記住關鍵時刻：長程代理的主動記憶代理

摘要

在長時程任務中，與決策相關的狀態往往散落在不斷擴展的軌跡中，而行動代理必須將其提取並加以執行。隨著軌跡增長，任務要求、環境事實、先前的嘗試、診斷結果以及未完成的子目標可能被埋沒在上下文視窗內，甚至被推至視窗外，從而在需要時無法影響決策。我們將這種失效模式稱為「行為狀態衰減」。本研究將記憶視為一種主動干預機制，而非被動檢索。一個獨立的記憶代理與未經修改的行動代理並行運行，從近期軌跡中更新結構化記憶庫，並決定是注入一則基於記憶的提醒還是保持沉默。該模組可即插即用於前沿行動代理與現有代理框架中。在 Terminal-Bench 2.0 與 τ^2-Bench 上，此方法分別提升了較弱與較強行動代理的 pass@1 指標，在 Terminal-Bench 上提高 8.3 個百分點，在 τ^2-Bench 上提高 6.8 個百分點。消融實驗顯示，選擇性干預優於被動記憶庫暴露、持續注入、僅顧問式指導以及一般檢索。作為朝向開放權重記憶策略的初步嘗試，我們使用監督式微調與群組相對策略最佳化在 SETA 上訓練 Qwen3.5-27B，提升了驗證獎勵，並在 Terminal-Bench 上實現了部分遷移。

English

In long-horizon tasks, decision-relevant state is often scattered across an expanding trajectory, while the action agent must surface it and act. As trajectories grow, task requirements, environment facts, prior attempts, diagnoses, and open subgoals can be buried in the context window or pushed beyond it, failing to influence decisions when needed. We call this failure mode "behavioral state decay". We study memory as an active intervention mechanism rather than passive retrieval. A separate memory agent runs alongside an unmodified action agent, updating a structured memory bank from the recent trajectory and deciding whether to inject a memory-grounded reminder or remain silent. The module is plug-and-play with frontier action agents and existing agent harnesses. Across Terminal-Bench 2.0 and τ^2-Bench, it improves pass@1 for both weaker and stronger action agents, with gains of +8.3 pp on Terminal-Bench and +6.8 pp on τ^2-Bench. Ablations show that selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval. As an early step toward open-weight memory policies, we train Qwen3.5-27B on SETA using SFT and GRPO, improving validation reward and achieving partial transfer to Terminal-Bench.