扎根智能体记忆:面向企业级智能体的环境探测式记忆管理
Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents
arXiv Identifier: arXiv:2609.11060 [cs.AI]
Primary Subject: Artificial Intelligence (cs.AI)
Secondary Subjects: Software Engineering (cs.SE)
Submission Date: September 10, 2026
Authors: Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz
文章背景与核心概要
随着持久化记忆系统逐步进入面向生产环境的 AI 智能体平台,长周期智能体得以在跨会话交互中持续积累经验。然而,传统的任务后记忆整理智能体 (Curator Agent) 往往仅局限于已完成的历史轨迹进行复盘,容易固化此前执行中的错误、过度泛化片面证据、或长期保留陈旧过时的无效知识。为此,本文提出了 环境探测式管理 (Environment-Probing Curation)——这是一种轻量且高度兼容生产部署的扩展方案,通过为异步运行的记忆管理智能体赋予最小权限的只读外部环境工具,使其能够主动校验、界定范围并实时刷新候选记忆,而完全无需重新训练底层模型、修改任务智能体架构或变更生产写权限。在基于 GitHub Copilot 真实 SDK 构建的测试环境中,该方法将数据库探索任务的通过率从 39% 提升至 73%,并在大幅削减工具调用量与推理开销的同时保持了出色的模式稳定性,为构建高可靠的企业级智能体记忆系统提供了开创性范式。
📋 核心概要
📋 Summary
持久化记忆系统使长周期 AI 智能体 (AI Agent) 能够在多个会话之间持续积累实操经验,但标准的任务后整理机制往往容易引入错误推论、过度泛化局部证据,或是保留陈旧过时的无效知识。本文提出了环境探测式管理 (Environment-Probing Curation),这是一种轻量且完美适配现有部署的扩展方案:它为异步运行的记忆整理智能体赋予遵循最小权限原则的只读外部工具,使其能够主动检验、界定范围并动态刷新候选记忆,而无需重新训练模型、无需改动任务智能体,也无需变更生产环境的写入权限。
Persistent memory systems enable long-horizon AI agents to accumulate experience across sessions, but standard post-task curation often introduces errors, overgeneralizes evidence, or retains stale knowledge. This paper introduces environment-probing curation, a lightweight, deployment-compatible extension that empowers asynchronous curator agents with least-privilege, read-only world tools. This allows the system to actively check, scope, and refresh memories without requiring model retraining, task-agent modifications, or changes to production write authorities.
在基于 GitHub Copilot (GHCP) SDK 构建的类生产测试环境(包含 CLBench 数据库探索与 90 项改编自 APEX 的管理咨询任务)中的评估表明,环境探测机制显著提升了任务通过率,大幅降低了查询开销与计算成本,并在杜绝架构漂移 (Schema Drift) 的前提下显著增加了任务智能体的奖励收益。
Evaluated on a production-like GitHub Copilot (GHCP) harness (using CLBench and 90 adapted APEX management-consulting tasks), environment probing significantly boosts pass rates, reduces query overhead and costs, and improves task-agent reward gains without schema drift.
📌 文档元数据
📌 Document Metadata
| 字段 | 详细信息 |
|---|---|
| arXiv 编号 | arXiv:2609.11060 [cs.AI] |
| 主要学科领域 | 人工智能 (cs.AI) |
| 次要学科领域 | 软件工程 (cs.SE) |
| 提交日期 | 2026 年 9 月 10 日 |
| 作者 | Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz |
| 论文链接 | 查看 PDF | 网页版本 | DOI |
Field Details arXiv Identifier arXiv:2609.11060 [cs.AI] Primary Subject Artificial Intelligence ( cs.AI)Secondary Subjects Software Engineering ( cs.SE)Submission Date September 10, 2026 Authors Susheel Suresh, Hazel Mak, Sahil Bhatnagar, Chhaya Methani, Alejandro Gutierrez Munoz Links View PDF | HTML Version | DOI
📄 论文摘要
📄 Abstract
💬 [原文引用 / Original Quote]: Persistent memory is entering production-oriented agent platforms to help long-horizon agents accumulate experience across sessions. Yet a post-task curator agent restricted to completed trajectories can preserve errors, overgeneralize partial evidence, or retain stale knowledge. We introduce environment-probing curation, a deployment-compatible extension that gives an existing asynchronous curator agent least-privilege, read-only world tools to check, scope, and refresh candidate memories. It requires no model retraining and leaves the task agent, retriever, memory representation, and production write authority unchanged.
