Google outlines cloud memory for assistants that it says it cannot read

single source· 1 articles · confidence: low · first seen 2026-09-23 00:00 UTC

What this means for you

Nothing to act on yet. There is no date, no named product, no API and no pricing, and the security properties are described only by Google — no outside review has been published. If you build cross-device assistant memory, note the design depends on keys held on the user's device.

Google DeepMind published a technical description of a server-side memory layer for its Private AI Compute platform, which it says will let an AI assistant keep context across devices without Google holding the keys to it. Data is stored encrypted in the cloud, and the decryption keys stay on the user's devices. When a model needs the data, an encrypted channel connects the device to a secure enclave, an isolated memory region in the cloud. The enclave decrypts briefly, handles the request, then re-encrypts. No launch date, product name, pricing or independent verification was given.

Key facts

  • ·The capability is described for Google's Private AI Compute platform, published 23 September 2026. source
  • ·Persistent memory is intended to carry context across a user's devices rather than living only on one device. source
  • ·The information needed to assist a user is sealed in dedicated encrypted cloud storage. source
  • ·Cryptographic keys that unlock that storage are held exclusively on the user's personal devices, according to Google. source
  • ·A secure enclave temporarily decrypts data in isolated memory to handle a request, saves new context, then re-encrypts it. source
  • ·No availability date, product name, pricing or third-party verification of the security properties was included. source

What the sources say

  • Google DeepMind Blog — Architecture post describing the cross-device memory design, with no availability date or outside review

Sources

The original reporting. Follow these — they did the work.

← the wire