LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation

WeintroduceLEGOMem, amodularproceduralmemoryframework for multi-agent LLM systems in workflow automation. LEGOMem distills successful executions into reusable full-task and subtask memories and allocates them to orchestrators and task agents to improve planning and execution. Across three retrieval variants, experiments on OfficeBench show consistent gains of 12–13 absolute points over memory-less and baseline methods, highlighting the importance of procedural memory for workflow automation.

Authors

• Dongge Han Microsoft
• Camille Couturier Microsoft
• Daniel Madrigal Diaz Microsoft
• Xuchao Zhang Microsoft
• Victor Rühle Microsoft
• Saravan Rajmohan Microsoft

Keywords

• Multi-agent systems
• Procedural memory
• LLM Agents
• Workflow