01 · Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory

在有预算的继承式智能体记忆验证中计划指针与记录指令形式

Research record	Details
Authors	Kazuki Nakayashiki
Published	2026-09-03
Sources	openalex
Focus	inherited agent memory, budgeted verification, record-directive form, retrieval steering, registered experiments 继承式智能体记忆、有预算的验证、记录指令形式、检索引导、注册实验

English

The paper offers empirically robust, reproducible findings about how short stored directives steer memory retrieval under strict budgets—useful for sections on memory-access policies and evaluation methodology. It does not address structured graph memory, multi-hop retrieval mechanics, or entity resolution, so a focused skim is sufficient unless the thesis will evaluate directive-form effects in depth.

中文

该文提供了关于短指令在有预算限制下如何引导记忆检索的可复现实证结果——对讨论记忆访问策略与评估方法的章节有参考价值。但其并未涉及结构化图记忆、多跳检索机制或实体解析，除非论文要深入评估指令形式效应，否则以有针对性的略读为宜。

Research synopsis

English

The paper empirically studies how short stored directives in an inherited agent memory influence which archived record an agent retrieves when only one archived source is allowed before action. Across twelve registered studies (14,760 attempts) on multiple model families, the author compares three directive forms (pointer, criterion, or both) and measures their effects on retrieval choice under strict budgets. Results are presented as descriptive effects on several commercial and research models, with extensive archival transparency (code, data, preregistration). The manuscript emphasizes empirical description and reproducibility rather than proposing mechanistic explanations.

中文

该论文实证研究了在继承式智能体记忆中，当智能体在行动前仅被允许拉取一个归档来源记录时，存储的短指令如何影响其检索到哪条记录。作者在十二项注册研究（14,760 次尝试）中比较了三种指令形式（指针、判准或两者）并测量它们在严格预算下对检索选择的影响。结果以描述性效应呈现，涵盖多个模型族，并提供了完整的归档与可重现材料（代码、数据、注册）。手稿侧重于实证描述与可复现性，而非机制性解释。

Thesis relevance

English

Overlap: the paper studies how store-encoded directives steer memory retrieval under strict budgets, which is directly relevant to the thesis’ concern with how memory-access policies affect which evidence is reused. Differences/limitations: it does not construct or evaluate a persistent structured graph memory, perform multi-hop retrieval, or address entity resolution or KG assembly. Complementarity: its rigorous, registered experimental design and archival practices offer a model for evaluating retrieval-steering interventions in a graph memory system and for measuring per-model variability in memory access.

中文

重合点：本文研究存储指令在有预算约束下如何引导记忆检索，这与论文关注的记忆访问策略如何影响可复用证据直接相关。差异/局限：该工作并未构建或评估持久的结构化图记忆，也未进行多跳检索或处理实体解析与知识图谱组装问题。互补性：其严格的注册实验设计与完整归档实践可作为在图记忆系统中评估检索引导干预、以及衡量不同模型间记忆访问差异的范式参考。

Writing and related work

English

Cite this paper when discussing memory-access directives, verification budgets, and empirical variability across model providers. Note the paper’s descriptive scope and strong reproducibility; avoid treating its observed effects as mechanistic explanations without further study.

中文

在讨论记忆访问指令、有预算的验证和跨模型供应商的实证差异时应引用此文。注意其结果为描述性且具有强可复现性；在未进一步研究机制前，不应将这些效应视为机理性解释。

Method and evaluation

English

Replicate the paper’s registered, high-integrity experimental protocol when testing how different retrieval signals (pointer, criterion, combined) affect node/edge selection in a constructed graph memory. Measure downstream impacts on multi-hop QA, bridge-entity identification, and end-to-end latency, and report per-model variability; archive runs and analysis for reproducibility. Consider adding controlled budgets (number of retrieved records/credits) and ablations that combine directive form with embedding-similarity or LLM-based entity-resolution.

中文

在测试不同检索信号（指针、判准、二者合并）如何影响构建中图记忆的节点/边选择时，可复用该文的注册高完整性实验方案。测量对下游多跳问答、桥接实体识别与端到端延迟的影响，并报告按模型划分的差异；归档运行与分析以保证可复现性。可考虑引入受控预算（可检索记录数/学分）与消融实验，将指令形式与嵌入相似度或基于LLM的实体解析结合评估。

Future directions

English

Evaluate how directive forms interact with structured graph memory retrieval policies: test whether pointers or criteria improve reuse of intermediate entities and multi-hop path traversal. Investigate combining directive signals with embedding-based and LLM-based entity resolution, and measure effects on QA accuracy and verification under different budget regimes.

中文

评估指令形式与结构化图记忆检索策略的交互：检验指针或判准是否有助于重用中间实体和多跳路径遍历。研究将指令信号与基于嵌入和基于LLM的实体解析结合，并在不同预算机制下衡量对问答准确性与验证性的影响。

---