Google’s Private AI Compute platform will introduce a server-side, encrypted persistent memory layer that preserves on-device privacy by keeping decryption keys exclusively on users’ personal devices. The design aims to enable long-term continuity of AI assistants across multiple devices while preventing access to stored data by third parties, including Google, according to the announcement.
What is changing and how it works
The update adds a persistent memory layer in the cloud that functions like a secure, encrypted vault. Key points described by the company:
- Stored information resides in dedicated, encrypted cloud storage.
- Cryptographic keys required to decrypt that data are held only on the user’s personal devices, so the company says no one else — not even Google — can access the contents.
- When an AI model needs the data, the device connects to the protected cloud environment via an authenticated, end-to-end encrypted channel.
- A protected, isolated execution environment (a “secure enclave”) temporarily decrypts the data in isolated memory to process the request, stores any updated context, and immediately re-encrypts it.
This approach attempts to combine the privacy benefits of on-device processing with the computational scale of cloud-hosted, frontier AI models.
Why this was needed
Large, advanced AI models often require more compute than a single personal device can provide. Earlier implementations of Private AI Compute and similar industry solutions operated in hardware-isolated cloud enclaves but were strictly stateless: they erased context at the end of each task. Such stateless operation cannot sustain the richer, continuous experiences users expect from personal AI assistants.
The new persistent memory layer is intended to let cloud-scale AI securely retain context over time and across devices while maintaining strong privacy protections.
Building trust and transparency
To bolster user trust, the announcement highlights several transparency measures:
- A tamper-proof public record of the server software will be published so that devices can verify the software is authentic and unmodified before sending personal data.
- The team has published technical updates on methods and included the results of an independent audit by a leading cybersecurity firm.
These resources are presented so that the wider privacy and security community can review and verify Private AI Compute’s protections.
Development credits
Google DeepMind, Platforms & Devices, and the Core and Cloud teams co-developed this research. The announcement also acknowledges executive sponsorship from Four Flynn, Jay Yagnik, and David Kleidermacher.
Conclusion
Adding private, persistent memory to Private AI Compute is intended to enable seamless, multi-device assistance while keeping decryption keys on users’ devices and minimizing third-party access to stored context. The published technical brief, architecture details, security proofs, and verification protocols are provided for external review.
