Causal Inference–Driven Repair of Memory Poisoning in Multi-Agent Collaborative Systems

Abstract

Memory poisoning in multi-agent systems may distort downstream decision processes by altering causal dependencies among shared states. This study introduces a structural causal model (SCM) to identify and repair corrupted memory segments. Agent memory variables are modeled as nodes in a directed acyclic graph learned through constraint-based causal discovery. Intervention tests estimate causal effect shifts caused by adversarial perturbations. Corrupted nodes are reconstructed using do-calculus–based counterfactual inference. Experiments were conducted on a 90-agent distributed task-planning simulation with injected poisoning rates of 10%–30%. Causal repair reduced downstream decision error from 34.6% to 12.8% under a 20% poisoning scenario. Recovery latency decreased by 41.3% compared with naive memory reset strategies. Causal modeling provides interpretable and effective repair mechanisms for collaborative agent memory integrity