M-Plan Framework Combines Long-Term Memory Reorganization With Global-Local Guidance on ALFWorld, Webshop, PDDL

M-Plan implements a closed-loop multi-agent system that generates milestones and local hints during execution while optimizing task traces for long-term storage after completion. The arXiv preprint 2610.27032 describes global-local guidance that supplies current-task context only when agent confusion is detected, avoiding retrieval of unfiltered historical noise. Experiments across multiple base models show consistent gains over retrieval-only and guidance-only MAS variants.

Prior MAS work such as ReAct and Reflexion demonstrated single-turn strength but plateaued on multi-turn horizons because memory remained either static or unadapted. M-Plan addresses both gaps by treating memory writes as an explicit post-task optimization step and by conditioning short-term hints on real-time execution state. This design aligns with observed failure modes in long-context LLM agents where raw retrieval injects irrelevant tokens.

Operationally the plug-and-play structure allows existing agent scaffolds to adopt the memory arrangement layer without retraining. Production pipelines that already log trajectories can insert the reorganization stage before vector ingestion, lowering noise in subsequent retrievals. The reported efficiency lift stems from fewer redundant steps once local hints replace broad memory scans.

Next steps include open release of the memory controller code and targeted evaluations on enterprise workflow benchmarks that extend beyond the three academic environments tested.