01 · Newer Is Not Truer: Three Presuppositions Behind the Recency Rule in Long-Term Agent Memory, and Why Retaining Both Records Relocates the Decision Rather Than Making It

更新的并非总是真实：长期智能体记忆中新近性规则背后的三项预设，以及为何保留双重记录只是将决策重新定位而非解决问题

Research record	Details
Authors	Pranay M. Mahendrakar
Published	2026-08-31
Sources	openalex
Focus	long-term agent memory, recency rule critique, write-time vs read-time decisions, provenance and temporal semantics 长期智能体记忆, 新近性规则批判, 写时与读时决策, 溯源与时间语义

English

The paper challenges a pervasive operational assumption about persistent agent memory and proposes concrete empirical questions and a reporting convention that are directly applicable to the thesis’s memory-update design and evaluation; a close, critical reading can expose hidden assumptions and help shape experiments.

中文

本文质疑关于持久化智能体记忆的普遍操作性假设，并提出可直接用于论文记忆更新设计与评估的具体实证问题和报告约定；仔细研读可揭示隐含预设并有助于实验设计。

Research synopsis

English

This theoretical paper examines the common default that, when persistent agent memory contains conflicting records, the newer record supersedes the older. The author identifies three presuppositions underlying that rule: (1) a key that determines which records are in conflict, (2) that recorded write order reflects when facts were true (transaction time vs valid time), and (3) that newer records are at least as trustworthy as older ones. The paper argues these presuppositions are under-justified, critiques the shift toward retaining both records as merely relocating the decision to read-time, and enumerates nine empirical studies and a reporting convention that would help resolve open questions. No experiments are reported.

中文

本文为一篇理论性讨论，审视在持久化智能体记忆中遇到冲突记录时普遍采用的新近记录优先规则。作者指出该规则基于三项预设： (1) 存在可识别冲突记录的键，(2) 写入顺序反映事实成立顺序（事务时间 vs 有效时间），(3) 新近记录至少与旧记录同样可信。文章认为这些预设证据不足，批评保留双重记录只是将决策转移到读时，并列举了九项可解决开放问题的实证研究和一项报告约定。文中不包含实验数据。

Thesis relevance

English

Overlap: the paper addresses fundamental policies for persistent agent memory, including provenance, temporal semantics, and whether to supersede or retain prior records—issues directly relevant to graph memory update rules and entity resolution policies in the thesis. Differences: it is a conceptual and literature-analytic argument without empirical evaluation and it does not propose or evaluate structured knowledge-graph construction, multi-hop retrieval, or local LLM-driven agentic RAG. Complementarity: its analysis of write-time vs read-time decisions and recommended empirical studies can inform the thesis’s update semantics, evaluation design, and reporting conventions.

中文

重叠点：本文探讨了持久化智能体记忆的基本策略，包括溯源、时间语义以及是覆盖还是保留先前记录——這些问题与论文中图记忆的更新规则和实体解析策略直接相关。差异：文章为概念性文献分析，缺乏实证评估，且未提出或评估基于知识图谱的构建、多跳检索或本地LLM驱动的智能体RAG方法。互补性：其关于写时与读时决策的分析及所建议的实证研究可为论文在更新语义、评估设计和报告规范上提供指导。

Writing and related work

English

Cite this paper when discussing memory-update policies, provenance, and assumptions behind supersession rules. Use its critique to motivate explicit statement of update semantics and to justify experiments that compare alternative retention strategies rather than implicitly assuming recency.

中文

在论述记忆更新策略、溯源以及覆盖规则的预设时引用此文。可用其批判性观点作为阐明更新语义的动机，并据此设计比较不同保留策略（而非默认新近优先）的实验证明。

Method and evaluation

English

Implement and compare three memory-update regimes in the thesis system: (a) newest-wins supersession, (b) retain-all with read-time adjudication, and (c) hybrid rules keyed by conflict identifier and valid/transaction time. Measure effects on graph coherence, bridge-entity fragmentation, multi-hop retrieval accuracy, and end-to-end latency. Instrument and report provenance metadata (write time, extraction confidence) so read-time policies can be objectively evaluated; follow the paper’s suggested studies and reporting convention as a checklist.

中文

在论文系统中实现并比较三类记忆更新策略：(a) 新近优先覆盖，(b) 全部保留并在读时裁决，(c) 基于冲突键与有效/事务时间的混合规则。度量这些策略对图一致性、桥实体碎片化、多跳检索准确性和端到端延迟的影响。记录并上报溯源元数据（写入时间、抽取置信度），以便对读时策略进行客观评估；可将该文提出的实证研究和报告约定作为实现检查表。

Future directions

English

Empirically operationalize the nine studies the paper proposes, applying them to the thesis’s evolving graph memory to quantify error accumulation and the trade-offs of retention policies. 2) Develop and evaluate read-time adjudication mechanisms (including provenance-aware ranking and LLM-judges) and compare their cost/precision trade-offs against write-time policies.

中文

将文章提出的九项研究实施为实证实验，应用于论文中的动态图记忆，以量化错误累积和保留策略的权衡。2) 设计并评估读时裁决机制（包括溯源感知排序和LLM裁判），并将其成本/精度权衡与写时策略进行比较。

---