01 · Corpus-Guided Dual-Path Propagation for Graph Retrieval-Augmented Generation

语料库引导的双路径传播用于图检索增强生成

Research record	Details
Authors	Baoxian Liu, Tong Wei
Published	2026-09-29
Sources	openalex
Focus	graph-based RAG, multi-hop retrieval, entity neighborhood propagation, relation-free graph retrieval 基于图的RAG, 多跳检索, 实体邻域传播, 无关系图检索

English

The paper presents a concrete neighborhood-construction and dual-path propagation scheme that could be adapted to the thesis’s incremental graph memory; its empirical evidence-recall improvements make it a potentially useful methodological contribution to integrate or compare against.

中文

该论文提出了具体的实体邻域构建与双路径传播方法，可适配到论文的增量图记忆中；其在证据召回上的实证提升使其成为值得深入阅读以整合或比较的方法性贡献。

Research synopsis

English

The paper addresses limitations of relation-free graph retrieval for multi-hop question answering, where relying on query-sentence similarity can miss bridging evidence or activate incidental entities. It proposes NexusRAG, which augments a relation-free Tri-Graph with a corpus-level entity neighborhood built from joint entity co-occurrence and semantic similarity. NexusRAG runs two complementary propagation paths: neighborhood-constrained semantic propagation through sentences to identify a query-relevant entity frontier, and direct structural propagation among neighboring entities to expand that frontier. Propagated entity weights are used to initialize passages for Personalized PageRank. The authors report consistent gains on three multi-hop QA benchmarks and a GraphRAG-Bench subset, with evidence-recall improvements of 4.2–8.1 points on the subset.

中文

本文针对基于关系空白（relation-free）图检索在多跳问答中的不足，提出 NexusRAG。作者通过联合实体共现与语义相似度构建语料库级别的实体邻域结构，以增强基于 Tri-Graph 的检索。NexusRAG 采用双路径传播：通过句子进行的邻域约束语义传播以识别与查询相关的实体前沿，以及在邻近实体间的结构性直接传播以扩展该前沿。传播得到的实体权重用于为 Personalized PageRank 初始化段落。作者在三项多跳 QA 基准和 GraphRAG-Bench 的一个子集上报告了稳定提升，子集上的证据召回提高 4.2–8.1 点。

Thesis relevance

English

Overlap: both papers work on graph-based RAG and improving multi-hop retrieval by leveraging entity relations beyond plain sentence similarity. Differences: NexusRAG builds a corpus-level, relation-free neighborhood graph for offline propagation, while the thesis studies an incrementally assembled, persistent graph memory created from LLM-driven agent interactions and explicit subject-predicate-object triples. Limitations: NexusRAG does not address agentic, incremental memory construction, entity-resolution strategies for merging surface forms, or integration with local LLM agent paradigms. Complementarity: NexusRAG’s neighborhood construction and dual-path propagation mechanisms could be adapted as a propagation and edge-weighting component within the thesis’s evolving graph memory.

中文

重叠点：两者都研究基于图的 RAG，并尝试通过利用实体间关系超越纯句子相似度来改善多跳检索。差异：NexusRAG 构建的是基于语料库的、无关系（relation-free）实体邻域图并在其上离线传播，而本论文关注的是由智能体交互和 LLM 提取的主谓宾三元组增量构建的持久图记忆。局限性：NexusRAG 未涉及智能体化、增量记忆构建、用于合并不同指称形式的实体解析策略，也未讨论与本地 LLM 智能体范式的集成。互补性：NexusRAG 的实体邻域构建和双路径传播方法可作为本论文中演化图记忆的传播机制与边权设定策略进行借鉴。

Writing and related work

English

Position NexusRAG as a pragmatic, corpus-level method that mitigates query-similarity blind spots by combining co-occurrence and semantic similarity; emphasize its relation-free design when comparing to KG-based or triple-centric approaches. When citing, contrast offline corpus graph construction with online, agent-driven graph memory to clarify scope.

