The Pulse
Google DeepMind Adds Private Memory to Cloud AI
Google DeepMind describes a server-side memory layer for Private AI Compute that stores encrypted user context while keeping decryption keys on personal devices.
AI.info Team ·
Google DeepMind moves private AI memory into the cloud
Google DeepMind says it has designed a server-side memory layer that lets AI assistants retain personal context across devices without placing the encryption keys in Google's hands. The company described the system on September 23, 2026, as an update to its Private AI Compute platform, which runs AI workloads inside hardware-isolated cloud environments.
The proposal addresses a limitation in the platform's earlier design. Private AI Compute could process sensitive requests in protected cloud infrastructure, but it was “stateless”: once a task ended, the system discarded its working context. The new architecture adds persistent storage for information an assistant may need later, while keeping the keys needed to decrypt that information on a user's devices.
Google DeepMind has not announced a consumer product launch or a release schedule for the memory system. Its post presents the work as a technical update and points readers to an updated Private AI Compute technical brief.
Keys stay on the user's devices
Under the proposed design, each user receives a dedicated encrypted memory store. The data is encrypted before it is saved, while the cryptographic keys required to access it remain under the control of the user's personal devices, according to the Google DeepMind announcement.
When an AI model needs that information, the device establishes an authenticated, end-to-end encrypted connection to a protected cloud environment. A secure enclave temporarily decrypts the relevant material in isolated memory, handles the request, writes any new context and encrypts it again before storage.
Google describes the arrangement as a combination of hardware-enforced secure enclaves, encrypted communication channels and per-user databases protected by device-derived keys. The company's claim is that the stored information remains inaccessible to anyone else, including Google. The architecture depends on the security of the user's devices, the enclave hardware and the software that controls access.
Why cloud memory changes the privacy problem
On-device processing offers a straightforward privacy boundary, but the most capable AI models often require more processing power than a phone, laptop or wearable can provide. A cloud assistant can perform heavier tasks, yet long-term memory creates a larger exposure: personal preferences, past conversations and information drawn from other devices must remain available after an individual request ends.
Google DeepMind says ordinary workarounds, such as asking an assistant to maintain a list of personal facts, do not provide the continuity required for a persistent assistant. Its proposed memory layer is intended to support examples such as opening assembly instructions on a laptop after viewing them through smart glasses, or continuing a conversation between mobile and web interfaces.
The technical distinction matters. The system is not simply a larger prompt history held by a cloud provider. It is designed as a separate encrypted storage layer that can be accessed by protected inference services only when the user's device authorizes the operation.
Software transparency becomes part of the boundary
Google DeepMind is pairing the memory design with additional verification measures. The company says it will publish a tamper-resistant public record of the server software so that devices can check whether the software is authentic and has not been altered before sending personal data.
The post says Google is providing an update on its technical methods, including the results of an independent audit by a leading cybersecurity firm. It also points readers to the system architecture, security proofs and verification protocols.
That disclosure plan gives outside reviewers a way to inspect more than the storage diagram. For a privacy system that promises Google cannot access user memory, the question is not only how the encryption works but also whether the deployed software matches the code and security claims presented to users.
Private AI Compute extends Google's earlier platform
Google DeepMind says it previously introduced Private AI Compute as a cloud processing platform for sensitive AI tasks. The system combined hardware-isolated cloud infrastructure with a way to process complex requests while preserving protections associated with on-device processing.
That earlier platform already used protected cloud enclaves to handle isolated requests. The September 2026 update adds persistence, changing the system from a place that handles isolated requests into one that can retain user-specific context over time.
The announcement leaves several practical questions unanswered, including which Google products will use the memory layer, how users will inspect or delete stored memories, what happens when a user loses access to every authorized device, and how the independent audit evaluates recovery and account-management procedures. For now, Google DeepMind has published the architecture, the software-verification plan and the technical brief; it has not identified a public deployment date.