Σ-Mem memory tracks AI agent trustworthiness, beats majority voting

Multi-agent LLM systems struggle to know which agents to trust. Σ-Mem, a new online reliability memory from researchers Peilin Feng, Suorong Yang, and Soujanya Poria, tackles this by storing two forms of evidence: historical competence for each peer and relationship quality across the peer set. Both are kept as real symmetric states and updated whenever a decision receives correctness feedback. Crucially, the update is mathematically bounded via Weyl's inequality, so the memory adapts online without any retraining of the underlying models. This makes Σ-Mem a reusable component: it can steer a central model's residual reasoning, route queries to the most reliable peer, or weight votes based on accumulated track records.

In experiments with five Qwen-family models, Σ-Mem adapted to counterfactual reliability shifts and generalized to unseen agents and domains. Direct readouts from the memory beat both majority voting and the best fixed peer across the full out-of-distribution evaluation set. Performance also improved consistently as more feedback accumulated, showing that the memory progressively builds actionable reliability knowledge. This establishes reliability memory as a foundation for adaptive coordination, moving multi-agent systems from naive consensus to evidence-based trust.