Thought Leadership
Anthropic is introducing “Dreaming” for Claude Agents. The memory feature analyzes workflows, identifies patterns, and doubles usage limits for subscribers.
At its developer conference “Code with Claude,” AI company Anthropic unveiled a technological innovation for its platform: so-called “Dreaming.” This is not a biological process but rather an automated procedure for memory optimization in “Managed Agents.” The feature is designed to enable artificial intelligences to systematically process past events and preserve relevant information for future tasks. At the same time, Anthropic is responding to the high demand on its computing capacity by doubling usage limits for professional subscribers.
Systematic pattern recognition through “dreaming”
The “Dreaming” feature is currently in a research preview phase and is exclusively available for Managed Agents on the Claude platform. Anthropic describes the procedure as a scheduled process in which the AI reviews past sessions and memory contents. In this “dream state,” the software identifies specific patterns that are often not recognizable to a single agent during the active processing of a task.
According to the developers, these patterns include recurring errors, workflows that various agents on a team have agreed upon, as well as shared preferences within a project structure. The goal is a restructuring of memory in order to maintain a high “signal-to-noise ratio.” Information that is crucial for the long-term success of a project is curated and permanently stored in the agent harness’s memory. Users can choose whether this process should run fully automatically or whether they would like to manually review the proposed changes to the AI’s memory.
Distinction from conventional context compaction
In the world of large language models (LLMs), the limited context window represents a technical hurdle. Until now, many models have used a procedure called “compaction.” This involves periodically analyzing long conversations to remove irrelevant data and retain only the core of the conversation in active memory.
“Dreaming,” however, goes beyond this simple cleanup. While compaction is usually limited to a single conversation with a single agent, dreaming works periodically across different sessions and agent instances. This is particularly significant for complex multi-agent orchestrations, where several specialized AIs work for hours or days toward a shared goal. Information learned by Agent A in an early phase of the project can thus be made usable for Agent B in a later phase through the “dream process.”
Claude agents take over coordination of computing power
Managed Agents represent a higher-tier alternative to direct use of the Messages API. Anthropic defines them as a preconfigured, managed agent framework (“Managed Infrastructure”) developed specifically for long-running projects. Developers no longer need to provide or manage the infrastructure for executing the agents themselves; instead, the Claude platform handles the coordination and scaling of computing power.
This infrastructure is designed to handle tasks that require continuous interaction over multiple hours. “Dreaming” functions here as a necessary maintenance instance to ensure that, despite the long runtime, agents do not lose focus on the original objectives or become inefficient due to outdated information.
Doubling of usage limits for Pro and Max subscribers
In addition to the functional innovations, Anthropic announced that it is significantly increasing usage limits for subscribers of the “Claude Pro” and “Claude Max” plans. The previous five-hour limits are being doubled. This decision follows ongoing frustrations within the user base, as the company’s computing infrastructure has struggled in recent months to keep pace with rapidly rising demand.
The doubling of limits is intended to ensure that developers and companies relying on Claude Code and the associated agent workflows can advance their projects without frequent interruptions caused by capacity bottlenecks. Anthropic emphasizes that the optimization of computing efficiency, including through features such as “Dreaming,” has helped make this capacity expansion possible.
Progress in orchestration and objectives
Alongside the introduction of “Dreaming,” two additional features were moved from the research phase into broader availability: “Outcomes” and “Multi-Agent Orchestration.” “Outcomes” allows developers to define specific success criteria for agent tasks against which the AI can independently measure its performance.
Multi-agent orchestration enables the seamless collaboration of various specialized instances. These features are now accessible to a wider circle of developers. Access to the new “Dreaming” feature, however, initially remains limited to selected participants of the research preview by request.
