Anthropic's 'Dreaming' Gives Claude Agents Overnight Memory Consolidation
Anthropic launched 'Dreaming'—a scheduled process that runs between agent sessions, reviews past activity, extracts patterns, and writes new memory entries for the next session to use. Harvey's legal AI saw task completion rates rise roughly 6x in internal testing.
Original sourceAnthropic unveiled a feature called **Dreaming** at its Code with Claude developer conference in San Francisco on May 6, 2026. The feature is a scheduled background process that runs between agent sessions, reviews everything an agent did in its last job, pulls patterns and lessons from those sessions, and writes new memory entries that the following session can access automatically.
The analogy Anthropic uses is hippocampal memory consolidation—the way a human brain replays the day's events during sleep and decides what to keep. Just as sleep consolidates short-term experiences into long-term memory, Dreaming converts raw session history into structured, searchable knowledge that persists across agent invocations. The system currently ships as part of Anthropic's managed agents offering and requires no configuration changes from developers to activate.
Early results are striking. Legal AI startup Harvey reported roughly a 6x improvement in task completion rates in internal testing once Dreaming was enabled—a number that would be extraordinary if it holds across diverse workloads. Anthropic is positioning this as a core part of its agentic infrastructure play, competing directly with third-party persistent memory tools like AgentMemory. The launch signals that Anthropic sees long-horizon memory not as a peripheral feature but as fundamental infrastructure for capable agents.
Dreaming is currently available to select managed agent partners and is expected to roll out more broadly later in Q2 2026. It represents one of the more concrete differences between running Claude via the API directly versus using Anthropic's managed agent platform—a distinction that will matter increasingly as enterprises choose their agentic infrastructure stack.
Panel Takes
The Builder
Developer Perspective
“A 6x task completion improvement from memory consolidation alone is a number that will drag every engineering team into a proof-of-concept. If it holds outside Harvey's environment, this reshapes how we think about agent session design entirely.”
The Skeptic
Reality Check
“One partner's 6x figure during internal testing is not a benchmark—it's a press release. Dreaming could also consolidate bad habits as readily as good ones, and there's no mention of mechanisms to detect or correct memory drift over time.”
The Futurist
Big Picture
“This is the first major step toward agents that actually improve at their jobs over time without human intervention. Dreaming is primitive compared to where memory consolidation will be in three years, but the direction is unmistakably right.”