Σ-Mem：用於基於LLM之多智能體系統的線上可靠性記憶

摘要

記憶對於長時程的大型語言模型代理（LLM agent）至關重要，然而現有的記憶系統主要保存互動內容，而非建模哪些代理值得信賴及其信賴條件。此限制在多代理系統中尤為重要，因為中央模型可能無法直接驗證看似合理或相互關聯的同儕回應。我們提出Σ-Mem，一種線上可靠性記憶，它記錄個別同儕的歷史能力證據，以及跨同儕集合的同儕關係證據。兩種類型的證據皆以實對稱狀態（real symmetric state）維護，並從決策後的正確性回饋中更新。根據Weyl不等式，每次事件級更新所導致的頻譜變化是有界的，使得系統無需重新訓練底層模型即可進行穩定的線上適應。Σ-Mem提供通用的寫入與讀取介面：同一記憶可用於中央模型的殘差引導（residual steering）、無回應同儕路由（response-free peer routing），或可靠性加權投票（reliability-weighted voting）。在五個Qwen系列模型中，Σ-Mem能適應反事實可靠性偏移（counterfactual reliability shift），並泛化至未見過的同儕與任務領域。直接記憶讀取在完整的分布外（OOD）評估集上也優於多數決投票與最佳固定同儕。此外，隨著更多正確性回饋的取得，效能持續提升，表明Σ-Mem能逐步累積可操作的可靠性資訊。這些結果確立了可靠性記憶作為基於LLM的多代理系統中適應性協調的可重用基礎。

English

Memory is central to long-horizon LLM agents, yet existing memory systems primarily preserve interaction content rather than modeling which agents can be trusted and under what conditions. This limitation is particularly important in multi-agent systems, where a central model may be unable to directly verify plausible or correlated peer responses. We introduce Σ-Mem, an online reliability memory that records historical competence evidence for individual peers and peer relationship evidence across the peer set. Both forms of evidence are maintained as real symmetric states and updated from post-decision correctness feedback. By Weyl's inequality, the spectral change caused by each event-level update is bounded, enabling stable online adaptation without retraining the underlying models. Σ-Mem provides a general write-and-read interface: the same memory can be used for residual steering of a central model, response-free peer routing, or reliability-weighted voting. Across five Qwen-family models, Σ-Mem adapts to counterfactual reliability shifts and generalizes to unseen peers and task domains. Direct memory readouts also outperform majority voting and the best fixed peer over the full OOD evaluation set. Moreover, performance improves consistently as more correctness feedback becomes available, indicating that Σ-Mem progressively accumulates actionable reliability information. These results establish reliability memory as a reusable foundation for adaptive coordination in LLM-based multi-agent systems.