Anthropic Launches 'Dreams' Feature for Self-Improving Claude Agents

During its Code with Claude developers' conference, Anthropic introduced a new capability called "Dreams" for its Claude AI platform. The feature allows managed agents to review historical interactions and reorganize their memory stores, theoretically improving performance over time through pattern recognition and memory consolidation.

According to official documentation from the company, the Dreams feature functions as an asynchronous job that reads an existing memory store alongside transcripts from up to 100 prior sessions. The system then produces a reorganized memory store with duplicates merged, stale entries replaced, and new insights surfaced. The original memory store remains untouched, creating a separate output that developers can review before adoption.

This addresses a practical problem in long-running agent deployments. As agents work alongside users over extended periods, they continuously write information to memory stores. These writes are local and incremental, which means memory stores accumulate duplicates, contradictions, and stale entries across multiple sessions (a problem that has plagued users for years, frankly). The dreaming process attempts to clean this up automatically.

Anthropic's documentation specifies that Dreams currently supports only Claude Opus 4.7 and Claude Sonnet 4.6 models. The feature is available through special beta headers for developers using the company's Managed Agents platform. Developers must request access to try it during the research preview phase.

The technical implementation requires specific API headers. All Managed Agents API requests need the managed-agents-2026-04-01 beta header, while Dreams additionally require the dreaming-2026-04-21 beta header. The SDK sets these automatically, but developers working directly with the API must configure them manually.

Running a dream takes minutes to tens of minutes depending on input size. Developers can provide custom instructions during execution, such as asking Claude to focus on coding preferences while ignoring temporary debugging conversations. The process can fail under certain conditions, including oversized memory stores, unavailable session data, or runtime timeouts.

Usage will be billed at standard API token rates, with costs varying based on the size and number of sessions processed. Anthropic cited the example of Harvey, a legal AI company that used the feature to identify file-type workarounds and tool-specific patterns across extended drafting sessions. The company reported a sixfold increase in task completion rates.

Independent reporting from ZDNET notes that the feature surfaces patterns that a single agent can't see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across teams. The outlet also highlights Anthropic's tendency to anthropomorphize its products, pointing to the company's history of naming features after human cognitive processes.

Anthropic has a documented pattern of this approach. In January, the company published a constitution for Claude intended to help shape the chatbot's decision-making. In August 2025, Anthropic launched a feature that lets Claude end toxic conversations with users for its own well-being. When it retired its Opus 3 model in January, Anthropic set it up with a Substack so it could blog on its own.

Alongside Dreams, the company also introduced an "Outcomes" feature that allows AI agents to evaluate their own work against predefined metrics. With outcomes, developers write a rubric describing what success looks like, and a separate grader evaluates the output against those criteria. When something isn't right, the grader pinpoints what needs to change and the agent takes another pass.

The physical experience of using Dreams involves waiting through the asynchronous process while monitoring status updates. Developers poll the dream by ID to check progress, watching the status move from pending to running to completed. The session is archived when the dream reaches a terminal state, so the transcript remains available afterward.

Once the dream completes, the output holds a rebuilt memory store. Developers can attach it to future sessions as a memory store resource in place of or alongside the input memory store. Alternatively, they can discard it entirely. The dream itself never deletes or modifies its inputs.

Whether this actually translates to meaningful improvements in production environments remains to be seen. The feature requires careful configuration, introduces additional API costs, and depends on having sufficient session history to analyze. For teams running short-lived agents or those with limited interaction data, the value proposition becomes less clear.

The naming choice itself invites questions about whether Anthropic is selling a technical capability or a narrative. Functionally, the dreaming feature makes sense: it refines an agent's pool of references for how it should work, which should ideally make it better at whatever task you give it. But calling it "dreaming" adds a layer of abstraction that doesn't necessarily help developers understand the underlying mechanics.

For now, the feature sits in research preview with access restrictions. Developers must request access, and the beta headers suggest this isn't a stable release. Whether users actually pay for it remains the real question, especially given the additional token costs and the complexity of managing memory stores across multiple agents.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn