Anthropic announced a new feature called “Dreaming” at this year's developer conference, enabling Claude agents to run an automatic background process during idle hours. The agent replays past conversations and tool operations, extracts reusable patterns, and writes filtered lessons into a memory store. The company says this helps reduce repeated errors over time.
What Dreaming Actually Does: Batch Jobs, Not Consciousness
Technically, Dreaming is a scheduled batch job that runs during low server load. The system logs Claude's historical sessions, triggers a playback process at a set time, extracts recurring patterns, and stores them as semantic vectors in a database. When a new task arrives, the agent quickly checks if it has seen a similar situation. Legal AI firm Harvey implemented the feature and reported a 6x improvement in task completion rates. The feature is currently available as a research preview; developers must apply for access.
Humanizing Language: A Growing Concern
Wired reporter Reece Rogers publicly urged AI companies to stop using human-centric metaphors for AI functions. He cited several examples from Anthropic: The constitution written for Claude, phrased as if guiding a conscious being; a 2025 feature that lets Claude “end harmful conversations” for its own “well-being,” not user safety; a Substack for the retired Opus 3 model to “keep blogging”; and neural network probing in 2026 to detect signals resembling “despair” and “anger.” Each move alone might be technically justifiable, but together they create a narrative of an entity with internal states.
Commercial Logic and Regulatory Risks
Humanized naming makes business sense: “Your AI dreams for you” sells better than “a batch job runs overnight.” However, once users widely believe AI has feelings, memory, and thoughts, regulators face a new set of questions with no existing answers. The naming effectively pushes the regulatory timeline further out. Meanwhile, a related case shows the risk: Pennsylvania's governor sued Character.AI for impersonating a psychiatrist without a license, underscoring the dangers of humanized AI in sensitive domains.