中文

将 NexusRAG 描述为一种务实的语料库级方法，通过结合共现与语义相似度来缓解仅依赖查询相似度的盲点；在与基于知识图谱或三元组的做法比较时强调其无关系（relation-free）设计。引用时应对比离线语料图构建与在线智能体驱动的图记忆以明确适用范围。

Method and evaluation

English

Consider adapting NexusRAG’s corpus-level neighborhood (co-occurrence + semantic similarity) to initialize or augment the thesis’s incremental graph edges and weights, then measure how that affects multi-hop retrieval and bridge-entity identification on MultiHop-RAG. Use NexusRAG-style propagated entity weights as priors for path-traversal algorithms (e.g., Personalized PageRank) within the evolving graph memory. Evaluate trade-offs in evidence recall, bridge-entity precision, reasoning-step count and end-to-end latency, and compare against the thesis’s proposed embedding-based and LLM-as-judge entity-resolution strategies.

中文

可考虑将 NexusRAG 的语料库级实体邻域（共现 + 语义相似度）用作增量图的初始边与权重，并衡量其对 MultiHop-RAG 上多跳检索及桥接实体识别的影响。将 NexusRAG 的传播得到的实体权重作为演化图记忆中路径遍历算法（例如 Personalized PageRank）的先验。评估证据召回、桥接实体精确度、推理步数和端到端延迟的权衡，并与论文中提出的基于嵌入和 LLM 评判的实体解析策略做比较。

Future directions

English

Extend NexusRAG concepts to agentic, incremental graph memory by applying the neighborhood construction and dual-path propagation to graphs assembled from LLM-extracted triples, and incorporate per-edge confidence scores. Experiment with integrating these propagation priors into different agent paradigms (ReAct, Plan-and-Execute, Self-Ask, Reflexion) and measure whether they improve reuse of prior reasoning paths and multi-hop QA accuracy.

中文

将 NexusRAG 的思想扩展到智能体化的增量图记忆：在由 LLM 提取三元组组装的图上应用实体邻域构建与双路径传播，并引入每条边的置信度分数。尝试将这些传播先验整合到不同智能体范式（ReAct、Plan-and-Execute、Self-Ask、Reflexion）中，评估其是否提高先前推理路径的复用率和多跳问答准确性。

---

02 · Capsule: Atomic, File-First Long-Term Memory with Bounded Drift for Autonomous AI Agents

Capsule：面向自治人工智能智能体的原子化、文件优先长期记忆与有界漂移

Research record	Details
Authors	Vikas Budde
Published	2026-09-30
Sources	openalex
Focus	long-term agent memory, file-first memory architecture, deduplication and context packing, hybrid retrieval indexes 长期智能体记忆, 文件优先记忆架构, 去重与上下文打包, 混合检索索引

English

High relevance: Capsule directly addresses the same operational failure modes (fragmentation, bloat, verifiability) and proposes concrete systems techniques (deduplication, token-budget packing, hybrid indexes) that can be applied or adapted to a graph-memory pipeline. The reported empirical evaluations and ablations are likely to contain useful procedures and metrics for the thesis.

中文

高度相关：Capsule 直接针对相同的操作失败模式（碎片化、膨胀、可验证性）提出具体系统技术（去重、令牌预算打包、混合索引），这些可应用或改编到图记忆管道中。其报告的实证评估与消融分析可能包含对论文有用的方法和度量。

Research synopsis

English

The paper identifies three failure modes of conventional Retrieval-Augmented Generation (RAG) for long-running autonomous LLM agents: semantic fragmentation from fixed-width chunking, append-only vector-store bloat, and opacity of embedding-only stores. It proposes Capsule, a dual-plane, file-first memory design: an auditable Storage Plane of atomic Markdown facts with YAML frontmatter stored on the filesystem, and a rebuilt, ephemeral Retrieval Plane (SQLite FTS5 / PostgreSQL tsvector plus vector embeddings). Key mechanisms include SHA-256 content-hash deduplication and a token-bounded knapsack optimizer for context composition. The authors report empirical improvements in prompt token efficiency, multi-hop fact recall, downstream QA accuracy with a frontier LLM, bounded long-term growth, and ablation results attributing gains to hybrid retrieval, packing, and deduplication.

