Google wants to give AI long-term memory in the cloud that, it says, even Google will not be able to access

A capable personal agent needs memory. Without it, the user has to repeatedly explain preferences, ongoing projects and previous decisions. Persistent memory, however, also creates one of the most sensitive databases a technology company can operate. Google DeepMind is therefore expanding its Private AI Compute architecture with persistent server-side memory designed so that even Google itself cannot read its contents through ordinary means. Information is to be stored in separate encrypted databases, and the keys required to decrypt it are derived from the user’s devices. When the model needs personal context, the device creates an authenticated, end-to-end encrypted channel into a protected cloud environment. The data is temporarily decrypted inside a hardware-isolated area, the model performs the computation and the updated context is encrypted again. Technically, this is an attempt to combine two advantages that have so far been in tension. Local processing gives strong control over privacy, but a phone or laptop lacks the compute needed for the largest models. The cloud provides performance, but traditionally requires the user to trust the operator not to use the data in other ways. Google says the new system replaces part of that trust with technical guarantees. It is publishing technical documentation, results of an independent audit and an immutable public record of server software, allowing a device to verify before sending data that it is communicating with an approved system version. That matters. Encryption alone is not enough if the data is later decrypted by ordinary server software that the operator can change. The secure environment must technically constrain the infrastructure administrator as well. The business significance could be substantial. Long-term memory is one of the conditions for turning a chatbot into a personal agent. A system that remembers projects, documents, habits and previous communication can provide much more useful assistance and at the same time create high switching costs. Portability will therefore be the next problem. Private memory locked into one ecosystem can tie a user to a provider even if it is technically well secured. Regulators are therefore likely to address not only privacy but also the ability to export and transfer context. For companies, the architecture is interesting beyond personal assistants. A similar principle can be used for sensitive corporate knowledge when a company wants the performance of a public cloud but does not want to expose the underlying content to the cloud operator. Long-term memory is therefore becoming a new infrastructure layer of AI. Competition will not be only about which model answers better, but also about whom users will trust with years of their work and personal context. KEY TERMS Secure enclave: A hardware-isolated environment in which sensitive data can be processed separately from the server administrator. End-to-end encryption: Protection of data so that only authorized endpoints can read it. Persistent AI memory: Long-term stored context that a model can use across different sessions and devices.

Google wants to give AI long-term memory in the cloud that, it says, even Google will not be able to access

The full article is available to users registered at Hard Skills