Anthropic: Claude Self-Corrects with 'Dreaming' Mode

Anthropic Introduces 'Dreaming' Mode for Claude

Anthropic has recently unveiled a new feature for its AI agents, called "Dreaming." This mode, which operates asynchronously, aims to automatically enhance the agents' performance by analyzing interaction history to optimize memory management.

In a landscape where tech giants like OpenAI and Google have already introduced similar features, Anthropic stands out with its unique approach. OpenAI, for instance, launched a Memory Consolidation feature in its Agents SDK, while Google integrated persistent memory features with Vertex AI Memory Bank in July 2025. Anthropic, on its part, has incorporated the "dreaming" mode into its Claude Managed Agents.

How 'Dreaming' Mode Works

This feature allows Claude users to create or deploy AI agents capable of self-improvement. By analyzing up to 100 past sessions, the agent can identify recurring patterns and organize its memory more effectively. The asynchronous process sorts and eliminates duplicates to create a reorganized memory store.

When a new call is made, the conversation history is provided to Claude, which processes the request and returns a response. Anthropic's SDK, a code library designed to facilitate program writing, manages the technical complexity by automatically compressing and summarizing the history. This keeps a lightweight history ready for use during the next call.

A Python code example provided by Anthropic illustrates how to create a dream. Inputs include the existing memory store and an array of sessions. Instructions can be passed to guide the dreaming process, and this feature is optimized for Claude Opus models (versions 4.8 and 4.7) and Claude Sonnet. The cost of using the Dreaming mode is charged according to the standard token rates of the API, depending on the chosen model.

Restricted Access to the Feature

Access to this feature is currently limited. Available in research preview within Managed Agents, it requires filling out an official form from Anthropic. Users must have an active Anthropic API account and a professional email address.

To obtain the UUID (Universally Unique Identifier) of the API organization, one must go to the Claude Console. In the bottom left of the interface, clicking on the user's name allows copying the organization ID code. Even if this feature is not used immediately, it is possible to click on "MCP Tunnels" to securely access external tools. Users can also explore Python and Claude API CLI for "SDK and tools."

Applications of 'Dreaming' Mode in Claude Managed Agents

The Dreaming Mode is particularly useful when agents make repeated errors or need to assimilate numerous interactions. It is suited for workloads requiring prolonged execution or cloud infrastructure. Here are three examples of use.

• Preventing Repeated Errors
  For instance, an AI agent that fails to send an email with a file that is too large can, thanks to the "Dreaming mode," identify the recurring issue and propose a solution. By detecting the pattern "File > 25 MB + Email = Failure," the agent can suggest creating a download link for large files, such as those from Google Drive or Dropbox, instead of attaching them directly.

  Technically, during Dreaming Mode, the model is fed two types of data: session logs with raw transcripts of failures, and a system directive to extract "playbooks" from errors. The model generates a structured text file, often in Markdown or JSON, containing the new rule, which is then stored in a persistent Memory Store.

  When a new task is performed, the harness, which surrounds the agent to verify its proper behavior, retrieves these rules and inserts them at the beginning of the "System Prompt." This allows the agent to have these rules in its immediate context.

• Knowledge Consolidation
  A small business using an AI agent to respond to customer inquiries can benefit from Dreaming Mode. For example, if the agent receives 50 questions about a new update over the week, it can activate Dreaming Mode over the weekend to analyze these conversations and draft a concise memory note summarizing the exact procedure.

  Triggering Dreaming Mode results in a scheduled task, such as a script that runs automatically over the weekend. Claude then produces the playbook, which is subsequently stored in a database constituting the agent's memory. Thus, when customers ask questions, the agent can respond immediately.

• Improving Agent Performance
  A small business using an AI agent for server infrastructure maintenance can also leverage Dreaming Mode. During this phase, the agent reorganizes its technical knowledge base, detects outdated instructions, archives old ones, and marks new ones as "Standard Priority." It removes unnecessary details from debugging logs to retain only the essentials, allowing the agent to respond more quickly and accurately while consuming fewer tokens.

However, the Dreaming mode has certain limitations. It may delete useful information, and large knowledge bases can incur token costs. In a business context, it is crucial for a manager to periodically validate the changes made by Dreaming Mode, especially on critical topics like security or compliance.