Reproduce “LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongge Han, Camille Couturier, Daniel Madrigal Diaz, Xuchao Zhang, Victor Rühle, Saravan Rajmohan
Microsoft researchers introduce LEGOMem, a modular procedural memory framework designed for multi-agent Large Language Model systems to automate complex workflows. The framework consistently improves task success rates by over 12% on the OfficeBench benchmark and enables smaller LLMs in multi-agent teams to match or exceed the performance of larger, memory-less LLM counterparts.
Key insights
- Modular procedural memory consistently and significantly boosts performance across diverse LLM team configurations and task difficulties, outperforming memory-less and adapted single-agent baselines.
- LEGOMem empowers smaller LLMs, allowing hybrid teams with smaller task agents to surpass memory-less larger LLM teams and even full small LLM teams to outperform memory-less hybrid teams.
- Orchestrator memory is crucial for effective multi-agent collaboration, enhancing high-level planning, task decomposition, and delegation, which leads to improved efficiency and reliability in execution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.