Anthropic has unveiled a new feature called “Dreaming” for its Claude AI platform, designed to help AI agents improve their performance by reviewing past work, identifying patterns, and refining future actions between active sessions. The announcement was made during the company’s developer event in San Francisco.
According to Anthropic, the feature allows AI systems to analyze previous interactions and update memory files containing user preferences, workflows, and contextual information. Instead of starting every session from scratch, the AI can continuously refine its understanding and improve task execution over time.
The company describes “Dreaming” as a step toward more autonomous and adaptive AI agents capable of learning from experience. The technology is intended to improve long-term performance in areas such as coding, research, finance, and enterprise workflows, where retaining context and recognizing recurring patterns can significantly enhance efficiency.
The feature was announced alongside several major infrastructure updates. Anthropic also revealed a partnership with SpaceX to access computing capacity from the Colossus 1 data center in Memphis, Tennessee, which houses more than 220,000 Nvidia processors. The additional infrastructure is expected to help Anthropic expand Claude’s capabilities and support increasing demand for products like Claude Code.
With the added compute power, Anthropic said it is doubling usage limits for Claude Code paid plans, removing peak-hour restrictions for Pro and Max subscribers, and increasing request volumes for its Claude Opus models.
The launch reflects a broader industry trend toward “agentic AI,” where AI systems move beyond simple chatbot interactions and begin operating as semi-autonomous agents capable of self-improvement, task delegation, and persistent contextual learning. Analysts view these developments as an important step toward more independent AI systems that can continuously optimize their behavior over time.
