An operations memory-enhanced multi-agent system for human-centric manufacturing process monitoring and decision support

Human-centric smart manufacturing requires process monitoring and decision support that remain adaptable and trustworthy under changing conditions, yet existing systems treat models, constraints, and operator interaction as loosely coupled components. This paper formulates local process monitoring and decision support as a memory-enhanced multi-agent system in which agents with distinct observability, timescales, and authority limits realize three coordinated loops: per-operation monitoring, event-triggered decision support, and memory governance. Long-horizon process knowledge is organized into episodic, semantic, and procedural memories and accessed via dual-level retrieval, where fast context-based lookup serves monitoring and deep retrieval serves decision support, while all parameter changes require operator authorization. Experiments on a robotic drilling cell show that recent episodic context improves monitoring accuracy, and memory-informed interval recommendation supports human-in-the-loop decisions with lower mean surface roughness and fewer violation-level outcomes.