Topic

From “Records” to “Assets”: How MemOS Builds the Memory Engine for OpenClaw Multi-Agent Systems

As intelligent agents increasingly enter real-world business scenarios, the traditional memory mechanism—where conversations end without lasting knowledge accumulation—can no longer support enterprise-level collaboration and continuous evolution. Drawing on my experience leading the MemOS project, this talk will share how to build a practical enterprise-grade memory engine around MemOS to address challenges in multi-agent scenarios, including high-cost forgetting, memory black boxes, loss of experience, and collaboration silos. The presentation will focus on MemOS’s design approaches in memory collection, intelligent deduplication, hierarchical retrieval, task summarization, skill evolution, and isolation with controlled sharing. It will also explore how conversational data can gradually evolve into reusable, manageable, and continuously evolving organizational assets, enabling stronger continuous learning and collaboration capabilities for enterprise agent systems. Outline: Four major bottlenecks of OpenClaw’s native memory in enterprise scenarios The core positioning of MemOS: turning memory into an evolving asset Multi-agent memory architecture design: isolation, sharing, and collaboration An intelligent closed loop from conversation to tasks to skills Key lessons from real-world implementation: deduplication, retrieval, accumulation, and governance The next evolution of the ShrimpPool agent system