's expanded architecture, are attempting to address through confidential computing rather than contractual assurances alone.