Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Σ-Mem is an online memory system for LLM multi-agent setups that records each peer’s competence and relationship reliability.

It updates trust evidence from correctness feedback while keeping spectral changes bounded for stable adaptation.

The memory supports routing, steering, and reliability-weighted voting, improving decisions when peer outputs cannot be directly verified.

Across five Qwen-family models, it adapts better to reliability shifts and outperforms majority voting and the best fixed peer, including on out-of-distribution tasks.