Anthropic lets Claude agents "dream" to learn from past mistakes

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 The system merges duplicate entries, removes outdated notes, and surfaces recurring patterns such as typical failure modes in a particular workflow or repeated team preferences.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.