Google DeepMind Gives Its Cloud Assistant Encrypted Memory

Google DeepMind has added persistent, server-side encrypted memory to its Private AI Compute platform. Before this, cloud processing was stateless. The context was wiped right after each task, and the assistant lost its history whenever a person switched devices. Now user data sits in protected hardware enclaves in the cloud. It gets decrypted only during processing, inside isolated memory, and is re-encrypted right away. The keys live solely on the person's own device. Google has published updated technical documentation and the results of an independent security audit, so the claims can be checked.

Keeping an assistant's memory across devices without letting the data leak out is hard. I run into exactly this problem when an agent needs to remember context beyond a single session. Google shows a working approach here. Only the user holds the keys. The server sees the data for just a few seconds of processing, inside a protected enclave. There is less hype in this than in a flashy model release. But this is precisely the kind of engineering that decides whether you can trust an agent with real long-term memory, not just a demo of one.