Anthropic adds self‑improving ‘dreaming’ system to Claude
Anthropic has launched a self-improving “dreaming” system for Claude Managed Agents to help AI agents refine memory and learn over time.
Anthropic is teaching AI agents to improve themselves between tasks. On 6 May 2026, Anthropic introduced a new research-preview feature called “dreaming” for Claude Managed Agents. The company says the system helps agents refine memories, identify patterns and improve performance over time with less human intervention.
The update arrives alongside new features such as outcomes, multi-agent orchestration and webhooks, which Anthropic says are designed to make enterprise AI agents more autonomous while remaining reviewable and governed.
Claude agents can now refine memories
Anthropic describes dreaming as a scheduled memory-curation process. Instead of storing raw interactions indefinitely, the system reviews prior sessions and memory stores, merges duplicate information, removes outdated entries and highlights recurring patterns such as repeated mistakes or team preferences.
The goal is to prevent what AI practitioners often call “memory rot”, where long-running agent systems become cluttered with stale or contradictory context. According to Anthropic, dreaming creates a feedback loop where agents gradually improve across projects and sessions rather than behaving like stateless assistants that restart from scratch every time.
The dreaming process runs asynchronously
The AI company says dreaming works as an asynchronous workflow. Developers can feed the system an existing memory store along with large batches of prior session transcripts. Claude then produces a reorganised memory layer that teams can either approve, reject or modify before deployment.
Anthropic noted that the original inputs remain untouched during the process, allowing enterprises to safely review updates before agents adopt them. During the research-preview phase, dreaming supports select Claude models and operates under default rate limits.
Typical runs reportedly take minutes rather than hours, making the feature practical for overnight or off-peak processing schedules.
Anthropic is pairing memory with self-correction
Dreaming is being rolled out alongside another feature called outcomes. Anthropic says outcomes allow agents to evaluate their own work against predefined quality rubrics. In internal testing, the company claims the system improved task success rates by up to 10 points compared with standard prompting workflows.
The company also reported improvements in complex file-generation tasks involving docx and pptx outputs, suggesting that agents become more reliable when memory refinement and self-evaluation are combined.
Anthropic believes these systems work best alongside multi-agent orchestration, where a lead agent delegates subtasks to specialist agents operating on shared projects.
Early enterprise use cases are emerging
Anthropic says early adopters are already testing the system in legal drafting, quality assurance and large-scale log analysis workflows.
In one example shared by the company, legal agents reportedly retained filetype preferences and tool-specific workarounds between sessions, helping improve completion consistency over time. Another enterprise team used orchestration features to analyse batches of engineering logs in parallel while surfacing only actionable anomalies for human review.
These remain company-provided examples rather than independently verified benchmarks, but they offer an early indication of how enterprises may deploy persistent agent systems in production environments.
Anthropic is pushing beyond stateless chatbots
The broader direction behind the update is becoming clearer. Anthropic is steadily repositioning Claude from a conversational assistant into a long-running enterprise agent platform capable of retaining context, coordinating subtasks and refining workflows over time.
Under CEO Dario Amodei, the company has repeatedly emphasised that future AI systems must combine stronger capabilities with governance, reviewability and policy controls. The managed-agent architecture reflects that approach by allowing organisations to inspect, audit and approve how agent memory evolves.
Why image and video tools are becoming AI’s biggest growth engine, not chatbots
Why enterprises are paying attention
For enterprises experimenting with AI agents, the practical value lies in consistency. One of the biggest problems with long-running AI workflows is that systems gradually accumulate noisy context, duplicate memory and conflicting instructions.
Anthropic’s dreaming system attempts to clean and restructure that memory layer before it degrades operational quality. For Indian enterprises and CIOs, the takeaway is less about a flashy AI feature and more about operational reliability.
Organisations piloting AI agents for customer support, compliance, reporting, or documentation workflows will increasingly need systems that can improve incrementally while remaining auditable and controllable.
Anthropic’s latest update suggests the AI industry is moving steadily toward agents that do not simply respond to prompts, but continuously learn how organisations work over time.