Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems
The paper proposes a memory layer that tracks which peer agents are reliable, instead of just storing what they said.
Σ-Mem records competence evidence for individual peers and relationship evidence across a peer set, then updates those records from correctness feedback after decisions. The authors say its bounded spectral updates allow online adaptation without retraining the underlying models. In evaluations with five Qwen-family models, it adapted to reliability shifts, generalized to unseen peers and domains, and beat majority voting and the best fixed peer on the full OOD set. HF Daily Papers' note
Σ-Mem records competence evidence for individual peers and relationship evidence across a peer set, then updates those records from correctness feedback after decisions. The authors say its bounded spectral updates allow online adaptation without retraining the underlying models. In evaluations with five Qwen-family models, it adapted to reliability shifts, generalized to unseen peers and domains, and beat majority voting and the best fixed peer on the full OOD set. HF Daily Papers' note
score 4