At its annual Code with Claude conference in San Francisco on May 6, 2026, Anthropic unveiled a new capability called Dreaming: a scheduled process that lets Claude agents review their past sessions, extract recurring patterns, and reorganize their memory without any human intervention. Early results are striking. Harvey, the AI-powered legal platform, reported task completion rates roughly six times higher after enabling the feature.
Key Takeaways
- Anthropic launches Dreaming in research preview: an automated memory consolidation process for Claude agents
- Agents analyze past sessions, merge duplicates, remove contradictions, and surface new observations to improve future performance
- Harvey (legal AI) saw a 6x increase in task completion; Wisedocs (medical review) cut processing time by 50%
Memory That Consolidates Itself Between Sessions
The metaphor is intentional. Just as the human brain consolidates experience during REM sleep, Dreaming does not simply accumulate information.
It sorts, consolidates, and discards what has become irrelevant.
In practice, Claude agents build up data in memory across sessions. Without a cleanup mechanism, those memories eventually contradict each other, duplicate, and lose relevance. Dreaming reads an existing memory store alongside past session transcripts, then produces a reorganized store: duplicates merged, obsolete entries replaced, fresh observations surfaced.
Alex Albert, head of research product at Anthropic, described Dreaming as analogous to the way individuals within organizations build skills through repeated work on a task.
The critical distinction from a standard memory system: Dreaming does not touch the underlying model weights. It updates the external memory layer that shapes Claude’s behavior in future sessions. Everything remains observable and auditable.
Harvey, Wisedocs, Netflix: Real Numbers From Early Access
Results from early adopters are already public.
Harvey uses Managed Agents to coordinate complex legal work, including long-form drafting and document creation. With Dreaming, their agents carry forward file-format-specific workarounds and tool usage patterns from one session to the next. Task completion rates increased roughly six times in their tests.
Wisedocs, which specializes in medical document review, cut processing time by 50% using Outcomes, another feature announced at the same event. Outcomes lets an evaluator agent score the output of a primary agent against criteria defined in advance.
Netflix relied on the third new capability: multi-agent orchestration. Its platform team built an analysis agent that processes logs from hundreds of builds across different sources, identifying recurring patterns across thousands of applications.
These deployments reflect an enterprise momentum Anthropic has been building for months, with Claude increasingly embedded in mission-critical workflows across legal, healthcare, and media.
Also on Horizon:
- Claude Code and AI Agents Are Blowing Up Dev Budgets
- OpenAI Launches Daybreak, Its AI Cybersecurity Platform
- Unity AI Open Beta: AI Agent Inside the Unity 6 Editor
Orchestration, Evaluation, Memory: One Closed Loop
Dreaming was not designed as a standalone tool.
Anthropic positions it alongside the two other features announced at the conference: multi-agent orchestration, which lets a primary agent distribute work to specialized sub-agents with their own contexts, and Outcomes, focused on output quality criteria.
Together, the three form what Anthropic describes as a continuous improvement loop: orchestration parallelizes the work, Outcomes evaluates the output, Dreaming consolidates and learns.
Dreaming is currently in research preview and requires a separate access request. Compatible models are claude-opus-4-7 and claude-sonnet-4-6. Outcomes and multi-agent orchestration are available in public beta.
The San Francisco Code with Claude conference on May 6 is the first of three stops: London on May 19, Tokyo on June 10. Ami Vora, Anthropic’s new chief product officer who succeeded Mike Krieger, delivered the keynote.
Follow the story on Horizon.
Pingback: Seedance 2.0: ByteDance's AI Video Tool Rattles Hollywood - Horizon IA