Causal Memory Policy: Making Memory Utility Identifiable by Intervening on Retrieval

📄 arXiv: 2610.02070v1 📥 PDF

作者: Arman Behnam, Binghui Wang

分类: cs.AI

发布日期: 2026-10-01


💡 一句话要点

提出因果记忆策略以解决记忆效用识别问题

🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)

关键词: 因果推断 记忆增强 语言模型 效用识别 机器学习

📋 核心要点

  1. 现有方法在记忆效用识别上存在不足,特别是未被检索的记忆效用无法评估。
  2. 论文提出因果记忆策略(CMP),通过对检索进行干预来识别记忆效用,采用自归一化逆倾向加权方法。
  3. 实验结果显示,CMP显著提高了必需与非必需记忆的区分度,AUC从0.54提升至0.66。

📝 摘要(中文)

增强记忆的大型语言模型必须决定保留哪些记忆,现有系统通过估计每个记忆对任务性能的影响来实现。然而,这些估计完全依赖于已检索的记忆。当记忆从未被检索时,存储级干预产生相同的结果,导致其效用无法识别。为此,本文提出因果记忆策略(CMP),通过对检索本身进行干预来恢复识别能力,并为具有已知倾向的记忆保留固定数量的上下文槽。CMP通过自归一化逆倾向加权在平衡分配设计下估计记忆效用。我们证明了通过检索的记忆效用的因果分解、估计器的无偏性和精确方差,以及不可逆操作下的最优决策规则。实验证明,在LongMemEval上54%的必需记忆和在LoCoMo上67%的必需记忆识别失败,CMP将必需与非必需记忆的区分度从0.54提升至0.66 AUC。

🔬 方法详解

问题定义:本文旨在解决增强记忆的语言模型中,未被检索的记忆效用无法识别的问题。现有方法仅依赖于已检索的记忆,导致存储级干预无法有效评估所有记忆的效用。

核心思路:提出因果记忆策略(CMP),通过对检索过程进行干预,保留具有已知倾向的记忆,从而恢复记忆效用的识别能力。此设计旨在克服现有方法的局限性,使得未检索记忆的效用也能被评估。

技术框架:CMP的整体架构包括记忆存储、检索干预和效用估计三个主要模块。首先,通过干预检索过程来选择记忆,然后利用自归一化逆倾向加权方法进行效用估计。

关键创新:CMP的核心创新在于通过干预检索而非仅依赖存储级干预来识别记忆效用。这一方法使得未检索记忆的效用得以评估,克服了传统方法的局限。

关键设计:CMP采用固定数量的上下文槽来存储记忆,并使用自归一化逆倾向加权来进行效用估计。设计中还考虑了平衡分配,以确保估计的无偏性和精确性。具体的损失函数和网络结构细节在论文中进行了详细描述。

🖼️ 关键图片

img_0
img_1
img_2

📊 实验亮点

实验结果表明,CMP在LongMemEval和LoCoMo数据集上分别识别失败率为54%和67%。通过CMP,必需与非必需记忆的区分度显著提高,AUC从0.54提升至0.66,展示了其在记忆管理中的有效性。

🎯 应用场景

该研究的潜在应用领域包括自然语言处理、智能助手和推荐系统等。通过提高记忆效用的识别能力,CMP可以帮助这些系统更好地管理和利用记忆,从而提升任务性能和用户体验。未来,CMP的框架可能会被扩展到其他领域,如强化学习和决策支持系统。

📄 摘要(原文)

Memory-augmented large language models must decide which memories to retain, and recent systems do so by estimating each memory's effect on task performance. However, these estimates rely entirely on retrieved memories. When a memory is never retrieved, store-level interventions produce identical outcomes, leaving its utility unidentified. This is a retrieval-level positivity violation, invisible to diagnostics that examine only memory operations. We introduce Causal Memory Policy (CMP), a causal framework that restores identification by intervening on retrieval itself, reserving a fixed number of context slots for memories sampled with known propensities. CMP estimates memory utility by self-normalized inverse propensity weighting under a balanced assignment design. We prove the causal factorization of memory utility through retrieval, the unbiasedness and exact variance of the estimator, and the optimal decision rule under irreversible operations. Empirically, identification fails for 54% of required memories on LongMemEval and 67% on LoCoMo, and the failure persists in a deployed memory system. CMP improves discrimination between required and non-required memories from 0.54 to 0.66 AUC. Finally, we show that identified memory utility alone is insufficient for retention decisions: per-query utility reaches 0.78 AUC on the query for which it is estimated, yet no aggregation available to a retention policy predicts a memory's value on unseen queries. Code is available at: https://anonymous.4open.science/r/cmp-release-D0C3/.