DeepMind's plan for cloud AI memory it cannot read

Google DeepMind laid out how it wants to give AI assistants a long-term memory that lives in the cloud without the company itself being able to read it. The problem is familiar to anyone who uses an assistant across a phone, a laptop, and glasses. You want it to carry your context everywhere, but storing that memory on a server usually means trusting whoever runs the server with it.

The approach, called Private AI Compute, keeps the encryption keys on your own devices and processes your data inside hardware-enforced secure enclaves in the cloud. When the model needs your information, the device opens an end-to-end encrypted connection to an isolated environment. The data is decrypted only inside that enclave, only for as long as the request takes, then re-encrypted. Google says even it cannot see what is stored, because the keys never leave your devices. To make that more than a promise, the team says it will publish tamper-evident records of the exact software running in the enclave along with independent security audit results, so outsiders can inspect the design rather than take the claim on trust. You can read the technical brief and check the architecture yourself. The whole guarantee rests on the enclave holding and the audits being real, which is exactly where the scrutiny should go.

Why it matters

If you have avoided letting an assistant keep a persistent memory because of where that data would sit, this is the design to examine closely, because the privacy depends entirely on the enclave and audit claims holding up, so read the brief before trusting it with anything sensitive.