跳转至

文章背景与核心概要

旨在成为长期陪伴者的对话智能体必须具备跨交互的持久记忆能力。然而，传统方法面临严峻挑战：简单地通过原始检索扩展上下文窗口会降低推理质量，而通过标准强化学习训练记忆智能体，则会在多阶段流水线中产生严重的信用分配瓶颈。

为了克服这些局限，作者提出了 SALIMORY 框架。该框架训练单一语言模型来管理一个认知结构化的记忆系统，涵盖用户事实、偏好和工作记忆。通过引入分层阶段式过程奖励和奖励分解对比细化机制，SALIMORY 为选择性过滤、整合和线索驱动召回等不同的记忆操作提供了独立的端到端监督。实验结果表明，SALIMORY 将记忆相关的故障减少了三分之一，端到端准确率比现有最先进模型提高了 10% 以上，并将“良好个性化”率提升了一倍以上。

SaliMory：为对话智能体构建认知记忆系统 (SaliMory: Orchestrating Cognitive Memory for Conversational Agents)

摘要 (Summary)

旨在作为长期陪伴者的对话智能体需要跨越所有交互的持久记忆。然而，传统方法面临着重大挑战：简单地通过原始检索扩展上下文窗口会降低推理质量，而通过标准强化学习训练记忆智能体则会在多阶段流水线中造成严重的信用分配瓶颈。

Conversational agents designed to act as lifelong companions require persistent memory across all interactions. However, traditional approaches face significant challenges: simply expanding context windows with raw retrieval degrades reasoning quality, while training memory agents via standard reinforcement learning creates a severe credit assignment bottleneck within multi-stage pipelines.

为了克服这些限制，作者引入了 SALIMORY，这是一个新颖的框架，它训练单一语言模型来管理一个认知结构化的记忆系统，该系统包含用户事实、偏好和工作记忆。通过采用分层阶段式过程奖励和奖励分解对比细化，SALIMORY 为不同的记忆操作——包括选择性过滤、整合和线索驱动召回——提供了隔离的端到端监督。实证结果表明，SALIMORY 将记忆相关的故障减少了三分之一，端到端准确率超过现有最先进模型 10% 以上，并将“良好个性化”率提高了一倍多。

To overcome these limitations, the authors introduce SALIMORY, a novel framework that trains a single language model to manage a cognitively structured memory system encompassing user facts, preferences, and working memory. By employing a hierarchical stage-wise process reward and reward-decomposed contrastive refinement, SALIMORY provides isolated end-to-end supervision for distinct memory operations—including selective filtering, consolidation, and cue-driven recall. Empirical results show that SALIMORY reduces memory-attributed failures by one-third, outperforms state-of-the-art models by over 10% in end-to-end accuracy, and more than doubles the "Good Personalization" rate.

文档元数据 (Document Metadata)

字段 详情
arXiv 标识符 arXiv:2606.04120 [cs.CL]
学科 计算与语言 (cs.CL)；人工智能 (cs.AI)
提交日期 2026年6月2日 (v1)；最后修订于 2026年8月24日 (v2)
DOI 10.48550/arXiv.2606.04120
Field Details
arXiv Identifier arXiv:2606.04120 [cs.CL]
Subjects Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Submitted on June 2, 2026 (v1); Last revised August 24, 2026 (v2)
DOI 10.48550/arXiv.2606.04120

作者 (Authors)

  • Kai Zhang
  • Xinyuan Zhang
  • Hongda Jiang
  • Shiun-Zu Kuo
  • Hyokun Yun
  • Ejaz Ahmed
  • Shereen Oraby
  • Ziyun Li
  • Sanat Sharma
  • Ann Lee
  • Ahmed A. Aly
  • Anuj Kumar
  • Raffay Hamid
  • Xin Luna Dong

摘要 (Abstract)

Conversational agents that serve as lifelong companions must maintain persistent memory across all interactions. However, simply expanding context windows with raw retrieval degrades reasoning quality, while training memory agents via standard reinforcement learning creates a severe credit assignment bottleneck in a multi-stage pipeline. To solve this, we introduce SALIMORY, a framework that trains a single language model to manage a cognitively-structured memory-spanning user facts, preferences, and working memory. By introducing a hierarchical stage-wise process reward and reward-decomposed contrastive refinement, SALIMORY provides isolated supervision for distinct memory operations (selective filtering, consolidation, and cue-driven recall) end-to-end. SALIMORY cuts memory-attributed failures by one-third, outperforms the state-of-the-art by over 10% in end-to-end accuracy, and more than doubles the Good Personalization rate.

全文及访问链接 (Full-Text & Access Links)

外部参考与工具 (External References & Tools)