中文

论文指出传统基于检索增强生成（RAG）的长期自治 LLM 智能体存在三类问题：固定宽度切分导致语义碎片化、追加式向量存储膨胀，以及仅嵌入存储的不可审计性。作者提出 Capsule：一种双平面、文件优先的长期记忆设计——可审计的存储平面由带 YAML frontmatter 的原子 Markdown 事实文件组成，检索平面为可重建的临时索引（SQLite FTS5 / PostgreSQL tsvector 与向量嵌入）。核心机制包括 SHA-256 内容哈希去重和基于令牌预算的“背包”上下文打包器。文中报告了在提示令牌效率、多跳事实召回、使用前沿 LLM 的下游问答准确性、长期增长有界性以及消融分析方面的改进，且将收益归因于混合检索、打包和去重组件。

Thesis relevance

English

Overlap: both the thesis and Capsule address long-term, persistent agent memory, semantic fragmentation and memory bloat, and the need for auditable, corrigible memory representations. Difference: Capsule focuses on file-first atomic facts, deduplication, and token-budgeted context packing with hybrid full-text+vector indexes, whereas the thesis centers on incrementally assembled Knowledge Graph (KG) graph memory, triple extraction, entity resolution, and multi-hop path traversal. Limitation: Capsule’s abstract does not describe structured graph construction, typed relations, or explicit entity-resolution strategies targeted at preserving multi-hop paths. Complementarity: Capsule’s deduplication, revision lineage, and hybrid indexing could be combined with the thesis’s KG extraction and graph-node canonicalization to reduce fragmentation and index bloat while improving verifiability.

中文

重合点：论文与本论文均关注长期、持续的智能体记忆，解决语义碎片化与记忆膨胀问题，并强调可审计、可纠正的记忆表示。差异：Capsule 侧重于文件优先的原子事实、去重，以及基于令牌预算的上下文打包和混合全文+向量索引；而本论文聚焦于增量构建的知识图谱（KG）图记忆、三元组抽取、实体解析与多跳路径遍历。局限性：Capsule 摘要未描述结构化图构建、有类型关系或专门用于保持多跳路径的实体解析策略。互补性：Capsule 的去重、修订谱系记录和混合索引可与论文中的 KG 抽取与图节点规范化结合，以减少碎片化与索引膨胀并改进可验证性。

Writing and related work

English

Position Capsule as a systems-level alternative to append-only vector stores that emphasizes auditability and budgeted prompt composition; highlight that its practical engineering choices (filesystem facts, content-hash deduplication, hybrid indexes) address operational issues complementary to structured graph memory. When citing, avoid implying Capsule provides KG-style typed relations or entity-resolution tailored to multi-hop path reuse unless validated by full text.

中文

将 Capsule 定位为对追加式向量库的系统级替代方案，强调其可审计性和预算化提示构成；指出其实用工程选择（文件系统事实、内容哈希去重、混合索引）在操作层面上与结构化图记忆互补。引用时应避免在未经过全文验证的情况下，暗示 Capsule 已提供 KG 风格的有类型关系或专门用于多跳路径复用的实体解析。

Method and evaluation

English

Consider adopting Capsule’s normalized content-hash deduplication to merge candidate KG nodes and to maintain revision lineages for graph edits. Evaluate a token-bounded knapsack context composer as an alternative to top-k chunk concatenation when assembling prompt context for multi-hop traversal; measure its effect on prompt token cost and multi-hop recall. Use a hybrid retrieval plane (FTS + vector embeddings) to support both surface-form matching for entity resolution and semantic retrieval for fuzzy matches. Reproduce Capsule’s memory-bloat simulation and component ablations to quantify long-run growth and the contribution of deduplication versus structured merging.

中文

可考虑采纳 Capsule 的规范化内容哈希去重，用于合并候选 KG 节点并维护图编辑的修订谱系。评估基于令牌预算的背包式上下文组合器，作为组装用于多跳遍历提示上下文的 top-k 拼接的替代方案，并测量其对提示令牌成本与多跳召回的影响。采用混合检索平面（全文搜索 + 向量嵌入）以同时支持实体解析所需的表面形式匹配与语义检索的模糊匹配。复现 Capsule 的记忆膨胀仿真与组件消融，以量化长期增长以及去重与结构化合并的相对贡献。

Future directions

English

Integrate Capsule’s file-first atomic facts and content-hash deduplication into the thesis’s incremental KG pipeline to test whether file-level canonicalization improves entity resolution and path coherence. Extend Capsule’s hybrid retrieval indexes to index extracted triples and relation types for efficient graph-aware retrieval. Empirically compare knapsack context packing versus graph-path-driven context assembly on MultiHop-RAG or similar multi-hop benchmarks.

