Memory is essential for long-horizon AI agents, but existing systems typically store interaction content rather than modeling which agents can be trusted and under what conditions. This gap becomes critical in multi-agent systems, where a central model often cannot directly verify plausible or correlated responses from peers.
To address this, researchers introduce Σ-Mem, an online reliability memory that records historical competence evidence for individual peers and relationship evidence across the peer set. Both forms are maintained as real symmetric states and updated from post-decision correctness feedback, allowing stable online adaptation without retraining the underlying models.
Σ-Mem provides a general interface that supports three applications: residual steering of a central model, response-free peer routing, and reliability-weighted voting. Experiments across five Qwen-family models show that Σ-Mem adapts to counterfactual reliability shifts and generalizes to unseen peers and task domains.
Direct memory readouts outperform majority voting and the best fixed peer over the full out-of-distribution evaluation set. Performance improves consistently as more correctness feedback becomes available, indicating that Σ-Mem progressively accumulates actionable reliability information.
The authors position Σ-Mem as a reusable foundation for adaptive coordination in LLM-based multi-agent systems, offering a new direction beyond content-based memory.