Google DeepMind has shared how it plans to add private, server-side memory to its Private AI Compute platform. The goal is an AI assistant that remembers your context across your phone, laptop and other devices while keeping on-device privacy standards. According to Google DeepMind, the encryption keys that unlock your stored data will live only on your personal devices, so the data is “inaccessible to anyone else, even Google.”
This is a technical update, not a product launch. But it tells us where Google’s assistant is heading.
The Short Version
- What: A new persistent memory layer for Private AI Compute, Google’s cloud setup for handling sensitive AI requests.
- Who: Google DeepMind. It will power Google’s AI assistant experiences.
- How: Your memory data sits in encrypted cloud storage. The keys stay on your devices, and the data only gets decrypted inside an isolated “secure enclave.”
- Why it matters: It goes after the biggest trade-off in personal AI. Until now you had to pick between an assistant that remembers everything and one that keeps your data private.
- When: Google didn’t give a launch date. The post describes how the system “will” work.
The Problem Google Is Trying to Solve
An assistant is only as useful as what it remembers. If it knows your preferences, your ongoing projects and what you asked for last week, it gets a lot more helpful. That kind of memory usually has to live in the cloud, especially if you want it on every device you own.
The cloud is also where privacy worries come from. On-device processing keeps data local, but a phone has limited storage and compute, and it can’t easily share context with your laptop. Google DeepMind calls this a longstanding dilemma in modern AI, and says the new architecture resolves it.
How It Works
Google DeepMind compares the new memory layer to “a secure digital vault in the cloud.” Here’s what happens when the AI needs your context:
- Your device opens an authenticated, end-to-end encrypted channel to a protected, isolated environment in the cloud.
- Inside that secure enclave, your data gets decrypted in isolated memory, but only long enough to handle the request.
- Any new context from the conversation gets saved.
- Everything is encrypted again right away.
A secure enclave is a locked-down part of a server that runs code apart from the rest of the system, so even the people running the servers can’t look inside. Because the keys stay on your hardware, Google says your information stays private “as if it never left your device.”
Why This Matters
What stands out here is who holds the keys. Most cloud services encrypt your data but keep the keys themselves, so in principle the provider can still read it. Google is saying it won’t be able to read what its own assistant remembers about you.
That puts Google in direct competition with Apple, which pitched its Private Cloud Compute on the same promise: cloud-scale AI with privacy guarantees close to on-device. Privacy used to be Apple’s main selling point here. Now it’s a basic requirement for any company that wants its assistant handling your email, calendar and personal history.
It also fits a wider trend. Memory is quickly becoming the feature that sets assistants apart, and every major lab is adding some version of it. The labs that can show memory is both useful and safe will earn the trust needed for AI agents that act on your behalf.
What to Watch
The blog post leaves some open questions:
- Verification: Can outside researchers check the enclave claims? Apple published parts of its system for inspection. Expect pressure on Google to do something similar.
- Key recovery: What happens to your memory if you lose all your devices? Keys that stay on your hardware make recovery harder.
- User control: How easily can you see, edit or delete what the assistant remembers?
- Rollout: Which products get this first, and when.
What It Means for You
For practitioners and businesses, this points to where enterprise AI is going. Persistent context that the provider can’t read makes it easier to bring AI assistants into regulated fields like healthcare, finance and law. For everyday users, the pitch is simple: an assistant that knows you better without Google being able to read what it knows.
The real test comes when this ships and security researchers start poking at it. Full technical details are available in Google DeepMind’s original post.