Hybrid User Memory Filtering Algorithm for LLM-Based Knowledge Management Systems: Reducing Contextual Noise in Industrial Automation

This paper presents a single-company field study on hybrid user memory filtering for Large Language Model (LLM)-based knowledge management systems, aiming to reduce contextual noise from irrelevant or outdated persistent memories. We propose the Hybrid Adaptive Filtering Engine (HAFE), which combines intent classification, ontology-based filtering, behavioral reuse prediction, and collaborative role-level comparison. HAFE was integrated into an industrial KM platform deployed at a major steel producer. In a field experiment with 120 engineers, HAFE reduced irrelevant memory retention by 41%, improved Mean Reciprocal Rank (MRR) by 12.5%, and increased user satisfaction (SUS score) by 18% (all p < 0.01). The results suggest that proactive memory quality control can improve effectiveness and user experience in this specific industrial KMS setting, while further cross-domain validation is required.