ConMem is a contribution-aware memory framework for long-horizon steel-equipment inspection. It segments heterogeneous logs into functional evidence units, estimates each unit’s diagnostic contribution with a Shapley-style method, and retains high-value evidence under a constrained memory budget. The paper reports 76.0% QA accuracy, outperforming the strongest directly comparable baseline. Compared with naive LLM baselines using an 8K context, ConMem reduces average input tokens by 88.2% and response time by 86.6%. Deployment evidence across three inspection cycles indicates that it preserved a weak early signal associated with seal wear for targeted field inspection.
No heat snapshots are available in the last 24 hours.