中文

将 Capsule 的文件优先原子事实与内容哈希去重集成到论文的增量 KG 管道中，以测试文件级规范化是否能改进实体解析与路径连贯性。将 Capsule 的混合检索索引扩展为对抽取的三元组和关系类型建立索引，以支持高效的图感知检索。在 MultiHop-RAG 或类似多跳基准上，实证比较背包式上下文打包与基于图路径驱动的上下文组装的效果。

---

03 · A Knowledge Graph-Driven Smart System for Incentive-Based Work Productivity Support

一种基于知识图谱驱动的激励型工作效率支持智能系统

Research record	Details
Authors	Theodora Stamoglou, Asimina Dimara, Ioannis Tzitzios, Nikolaos Kladovasilakis, Konstantinos G. Fouskas, Christos‐Nikolaos Anagnostopoulos
Published	2026-09-30
Sources	openalex
Focus	knowledge graphs, GraphRAG, personalized interventions, explainable recommendations 知识图谱（KG）, GraphRAG, 个性化干预, 可解释性推荐

English

Relevant for KG-grounding, rule-based enrichment, and user-centered evaluation metrics, but domain-specific (productivity) and small-scale; it does not address the thesis’ core technical questions about agentic graph memory, entity-resolution trade-offs, or multi-hop QA performance, so a skim suffices to extract applicable methods and metrics.

中文

在 KG 支撑、基于规则的图增强和面向用户的评估指标方面具有参考价值，但属于特定领域（生产力）且规模较小；未涉及论文的核心技术问题（智能体图记忆、实体解析权衡或多跳问答性能），因此查阅要点即可。

Research synopsis

English

This paper introduces ADAPTS, a Knowledge Graph (KG)-driven smart system for incentive-based workplace productivity support. The framework fuses indoor environmental sensors, wearable-derived physiological and behavioural indicators, and questionnaire profiles into an RDF/OWL Knowledge Graph. A rule-based reasoning layer derives high-level semantics and a Daily Productivity Score (DPS). The enriched graph is accessed via a Graph Retrieval-Augmented Generation (GraphRAG) mechanism to build a Unified User Context that grounds an LLM for personalized interventions and adaptive incentives. User interactions are continuously incorporated into the KG for iterative personalization. Evaluated on data from 25 participants, the authors report that KG-grounded recommendations were preferred over non-grounded LLM outputs on relevance, personalization, explainability, and trust.

中文

本文提出 ADAPTS，一种基于知识图谱（KG）的激励型工作效率支持系统。该框架将室内环境测量、可穿戴设备的生理与行为指标及问卷用户画像融合到 RDF/OWL 知识图谱（KG）中。基于规则的推理层生成高层语义并计算每日生产力得分（DPS）。通过 GraphRAG 机制利用富化后的图构建统一用户上下文，以此为 LLM 提供支撑，生成个性化干预和自适应激励。用户交互持续写入知识图谱（KG）以实现迭代个性化。在 25 名参与者的数据上评估，作者报告 KG 支持的推荐在相关性、个性化、可解释性和信任度上优于非图支撑的 LLM 推荐。

Thesis relevance

English

Overlap: both use Knowledge Graphs and GraphRAG to ground LLM outputs and continuously incorporate user interactions into a graph. Differences: ADAPTS is a domain-specific, sensor-and-user-profile application using an RDF/OWL schema and rule-based reasoning, evaluated on 25 participants; it does not address multi-hop retrieval tasks, agentic paradigms, or entity-resolution trade-offs central to the thesis. Complementarity: the paper offers practical examples of KG-grounding, rule-based enrichment, and user-facing metrics (explainability, trust) that could inform evaluation axes and graph enrichment strategies in the thesis.

中文

重合点：两者都使用知识图谱（KG）和 GraphRAG 为 LLM 输出提供支撑，并将用户交互持续写入图中。差异：ADAPTS 是面向特定场景的应用，采用 RDF/OWL 模式与基于规则的推理，在 25 名参与者上评估；它并未处理论文关注的多跳检索任务、智能体范式或实体解析权衡。互补性：该工作在 KG 支撑、规则性图增强以及以用户为中心的评估指标（可解释性、信任）方面提供了实用示例，可为论文的评估维度和图增强策略提供参考。

Writing and related work

English

Position this paper as an application demonstrating the usability and user-trust benefits of KG-grounded LLM recommendations rather than as a retrieval- or multi-hop-reasoning advance. Emphasize its use of RDF/OWL and rule-based enrichment when contrasting with systems that assemble graph memory from agent interactions.

