Entropy-Based Anomaly Detection for Memory Integrity Preservation in Multi-Agent Systems

Abstract

Memory poisoning often alters statistical properties of shared state representations. This study proposes an entropy-based monitoring mechanism to detect anomalous memory distributions in collaborative agent environments. Shannon entropy and conditional entropy metrics are computed over memory state vectors at each synchronization step. A divergence threshold based on Kullback–Leibler distance identifies suspicious memory updates. Experiments were conducted on distributed planning simulations with 150 agents and controlled poisoning injection rates from 5% to 30%. Detection precision reached 93.1% and recall 88.7% at a divergence threshold of 0.27. Early intervention reduced system-wide contamination by 48.3% compared with no monitoring. Entropy-driven detection provides a lightweight and scalable solution for preserving memory integrity.