AI memory systems

Google Research has published a method called the sleep framework, aimed at helping large language models absorb new knowledge during continual learning without forgetting what they already know. The approach draws inspiration from how sleep consolidates memory in biological systems, and it is drawing attention among AI researchers debating how to make model updates less prone to catastrophic forgetting.

Researchers have analysed 179 real-world incidents involving AI memory systems to identify what they call the technology's 'negative space' — the failure modes and gaps that current safety discussions overlook. By systematically cataloguing cases where AI systems retained, leaked or misused stored information, the team aims to give developers a clearer map of risks that often go unexamined, and to inform safer memory design going forward.

Microsoft has laid out an AI PC roadmap spelling out memory and hardware tiers for upcoming AMD and Intel chips. Entry-level machines would carry 8 GB of RAM, Copilot+ ready systems 16 GB or more, builder PCs higher, and supercomputers 128 GB and above. The specification guidance signals how Microsoft intends to standardise AI features across the Windows PC ecosystem, and hardware watchers are debating what the memory thresholds mean for future laptop and desktop designs.

AI agents require a dedicated operating environment—the infrastructure layer that manages their memory, tools, permissions and coordination—much as applications once needed operating systems. The piece argues this layer is now the critical missing piece of enterprise AI, and asks which companies or engineers will own its construction, raising questions about standards, vendor lock-in and where value will accrue in the agent economy.