中文

将该文定位为展示 KG 支持的 LLM 推荐在可用性与用户信任方面优势的应用研究，而非检索或多跳推理方面的进展。比较时可强调其采用 RDF/OWL 与基于规则的图增强方法，以区分于由智能体交互构建图记忆的系统。

Method and evaluation

English

Consider using their RDF/OWL representation and rule-based enrichment as a baseline for graph construction, while replacing or extending the curated rules with LLM-extracted triples to match the thesis’ incremental-assembly premise. Adopt their user-facing metrics (perceived relevance, personalization, explainability, trust) alongside the thesis’ technical metrics (bridge-entity ID, retrieval accuracy, graph coherence). If possible, replicate continuous incorporation as a controlled simulation to measure whether structured graph memory improves over repeated interactions.

中文

可考虑将其 RDF/OWL 表示与基于规则的富化作为图构建基线，同时用或扩展为 LLM 提取的三元组以符合论文的增量组装设定。将其面向用户的指标（感知相关性、个性化、可解释性、信任）与论文的技术指标（桥接实体识别、检索准确性、图一致性）并列评估。若可行，可复现其持续写入机制作为受控模拟，以测量结构化图记忆在重复交互中的累积效用。

Future directions

English

Extend evaluation beyond the productivity domain and the small participant set, integrate automatic entity-resolution mechanisms, and test the approach on multi-hop retrieval benchmarks such as MultiHop-RAG. Also explore hybrid pipelines that combine rule-based inference with LLM-derived triples to balance precision and scalability.

中文

将评估扩展到生产力之外的领域并扩大样本量，集成自动实体解析机制，并在多跳检索基准（例如 MultiHop-RAG）上测试该方法。还可探索将基于规则的推理与 LLM 生成的三元组结合的混合管线，以在精度与可扩展性间取得平衡。

---

04 · Lost in Conversation or Lost in Translation? Diagnosing Multi-Turn Degradation in RAG

是在对话中迷失还是在翻译中迷失？诊断检索增强生成（RAG）在多轮设置中的性能退化

Research record	Details
Authors	Pranav Handa, Ariful Azad
Published	2026-09-29
Sources	openalex
Focus	multi-turn evaluation, retrieval-augmented generation (RAG), GraphRAG, multi-hop QA, conversational robustness 多轮评估, 检索增强生成（RAG）, GraphRAG, 多跳问答, 对话鲁棒性

English

Highly relevant: provides a large-scale diagnostic framework, quantitative degradation metrics, and a failure-mode taxonomy that directly inform evaluation design and motivation for developing persistent graph memory and entity-resolution strategies in the thesis.

中文

高度相关：该论文提供了大规模诊断框架、量化的性能退化指标和失败模式分类，直接有助于为本论文中开发持久化图记忆与实体解析策略设计评测方案与论证依据。

Research synopsis

English

The paper investigates an evaluation mismatch: RAG and GraphRAG are usually benchmarked on single-turn, fully specified queries, whereas real users often build multi-hop questions through follow-up turns. The authors conduct a large-scale simulation study by converting multi-hop QA questions into underspecified conversational sequences and evaluating ten LLM assistants with eight retrieval systems across 1.5 million simulated conversations. They report widespread multi-turn performance degradation (up to 21% relative drop) and increased unreliability (up to 47%). Two failure modes are identified: “lost in translation” (rephrasing harms retrieval) and “lost in conversation” (retrieval succeeds but the LLM fails to synthesize evidence across turns).

