Glonce

Google DeepMind adds encrypted server-side memory to Private AI Compute

securityThe Glonce editorsGoogle DeepMind🔥74Glonce rating74Impact70Novelty70Relevance95Surprise45Players90Credibility88
Google DeepMind adds encrypted server-side memory to Private AI Compute

Google DeepMind has published a technical update describing a new capability for its Private AI Compute platform: a persistent, encrypted server-side memory layer that lets an AI assistant carry context across a user's devices while keeping the privacy guarantees normally associated with on-device processing. Previously, Private AI Compute's cloud enclaves, like comparable systems elsewhere in the industry, were strictly stateless, wiping all context the moment a task finished; workarounds such as having the AI keep a list of saved facts and preferences were not enough to support continuous, rich assistance. The new memory layer works like a secure digital vault in the cloud. Information needed to assist the user is sealed inside dedicated, encrypted storage, and the cryptographic keys needed to unlock that storage are held exclusively on the user's own devices, so the data is described as inaccessible to anyone else, including Google. When an AI model needs to draw on stored context, an authenticated, end-to-end encrypted channel connects the device to an isolated cloud environment called a secure enclave. That enclave temporarily decrypts the data in isolated memory to handle the request, saves any new context, and immediately re-encrypts it. Google frames the shift as necessary because frontier AI models often require far more computing power than a single device can supply, so cloud-scale processing has to be reconciled with device-level privacy. To support trust in the system, Google is publishing an updated technical whitepaper alongside a tamper-proof public record of its server software, which will let devices running Private AI Compute verify that the software is authentic and unaltered before sending any personal data. Google also says it is sharing an update on its technical methods, including results from an independent audit performed by an unnamed cybersecurity firm, and is inviting the broader privacy community to review the architecture, security proofs and verification protocols. The work is credited as co-developed by Google DeepMind together with the Platforms & Devices, Core and Cloud teams, with Four Flynn, Jay Yagnik and David Kleidermacher acknowledged as executive sponsors. The post does not name the specific products or devices that will carry the feature, does not give a rollout timeline, and does not name the auditing firm.

Key facts

  • Private AI Compute is getting a persistent, encrypted server-side memory layer so an AI assistant can retain context across a user's devices.
  • Cryptographic keys to unlock the encrypted memory vault stay only on the user's personal devices, making the data inaccessible to anyone else, including Google.
  • Previously, Private AI Compute's cloud enclaves were strictly stateless and wiped all context once a task ended, unlike the new persistent design.
  • Google is publishing an updated technical whitepaper and a tamper-proof public record of its server software so devices can verify software authenticity before sending personal data.
  • The update references results from an independent audit by an unnamed cybersecurity firm; the work is credited to Google DeepMind, Platforms & Devices, Core and Cloud teams, with Four Flynn, Jay Yagnik and David Kleidermacher as executive sponsors.

Why it matters

The change targets a real limitation: cloud AI processing has historically had to choose between the privacy of on-device computation and the extra power of the cloud, and previously Private AI Compute's cloud enclaves were stateless, discarding context the moment a task finished. Adding a persistent, encrypted memory layer aims to let an assistant remember things across sessions and devices while keeping the same privacy bar as local processing, which Google frames as necessary because frontier models need more compute than a single device can supply.

Who it affects

The update concerns anyone who uses AI assistants built on Google's Private AI Compute platform, since it changes how their data is retained and protected in the cloud. It is also relevant to the privacy and security research community, which Google is explicitly inviting to review the architecture, security proofs, and verification protocols.

How to use it

There is no consumer-facing action here: the source describes a platform-level architecture change rather than a product a user turns on. Google says it is publishing an updated technical whitepaper and a tamper-proof public record of its server software so that devices running Private AI Compute can verify the software is authentic and unaltered before sending personal data, and it invites outside review of the system.

How solid is it

The account comes directly from Google DeepMind's own blog post announcing the update, alongside a technical whitepaper and system architecture documentation it says it is publishing. It also cites results from an independent audit by a cybersecurity firm, though the firm is not named. No numeric details, such as latency, storage limits, or timing, are given anywhere in the source.

Risks and caveats

The post does not name which specific products, apps, or devices will carry this server-side memory feature, nor does it give a timeline for when it becomes available to users. The name of the cybersecurity firm behind the cited independent audit is also not disclosed, so the audit's scope and rigor cannot be checked from this source alone.

“frontier AI models often require far more computing power than any one device can provide”

— Google Private AI Compute team