Sigma-Mem Gives Multi-Agent Systems a Reliability Memory That Tracks Which Peers to Trust and When

arXiv·medium signalXBlueskyLinkedInCopy link

Sigma-Mem (arXiv 2607.27958, July 30) targets a gap in agent memory: existing systems store interaction content but not which peer agents are competent under which conditions, a real problem when a central model cannot directly verify plausible or correlated peer responses. It maintains competence evidence per peer and relationship evidence across the peer set as real symmetric states updated from post-decision correctness feedback, with Weyl's inequality bounding the spectral change per update so adaptation is stable online without retraining. Across five Qwen-family models it adapts to counterfactual reliability shifts, generalizes to unseen peers and domains, and direct memory readouts beat both majority voting and the best fixed peer across the full out-of-distribution set.