Descriptions:
Anthropic product manager Mahes introduces memory as the next major primitive for frontier AI agents, positioning it as the missing ingredient for continuous self-learning in long-horizon, multi-agent systems. The talk situates memory alongside previously launched primitives — MCP, Skills, and the Agent SDK — and explains how agents can use it to track success criteria, learn from environment-specific patterns, and share learnings across concurrent agent instances running in shared environments.
The centerpiece announcement is Dreaming, a new product launching in research preview within Anthropic’s managed agents API. Dreaming is a batch asynchronous process that reviews recent agent session transcripts, identifies recurring mistakes and successful strategies, and automatically produces updated, organized memory content. Early adopters report striking results: Harvey saw a 6x increase in task completion rates on realistic legal benchmark scenarios, while Rockutin cut first-pass errors in internal knowledge agents by 90% and also achieved better token efficiency and lower latency.
The presentation also covers the architecture behind Anthropic’s memory system — modeling memory as files in a filesystem, distinguishing between content, structure, and process layers — and discusses the limitations that motivated Dreaming, including siloed per-session learning and inefficient large-scale memory maintenance. The talk frames self-managed, shared memory as foundational infrastructure for the next generation of autonomous, continuously improving agent swarms.
📺 Source: Claude · Published May 08, 2026
🏷️ Format: Keynote Launch
