Advancing Private AI Compute with secure, server-side memory

Google DeepMind introduced a persistent memory layer for its Private AI Compute architecture, addressing the challenge of providing long-term AI assistance across devices without compromising privacy. The system uses a secure digital vault in the cloud, where data is encrypted and accessible only via cryptographic keys stored exclusively on user devices. This approach ensures that even Google cannot access the stored information, maintaining strict on-device privacy standards. The update aims to bridge the gap between the need for advanced AI capabilities and the requirement for robust data protection.

The new capability resolves a longstanding issue in AI development: enabling continuous, context-aware assistance while adhering to privacy constraints typically limited to on-device processing. When an AI model requires data to assist a user, an authenticated, end-to-end encrypted channel connects the device to a protected cloud environment called a 'secure enclave.' This isolated space temporarily decrypts the data, processes the request, and immediately re-encrypts it, ensuring privacy is preserved throughout the interaction.

Historically, local on-device processing has been the preferred method for maintaining privacy, but frontier AI models often demand computing power beyond what a single device can provide. Google’s Private AI Compute platform previously allowed complex tasks to be processed in hardware-isolated cloud enclaves, though these solutions were 'stateless,' wiping all context after each task. The new persistent memory layer addresses this limitation by securely retaining context over time and across devices, enabling richer, more continuous AI experiences.

To build user trust, Google DeepMind has published a tamper-proof public record of its server software and shared results from an independent cybersecurity audit. Devices running Private AI Compute can verify the authenticity of the software before sending personal data, reinforcing transparency and accountability. The research was co-developed by teams across Google DeepMind, Platforms & Devices, Core, and Cloud, with executive sponsorship from Four Flynn, Jay Yagnik, and David Kleidermacher.