Descriptions:
Ravi, who leads the API Knowledge team at Anthropic, presents a focused technical talk on Memory and Dreaming — two recently launched features for Claude Managed Agents designed to enable agents to learn and improve across sessions. The core thesis: every agent task should perform better than the last, with knowledge transferring not just within a single agent but across entire swarms working on shared problems.
Anthropicʼs memory architecture leverages Claudeʼs native proficiency with file systems — agents read, write, and reorganize structured files using tools like bash, with Opus 4.7 cited as the current state-of-the-art for this pattern. The system includes version-controlled audit trails, per-agent attribution, and a standalone CRUD API supporting enterprise operations like bulk export and redaction. Real-world results are cited: Racketin saw a 97% decrease in first-pass errors after deploying memory in production, and Wise Docs reduced recurring failures in its document verification pipeline using cross-session memory.
Dreaming, available in research preview, is a fully decoupled feedback loop that analyzes session transcripts at scale — nightly, hourly, or triggered by events — to detect patterns of inefficiency and mistake across agents, then proposes globally optimized memory updates rather than locally optimal ones. Harvey reported a 6x increase in completion rates on its legal benchmark after enabling Dreaming. Together, the two features represent Anthropicʼs most substantial step toward self-improving agent systems.
📺 Source: Claude · Published May 21, 2026
🏷️ Format: Deep Dive
