Google DeepMind has outlined a new server-side memory layer for Private AI Compute, aiming to let AI assistants retain context across devices without giving the cloud provider access to a user’s stored information. The architecture keeps the cryptographic keys needed to unlock memory on the user’s own devices, while encrypted data sits in dedicated cloud storage and is opened only inside protected secure enclaves when needed.
That design extends the privacy logic behind on-device personalization, similar in spirit to work on private AI that learns user habits locally. The difference is scale: DeepMind wants cloud-class models to remember useful context over time while preserving protections normally associated with local processing.
The company says authenticated, end-to-end encrypted channels connect devices to isolated cloud environments, where data is briefly decrypted for inference and then immediately re-encrypted. Per-user databases and device-derived keys are intended to prevent even Google from reading the stored memory.
The update also speaks to broader questions around consent and control over personal data. DeepMind says devices will verify server software against a tamper-proof public record before sending personal information, and it is publishing updated technical documentation and audit results.
This is an architectural announcement, not evidence that persistent private memory is already broadly deployed in consumer products.