Anthropic has quietly solved one of the most frustrating problems in AI agent development—and named it in a way that might undo all the goodwill. At its Code with Claude conference this week, the company unveiled "dreaming," a feature that lets Claude Managed Agents pause their work, review what just happened, and decide what to remember. To an engineer, it's a clever memory management system. To a critic, it's another step toward treating AI like it's alive.
The practical case is airtight. Large language models struggle with extended tasks because context windows—the amount of information they can hold at once—are finite. Current solutions involve "compaction," where the model periodically strips out irrelevant details from its working memory. Anthropic's dreaming takes a different approach: instead of just compressing what's already there, the system steps back, reviews entire sessions, and curates specific memories for permanent storage. When the agent wakes up and resumes work, it loads those curated memories back into context, essentially rebuilding relevance from a curated library rather than a compressed transcript.
For developers building multi-agent workflows that run for hours or days, this is a meaningful shift. Without dreaming, agents lose track of earlier decisions, repeat work, or hallucinate context that was supposed to be preserved. With it, Anthropic claims the system can maintain coherent memory across projects that would otherwise exceed practical context limits. The feature is currently in research preview and limited to Managed Agents on the Claude Platform—not available through the standard Messages API.
But the branding is audacious. Anthropic could have called this "offline consolidation," "memory curation," or "session review." Instead, it chose a word with millennia of metaphysical baggage. "Dreaming" implies consciousness, subconscious processing, something happening behind the scenes that the system itself doesn't fully understand. For a company that positions itself as a safety leader, this framing is a liability. Critics will point to it as evidence that Anthropic is anthropomorphizing its products, undermining the technical precision the company usually projects.
The tension is genuine: Anthropic's engineers clearly understand what's happening—scheduled memory consolidation is a straightforward engineering pattern. But the marketing chose evocative language over precise language, presumably because "dreaming" is more memorable than "batch memory compaction." Whether that tradeoff pays off depends on whether the feature works well enough that the name becomes irrelevant. If dreaming delivers reliability improvements, the controversy fades. If it produces strange artifacts or unexpected memory decisions, the name becomes ammunition.
For now, developers can access the feature through Managed Agents and experiment during the research preview period. The real test comes when Anthropic decides whether to ship this more broadly—and what to call it when it does.