Meta-Learned Adaptive Memory Filtering for Robust Multi-Agent Collaboration

Abstract

Static defense rules may fail under evolving poisoning strategies. This work introduces a meta-learning framework that adapts memory filtering parameters based on historical contamination patterns. A meta-optimizer updates filtering thresholds using second-order gradient estimation across poisoning episodes. The framework was evaluated on cooperative resource management simulations with 50 agents across 200 poisoning scenarios. Adaptive filtering reduced cumulative task performance loss by 44.8% compared with fixed-threshold filtering. Convergence speed improved by 19.5% under dynamic attack conditions. Meta-adaptive filtering enhances resilience against non-stationary memory poisoning strategies.