Summary
Kevin, an engineer at Anthropic, presents a live workshop introducing two new features in Anthropic’s Claude managed agents platform: persistent memory stores and a background consolidation process called dreaming. Published on Anthropic’s official Claude channel, the session walks through both features hands-on using the CLI and the Anthropic console interface, with a downloadable workshop repository for attendees to follow along.
Memory stores are persistent, file-system-like storage resources that attach to agent sessions, enabling agents to read and write information across multiple independent sessions rather than starting fresh each time. Kevin demonstrates the contrast directly: a stateless agent forgets everything between sessions, while an agent with a mounted memory store can answer questions about information shared hours earlier. Memory files are versioned, browsable and editable through the console UI, and accessible via dedicated API endpoints — giving developers fine-grained control over what their agents remember.
The dreaming feature addresses a compounding problem with naive memory accumulation: agents tend to write redundant, stale, or unorganized information over time, causing memory stores to grow unbounded without a cleanup mechanism. Dreaming is an asynchronous multi-agent harness that runs over a specified memory store and a set of session transcripts, fact-checking entries, consolidating duplicates, and distilling new insights that individual agents may have missed. Together, memory stores and dreaming position Anthropic’s managed agents platform as a serious infrastructure layer for long-running, stateful production workflows — a meaningful capability expansion for enterprise teams building persistent AI agents.
📺 Source: Claude · Published May 23, 2026
🏷️ Format: Keynote Launch
