Google DeepMind describes encrypted server-side memory for Private AI Compute
Google DeepMind has described a new encrypted server-side memory architecture for the Private AI Compute platform. Data is stored in encrypted storage in the cloud, but the decryption keys remain exclusively on the user's devices, so not even Google has access to the content.
Google DeepMind has described a technical update to the Private AI Compute architecture intended to bring persistent AI assistant memory that works across a user's devices to this platform. According to the company, the solution combines the assistant's long-term continuity with a level of privacy protection that had so far been reserved only for processing performed directly on the device.
The new memory layer is designed to function as an encrypted vault in the cloud: information needed for assistance is stored in dedicated encrypted storage, while the cryptographic keys to unlock it remain exclusively on the user's personal devices. According to Google DeepMind, as a result no one else has access to the data, including Google itself.
The described mechanism works as follows: when the AI model needs access to stored information, an authenticated, end-to-end encrypted channel is established between the user's device and an isolated protected environment in the cloud, referred to as a \"secure enclave\". This environment temporarily decrypts the data in isolated memory, processes the request, stores any new context, and immediately re-encrypts the data.
The source article only addresses the technical description of the architecture; it does not state a specific implementation date, available products, or further details. See the source article for details.
Why it matters
The described approach addresses the dilemma between an AI assistant's long-term memory across devices and the privacy protection that until now was offered only by processing performed directly on the device – according to Google DeepMind, the stored data remains unreadable even to Google itself, since the encryption keys are held only by the user's device. For users and companies considering AI tools with long-term memory, this represents a concrete technical mechanism that can be assessed when deciding whether to trust the storage of sensitive data, although information on availability and deployment in specific products is still missing.
Two audiences, two different impacts
What this means
For individuals
If this technology is deployed in products that users use, it should mean that the AI assistant will remember context across devices, without Google itself having access to the content of this memory.
For a business
Companies considering deploying Google AI assistants with long-term memory are given a technical description of the architecture, which they can factor into their assessment of data protection risks and regulatory compliance – so far without information on specific availability for enterprise customers.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.