中文

本文研究了一个评估不匹配的问题：RAG 与 GraphRAG 通常在单轮、完全指定的问题上进行评测，而真实用户常通过后续回合构建多跳问题。作者通过将多跳问答问题转换为不充分的对话序列，开展大规模仿真实验，评估十个 LLM 助手与八种检索系统，总计约 150 万次模拟对话。结果显示多轮交互普遍导致性能下降（相对下降可达 21%）并增加不可靠性（可达 47%）。论文识别出两种失败模式：“在翻译中迷失”（重述扭曲了检索查询）和”在对话中迷失”（检索成功但 LLM 无法跨回合合成证据）。

Thesis relevance

English

Overlap: both works address multi-hop QA and the limitations of standard RAG-style evaluation in multi-turn scenarios, and the candidate highlights retrieval vs. synthesis failure modes that are relevant to persistent memory design. Differences: the paper is diagnostic and large-scale but does not propose or evaluate persistent structured graph memory, entity resolution strategies, or the agentic paradigms (ReAct, Plan-and-Execute, Self-Ask, Reflexion) central to the thesis. Complementarity: its multi-turn simulation protocol and failure-mode taxonomy can motivate evaluation scenarios and ablations for the thesis’s graph memory and entity-resolution components.

中文

重合点：两者都关注多跳问答以及标准 RAG 风格评测在多轮场景下的局限性，论文中关于检索与合成失败模式的划分对持久化记忆设计具有参考价值。差异：该论文侧重诊断性的大规模分析，并未提出或评估持久化的结构化图记忆、实体解析策略，亦未研究论文中关注的智能体范式（ReAct、Plan-and-Execute、Self-Ask、Reflexion）。互补性：其多轮仿真协议与失败模式分类可用于为本论文的图记忆与实体解析设计评测场景与消融实验提供依据。

Writing and related work

English

Use this paper to justify multi-turn evaluation and to motivate why a persistent, query-aware memory might be needed. Cite its failure-mode taxonomy when arguing for separating retrieval quality from LLM synthesis capability. Note that the study is simulation-based, which should be mentioned when discussing ecological validity.

中文

可用此论文为多轮评估提供论证，并以此作为说明为何需要持久化、查询感知的记忆的动机。引用其失败模式分类以支持将检索质量与 LLM 合成能力区分开讨论。注意该研究基于仿真，应在讨论生态有效性时加以说明。

Method and evaluation

English

Adopt the paper’s multi-turn simulation protocol to stress-test the thesis system across incremental interactions. Instrument experiments to distinguish ‘lost in translation’ (measure query drift and retrieval recall per turn) from ‘lost in conversation’ (measure retrieval hits versus answer synthesis success). Apply their large-scale, cross-system evaluation style as a template for comparing entity-resolution strategies and agentic paradigms under realistic conversational drift.

中文

可采用论文的多轮仿真协议对论文系统在增量交互下进行压力测试。设计实验以区分“在翻译中迷失”（衡量查询漂移与每回合检索召回）与“在对话中迷失”（比较检索命中与答案合成成功率）。借鉴其大规模、多系统评估方式，作为比较不同实体解析策略与智能体范式在对话漂移下表现的模板。

Future directions

English

Extend their diagnostics to evaluate whether structured, persistent graph memory and entity-resolution methods reduce the identified failure modes. Run analogous simulations comparing query-aware graph memory, vector RAG, and agentic RAG variants to measure changes in retrieval drift and synthesis reliability. Validate key findings with smaller-scale human-in-the-loop conversation data.

中文

将其诊断方法扩展为评估结构化持久图记忆与实体解析方法能否降低上述失败模式。进行类似的仿真，比较查询感知图记忆、向量 RAG 与具有智能体的 RAG 变体，评估检索漂移与合成可靠性的变化。用小规模人机交互数据验证主要结论的可复现性。

---

05 · Salience, Ranking, and Metabolism: Three Conflated Signals in Long-Running Agent Memory Systems

显著性、排序与代谢：长期运行智能体记忆系统中被混淆的三种信号

Research record	Details
Authors	Baofeng Zhao
Published	2026-09-30
Sources	openalex
Focus	agent memory management, salience vs. retention signals, production telemetry and audit tools, memory-store degeneration 智能体记忆管理, 显著性与保留信号, 生产遥测与审计工具, 记忆存储退化

English

Practical, production-derived insights into memory lifecycle and concrete audit tools are directly relevant to evaluating long-term graph memory; the paper’s metrics and failure modes could materially inform experimental design and diagnostics.

