Google DeepMind Gives Its Cloud Assistant Encrypted Memory

Google DeepMind’s Private AI Compute now includes server-side persistent storage with encryption. Previously, cloud computation was stateless — no state was kept between tasks, and users had to start over from scratch each time they used different devices. With this new feature, user data is stored in secure hardware enclaves in the cloud. When it’s accessed for computation, it’s decrypted into isolated memory and immediately re-encrypted. All keys are kept locally on the user’s device. You can read more about the technical details in their updated docs, or see an external security audit of the system.

Maintaining a user’s private context across devices is challenging. In my work, I often need agents to keep track of information over multiple sessions. Here’s how Google does it:

Only the user has access to the keys. Data is only accessible to the server for a short period of time (while it’s being processed within a secure enclave). This isn’t as sexy as releasing a new model, but it’s crucial if we want to build agents that have true long-term memory, rather than just showing off a single demo.