跳转至

文章背景与核心概要

在固定内存预算下运行的智能体记忆系统，其性能通常分为两个阶段：保持（Retention）和检索（Retrieval）。传统的以检索为中心的模型隐式地假设必要证据总能在内存驱逐过程中存活下来。然而，本文挑战了这一假设，识别并隔离了一种名为“结构性间接前置条件驱逐（Structurally Indirect Prerequisite Eviction）”的检索前失效模式。在预算限制下，那些与给定查询仅有弱对齐的上游内存块会被过早丢弃，导致后续检索在开始之前就已宣告失败。

为了解决这一问题，作者引入了：1. 该失效模式的操作性定义；2. 一个可复现的确定性基准测试以及每个随机种子的追踪诊断；3. 依赖感知语义垃圾回收（Dependency-aware Semantic Garbage Collection, DSGC），这是一种旨在保护关键前置条件的单跳图感知规则。实验结果表明，DSGC 显著提升了全链条保持率，并在鲁棒性检查中揭示了单跳规则有效与退化的预算区间和缩放阈值。

When Retrieval Fails Before It Begins: Structurally Indirect Prerequisite Eviction as a Retention Failure in Agentic Memory

Authors: Minkyu Song
Published: July 5, 2026 (Accepted at the ICML 2026 Workshop on Failure Modes of Agentic AI, FAGEN@ICML 2026)
Identifiers: arXiv:2608.20400 [cs.AI] | DOI: 10.48550/arXiv.2608.20400
Resources: Code Repository

摘要与总结

在固定内存预算下运行的智能体记忆系统通常包含两个阶段：保持和检索。现有的以检索为中心范式隐式地假设必要证据能够在驱逐过程中存活，但我们通过隔离一种检索前的失效模式挑战了这一假设：结构性间接前置条件驱逐。在这种模式下，与查询弱对齐的上游块在预算压力下被丢弃。我们提供了该失效的操作性定义、可复现的确定性基准测试以及按种子划分的追踪诊断。最后，我们评估了依赖感知语义垃圾回收（DSGC），这是一种单跳图感知规则。在我们的主测试集中，在词汇编码器下，DSGC 将全链条保持率从 0.03 提升至 0.90，在句子编码器下从 0.23 提升至 1.00。鲁棒性检查随后识别了单跳规则保持有效或发生退化的预算和缩放机制。我们发布的管道和失效事后分析支持了将检索前的保持作为独特失效边界的机理分析。

Summary

In agentic memory systems operating under a fixed memory budget, performance is typically divided into two stages: retention and retrieval. Traditional retrieval-centric models implicitly assume that necessary evidence will always survive the memory eviction process.

This paper challenges that assumption by identifying and isolating a pre-retrieval failure mode termed Structurally Indirect Prerequisite Eviction. Under budget constraints, upstream memory blocks that are only weakly aligned with a given query are prematurely discarded, rendering subsequent retrieval impossible before it even begins.

To address this, the author introduces: 1. An operational definition of the failure mode. 2. A reproducible, deterministic benchmark along with per-seed trace diagnostics. 3. Dependency-aware Semantic Garbage Collection (DSGC), a one-hop graph-aware rule designed to protect critical prerequisites.

Key Results

  • Performance Gains: DSGC dramatically improves full-chain retention rates—jumping from 0.03 to 0.90 using a lexical encoder, and from 0.23 to 1.00 using a sentence encoder.
  • Robustness Boundaries: Subsequent robustness evaluations reveal the specific budget regimes and scaling thresholds where the one-hop rule holds strong versus where it degrades.

摘要

Agentic memory under a fixed budget involves two stages: retention and retrieval. Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: structurally indirect prerequisite eviction, in which upstream blocks weakly aligned with the query are discarded under budget pressure. We provide an operational definition of this failure, a reproducible deterministic benchmark, and per-seed trace diagnostics. Finally, we evaluate Dependency-aware Semantic Garbage Collection (DSGC), a one-hop graph-aware rule. In our main suite, DSGC improves full-chain retention from 0.03 to 0.90 under a lexical encoder and from 0.23 to 1.00 under a sentence encoder. Robustness checks then identify the budget and scaling regimes where the one-hop rule holds or degrades. Our released pipeline and failure postmortem support mechanistic analysis of retention before retrieval as a distinct failure boundary.

Abstract

Agentic memory under a fixed budget involves two stages: retention and retrieval. Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: structurally indirect prerequisite eviction, in which upstream blocks weakly aligned with the query are discarded under budget pressure. We provide an operational definition of this failure, a reproducible deterministic benchmark, and per-seed trace diagnostics. Finally, we evaluate Dependency-aware Semantic Garbage Collection (DSGC), a one-hop graph-aware rule. In our main suite, DSGC improves full-chain retention from 0.03 to 0.90 under a lexical encoder and from 0.23 to 1.00 under a sentence encoder. Robustness checks then identify the budget and scaling regimes where the one-hop rule holds or degrades. Our released pipeline and failure postmortem support mechanistic analysis of retention before retrieval as a distinct failure boundary.

元数据与分类

Metadata & Classification