中文

基于生产的实务见解与具体审计工具对评估长期图记忆极具参考价值；其度量与失效模式可实质性指导实验设计与诊断。

Research synopsis

English

This paper examines a common engineering practice that conflates three distinct functions—salience, ranking, and metabolism—into a single “importance” signal inside long-running agent memory systems. Drawing on 14+ months of production telemetry from a single conversational system (tens of thousands of stored items and association links), the author documents failures caused by the conflation (e.g., extreme semantic skew, high fraction of isolated memories, skewed access distribution). The contribution is a signal taxonomy with four invariants (I0–I3), six design laws, and two same-day audit tools (the Silence Test and the RAG Test) intended to detect and prevent memory drift; no benchmark claims are made.

中文

本文检视了一种将显著性、排序与代谢三种功能合并为单一“重要性”信号的工程惯例，及其在长期运行智能体记忆系统中的后果。作者基于一套单一对话系统的14+个月生产遥测（数万条记忆条目与关联链接）记录了因混淆产生的失效实例（如严重的语义偏斜、大量孤立记忆、访问分布高度集中）。论文贡献了一个信号分类（不变量 I0–I3）、六条设计法则，以及两种同日审计工具（Silence Test 与 RAG Test）用于检测和防止记忆漂移；未主张任何基准优越性。

Thesis relevance

English

Overlap: the paper addresses lifecycle and retention policies for long-term agent memory, which directly affect graph coherence and the reuse of prior reasoning paths in a query-aware graph memory. Differences: it focuses on signal taxonomy, production telemetry, and audit tooling rather than on structured graph construction, entity resolution, or multi-hop QA performance. Complementarity: the audit methods and metrics (e.g., access Gini, island ratio, structure-aware coherence) could be incorporated into the thesis’s evaluation suite to quantify how memory metabolism choices impact graph fragmentation, retrieval efficiency, and verifiability.

中文

重合点：论文讨论了长期智能体记忆的生命周期与保留策略，这些策略会直接影响图记忆的连贯性以及先前推理路径的重用。差异：论文侧重于信号分类、生产遥测与审计工具，而非结构化图构建、实体解析或多跳问答性能。互补性：其审计方法与度量（如访问基尼系数、孤岛比率、结构感知连贯性）可并入论文的评估体系，用以量化记忆代谢策略对图碎片化、检索效率和可验证性的影响。

Writing and related work

English

Position this paper in related work as a practical caution about lifecycle and retention signals: argue that reporting aggregate ‘importance’ is insufficient and that structured-memory work should measure signal separation and store-level coherence. Cite the paper’s invariants and audit tools when motivating evaluation metrics.

中文

在相关工作中将该论文定位为关于生命周期与保留信号的实用警示：指出仅报告聚合的“重要性”指标不足，结构化记忆研究应测量信号分离与存储级连贯性。在论证评估指标时引用其不变量与审计工具。

Method and evaluation

English

Apply the paper’s audit metrics (access-distribution Gini, island ratio, structure-aware coherence) to the evolving graph memory and report them alongside multi-hop retrieval accuracy and latency. Use the Silence Test and RAG Test to detect whether retention policies cause useful bridge entities or reasoning paths to be retired. Experimentally compare conflated vs. separated signaling (salience/ranking/metabolism) and measure downstream effects on entity-resolution quality and end-to-end QA.

中文

将论文的审计度量（访问分布基尼系数、孤岛比率、结构感知连贯性）应用于演化中的图记忆，并与多跳检索准确率和延迟一起报告。使用 Silence Test 与 RAG Test 检测保留策略是否导致有用的桥接实体或推理路径被淘汰。通过实验比较合并信号与分离信号（显著性/排序/代谢），并测量其对实体解析质量和端到端问答性能的下游影响。

Future directions

English

Adapt the proposed signal taxonomy and audit tools to graph-based memory stores and evaluate their sensitivity to entity-resolution policies. Validate whether separating salience, ranking, and metabolism improves multi-hop retrieval and verifiability on benchmarks such as the MultiHop-RAG set.