In a production-like GitHub Copilot (GHCP) harness built on its SDK, we compare stateless execution, full in-context learning, GHCP + Mem, and GHCP + Mem (w/ Env Probing) on CLBench database exploration and 90 adapted APEX management-consulting tasks.
- On CLBench: Probing raises pass rates from \(39\%\) to \(73\%\) and pass-discounted rewards from \(8.60\) to \(22.60\), while reducing queries from \(8.8\) to \(4.7\) per question and task-agent costs from \(\$3.38\) to \(\$1.68\).
- Across APEX Worlds: All 18 memory-versus-baseline mean reward comparisons are positive, task-agent tool calls fall by \(16\text{--}75\%\), and probing delivers the best task-agent reward gain per dollar in five out of six worlds.
- Model Stability: Probing attains higher mean reward than standard GHCP + Mem on both Sonnet 4.6 and Opus 4.7 without schema drift.
Environment probing successfully transforms existing agent-memory curation into an environment-informed, auditable process while preserving a compact task-time interface.
持久化记忆正逐步迈入面向生产的智能体平台,以帮助长周期智能体在跨会话交互中沉淀宝贵经验。然而,如果仅将任务后整理智能体限制在已完成的历史轨迹中,就极易固化推理错误、过度泛化局部证据,甚至保留陈旧过时的信息。我们提出了环境探测式管理:一种面向实际部署的扩展方案,它赋予现有异步整理智能体最小权限的只读真实世界工具,以主动核验、限定作用域并刷新候选记忆。该方案完全不需要重新训练模型,且保持了任务智能体、检索器、记忆表征格式以及生产写入权限的原封不动。
在基于其实际 SDK 构建的类生产 GitHub Copilot (GHCP) 运行环境中,我们在 CLBench 数据库探索和 90 项改编的 APEX 管理咨询任务上,对比了无状态执行、完整上下文学习 (In-Context Learning)、GHCP + Mem 以及 GHCP + Mem(带环境探测)四种架构:
- 在 CLBench 上: 环境探测将通过率从 \(39\%\) 大幅提升至 \(73\%\),折扣奖励从 \(8.60\) 跃升至 \(22.60\),同时将每道题的查询量从 \(8.8\) 次缩减至 \(4.7\) 次,任务智能体开销从 \(\$3.38\) 降至 \(\$1.68\)。
- 在各 APEX 环境中: 所有 18 组记忆对比基线的平均奖励差异均为正值,任务智能体的工具调用次数减少了 \(16\text{--}75\%\),且在六个测试世界中的五个世界里实现了最高的“单位美元任务奖励收益”。
- 模型稳定性表现: 无论在 Sonnet 4.6 还是 Opus 4.7 模型上,带探测的环境管理均获得了高于标准 GHCP + Mem 的平均奖励,且没有发生任何架构漂移。
环境探测成功将现有的智能体记忆整理升级为兼具环境感知与可审计性的稳健流程,同时在任务运行时保持了紧凑高效的接口契约。
🔗 相关资源与工具
🔗 Additional Resources & Tools
- 代码、数据与媒体: 可通过 Hugging Face、CatalyzeX 代码检索 以及 DagsHub 查看获取情况。
- 交互式与文献检索工具: 可通过 Connected Papers、Litmaps、scite 智能引用 和 alphaXiv 进行深入探索。
- Code, Data & Media: Check availability via Hugging Face, CatalyzeX Code Finder, and DagsHub.
- Interactive & Bibliographic Tools: Explore via Connected Papers, Litmaps, scite Smart Citations, and alphaXiv.