Anthropic's agent "dreaming" sparks developer concerns over memory reliability and costs

At an AI DevCon session in London, Anthropic staff member Lamis Mukta unveiled a memory management innovation called "dreaming"—an asynchronous, batch-mode process designed to clean up, consolidate, and refine the stored knowledge that agents accumulate over time. The concept emerged from earlier approaches like static configuration files and agent-managed memory tools, which struggled with scaling and staleness.

How dreaming works#

Anthropic's dreaming process runs continuously in the background with dedicated compute resources, analyzing agent behavior patterns and proposing memory-store updates. The system identifies recurring failure modes—wrong units, missing context, broken tool configurations, stylistic quirks—and suggests targeted memory improvements for human review.

The architecture builds on progressive disclosure, a tiered retrieval system where agents scan lightweight metadata first before accessing full content if needed, reducing unnecessary context window consumption. As Mukta framed it: "The way I like to think about it is as if I'd had a bookshelf in my room, and every time someone talks to me, I can kind of scan and look at my list of books and see if any of the titles might be relevant to the conversation."

This out-of-band approach sidesteps a core limitation of in-band memory management, where agents must simultaneously execute tasks and update their own knowledge bases.

Industry skepticism: five key concerns#

Developers and AI architects interviewed by The New Stack raised serious reservations:

• Bad memories outlive sessions. Jayakumar Ramalingam, an independent AI systems researcher, warns that while hallucinations normally die within a single session, a flawed memory entry can poison thousands of future sessions. Human review at scale becomes a "rubber stamp for recommendations nobody has time to reconstruct," he argues. He calls for memories to carry provenance, evidence, and expiration dates rather than merely recurring patterns.

• Missing operational fundamentals. Ramalingam insists agent memory requires "versioned infrastructure" discipline—tracking who changed what, why, and how to roll back. Without those controls, "it becomes an unaudited configuration file with an AI attached."

• Dreaming misses the root problem. Vladimir Beskorovainyi, an enterprise AI architect, contends dreaming catches surface-level errors (wrong em dashes, typos) but misses costlier failures: agents reaching for the wrong tool for reasons that defensible. He notes the real issue is that recent batch updates cause the memory store to drift toward "whatever the agent fleet happened to do lately," eroding correctness over time.