Anthropic lets Claude agents "dream" to learn from past mistakes
May 10, 2026
At the Code with Claude conference on May 6, 2026, Anthropic introduced a "dreaming" system for managed agents. It curates memories from past sessions and, according to Harvey, lifted task completion rates roughly sixfold.
Anthropic gives Claude agents a scheduled self-reflection
At the second edition of the Code with Claude developer conference in San Francisco, Anthropic introduced a feature called Dreaming for Claude Managed Agents on May 6, 2026. It runs as a scheduled background process that combs through past sessions and an agent's memory store, extracts patterns, and produces a curated memory layer that teams can review, modify, or reject.
What Dreaming actually does
According to Anthropic and Simon Willison's live blog from May 6, 2026, Dreaming is not a training phase but a memory-curation step. The system merges duplicate entries, removes outdated notes, and surfaces recurring patterns such as typical failure modes in a particular workflow or repeated team preferences.
Three layers working together
According to 9to5Mac and Anthropic's own announcement, the new generation of managed agents covers three building blocks. First, memory: cross-agent memory that lives across sessions. Second, outcomes: a scoring system that tracks success and failure of actions. Third, dreaming: the scheduled consolidation of those memories.
Numerical results
Anthropic cites two early adopters. Harvey, an AI provider for legal work, reports a roughly sixfold increase in task completion rate after introducing dreaming. Wisedocs, focused on medical document review, halved its average review time per document, according to Anthropic.
Why it matters
Long-term memory has been the bottleneck for agentic AI for months. If an agent starts from scratch every day, it cannot learn anything about the system it operates in. Dreaming places a layer between the model and the task world that condenses recurring experiences and makes them reusable.
Anthropic positions itself this way against providers like OpenAI and Google, both of which have announced extended memory mechanics in recent weeks. Unlike classic RAG architecture, which loads external documents, dreaming is about processing internal session data, with governance: teams can review the consolidation proposals before they reach production agents.
In plain language
Imagine an office assistant who jots down notes every day on tasks, mistakes, and small tricks. At the end of the week, the assistant sits down, organizes the notes, throws away duplicates, and writes a clean cheat sheet. That is exactly what Anthropic's dreaming does for Claude agents: they tidy up their own notes and become faster and more accurate on the next assignment.
A practical example
A mid-sized German law firm with 60 lawyers deploys a Claude managed agent for contract review in 2026. In the first four weeks, the agent collects notes on typical clause pitfalls in German tenancy law during each session. After the first dreaming run, a curated memory layer emerges containing recurring review patterns, firm-specific templates, and avoidable mistakes. The lawyers review the proposals, reject two of them, and accept the rest. As a result, average processing time per tenancy contract drops from 38 minutes to 24, and the rate of after-the-fact corrections is cut in half.
Scope and limits
First, dreaming is in a research-preview phase according to Anthropic. It does not promise production SLAs and may change.
Second, the cited numbers are vendor-reported. A sixfold task completion rate at Harvey and a 50 percent reduction in review time at Wisedocs are not validated by independent audits and refer to specific workflows that may not be representative of other industries.
Third, memory consolidation remains a potential security risk: if an agent uncritically learns patterns introduced through crafted inputs, prompt injection can over time eat its way into the memory store. Anthropic addresses this with an approve, modify, or reject interface, but the ultimate responsibility lies with the operator.
SEO and GEO keywords
Anthropic, Claude, Dreaming, Managed Agents, Code with Claude, Memory Consolidation, Agentic AI, Harvey, Wisedocs, AI Agents, Outcomes, Multi-Agent Orchestration, 2026.
π‘ In plain English
Anthropic gives its Claude agents a break to tidy up their own memories. Old sessions become a clean cheat sheet that humans can review before the agent uses it again.
Key Takeaways
- βAnthropic introduced dreaming on May 6, 2026 at the Code with Claude conference in San Francisco.
- βDreaming is a scheduled memory-curation process for Claude managed agents, not a new model.
- βIt joins memory and outcomes as three new building blocks for cross-agent learning.
- βHarvey reports a roughly sixfold task completion rate after rollout, according to Anthropic.
- βWisedocs halved review time for medical documents.
- βThe feature is in research preview and uses an approve/modify/reject interface for governance.
Sources & Context
- Anthropic introduces "dreaming," a system that lets AI agents learn from their own mistakes (VentureBeat, May 6, 2026)
- Anthropic updates Claude Managed Agents with three new features (9to5Mac, May 7, 2026)
- Live blog: Code w/ Claude 2026 (Simon Willison, May 6, 2026)
- Anthropic will let its managed agents dream (The New Stack)
- Anthropic is letting Claude agents 'dream' so they don't sleep on the job (SiliconANGLE, May 6, 2026)
- Anthropic Unveils 'Dreaming' Feature to Help Its AI Agents Self-Improve (US News, May 6, 2026)