Anthropic gave agents the ability to dream. Then developers woke up.

At Anthropic's AI DevCon in London this summer, Lamis Mukta discussed the latest advancements in memory management for AI agents. Mukta highlighted that current models cannot automatically understand how to succeed in an organization or perform specific tasks without human guidance. To address this, Anthropic developed a "dreaming" process that runs asynchronously in the background to update and maintain memory systems for agents.

Traditional memory management techniques, such as using a CLAUDE.md file containing instructions and preferences, can become cumbersome and difficult to manage over time, especially as files grow lengthy. Mukta proposed a new approach where agents autonomously manage their own memory systems, deciding when to read, write, and update memories. This in-band process occurs within a single conversation session and utilizes progressive disclosure to limit the amount of information presented to the model at any given time.

However, Mukta noted that this approach still relies on human oversight and opinionated decision-making about what memories are necessary. Additionally, memories can become outdated or irrelevant, and incorrect or malicious memories can introduce significant risks. To mitigate these issues, Anthropic introduced "dreaming," an out-of-band asynchronous process that consolidates memory, removes irrelevance, and helps agents learn over time.

Dreaming is designed to address the limitations of in-band memory management, where agents must balance completing tasks and updating memory systems simultaneously. By consolidating memory and cutting out irrelevant information, dreaming allows agents to focus on learning and improving their performance over time. While Mukta believes that dreaming has the potential to significantly enhance Anthropic's Managed Agents memory and API approach, staff software engineer Jayakumar Ramalingam cautions that dreaming may create a dangerous promotion path, where bad memories can persist and influence future sessions.

Enterprise AI architect Vladimir Beskorovainyi agreed with Anthropic's focus on dreaming, but argued that the process addresses the wrong underlying problem. He believes that memory should be treated more like versioned infrastructure rather than artificial cognition, emphasizing the importance of provenance, evidence, and expiration conditions for each memory. Ramalingam and Beskorovainyi agree that for memory to be considered production-ready, it should be auditable, traceable, and easily reversible.