中文

将该信号分类与审计工具适配到基于图的记忆存储，并评估其对实体解析策略的敏感性。验证将显著性、排序与代谢分离是否能在如 MultiHop-RAG 之类的基准上提升多跳检索与可验证性。

---

06 · Incident Knowledge Graphs for Site Reliability Engineering: Connecting Alerts, Runbooks, Services, Deployments, and Postmortems

事故知识图谱（IKG）用于站点可靠性工程：连接告警、运行手册、服务、部署与事后分析

Research record	Details
Authors	Venkata Praveen Annam
Published	2026-09-29
Sources	openalex
Focus	incident knowledge graph, graph+embedding hybrid retrieval, site reliability engineering (SRE), persistent institutional memory 知识图谱（KG）、图+向量混合检索、站点可靠性工程、持久化机构记忆

English

The paper provides concrete schema, extraction/construction methods, hybrid retrieval architecture, and deployment-based evaluation metrics that are directly useful as engineering baselines and for designing realistic evaluations of persistent graph memory in the thesis. Its operational findings about maintaining a ‘living’ graph are particularly relevant to long-term memory maintenance.

中文

该论文提供了具体的图模式、抽取与构建方法、混合检索架构以及基于部署的评估指标，可直接作为工程基线并用于设计关于持久化图记忆的现实评估。其关于保持“活跃”图的运维发现对长期记忆维护尤为相关。

Research synopsis

English

The paper addresses fragmentation of operational context during production incidents by proposing Incident Knowledge Graphs (IKG), a continuously updated property graph that represents alerts, runbooks, services, deployments, and postmortems as typed nodes and relations. Retrieval combines graph traversal with dense embedding search to surface contextually relevant information during live incidents. Evaluation uses a held-out benchmark of 640 on-call retrieval queries and a twelve-month pilot; the hybrid approach reports 0.86 precision@5 and 0.83 MRR and is associated in deployment with reductions in median time-to-runbook (9.8→1.6 minutes) and median time-to-resolution (88.0→46.0 minutes) across 187 incidents. The paper documents the schema, extraction/construction pipeline, retrieval architecture, evaluation, and organizational factors affecting graph maintenance.

中文

本文针对生产事故期间运维上下文分散的问题，提出了事故知识图谱（IKG），将告警、运行手册、服务、部署和事后分析表示为持续更新的有类型节点与关系的属性图。检索将图遍历与稠密向量检索相结合，以在现场事故中提供相关上下文信息。评估包括640条留出检索查询基准和为期十二个月的试点；混合检索方法在部署评估中给出0.86的precision@5和0.83的MRR，并在实际使用中将查找相关运行手册的中位时间从9.8降至1.6分钟，将中位解决时间从88.0降至46.0分钟（基于187次跟踪的事故）。文章还描述了图模式、抽取与构建流程、检索架构、评估结果以及影响图作为活跃机构记忆的组织实践。

Thesis relevance

English

Overlap: both construct continuously updated property graphs to centralize heterogeneous operational data and use hybrid graph/embedding retrieval to improve contextual retrieval. Differences: this paper is an applied SRE system focused on live-incident retrieval and deployment metrics, and it does not address agentic LLM-driven RAG workflows, explicit extraction of S–P–O triples for reuse in multi-hop QA, nor the thesis’s planned empirical comparison of entity-resolution strategies. Complementarity: the IKG schema, extraction pipeline, and deployment-driven evaluation metrics can inform the thesis’s engineering baseline, retrieval architecture, and practical maintenance considerations for long-lived graph memory.

中文

重合点：两者都构建持续更新的属性图以集中异构运维数据，并采用图与向量混合检索提升上下文检索能力。差异：本文是面向SRE的应用系统，侧重现场事故检索与部署指标，未涉及智能体驱动的检索增强生成（RAG）工作流、用于多跳问答可重用的主谓宾三元组抽取，也未进行论文中计划的实体消歧策略对比实验。互补性：IKG的图模式、抽取与构建流水线以及基于部署的评估指标可为论文在工程基线、检索架构和长期图记忆维护的实践性考量提供参考。

Writing and related work

English

Position this paper as an applied, deployed example of constructing and operating a living knowledge graph; cite it for practical schema design, continuous extraction, and empirical retrieval baselines. Note that its domain-driven choices and operational constraints may differ from research settings for multi-hop QA and agentic systems.

中文

将该论文作为构建并运营“活跃”知识图谱的应用部署案例来定位；可引为图模式设计、持续抽取和经验检索基线的实用参考。需注意其领域驱动的设计和运维约束可能与多跳问答和智能体系统的研究设置不同。

Method and evaluation

English

Adopt the paper’s hybrid graph-traversal plus dense-embedding retrieval as a strong baseline when comparing structured graph memory to vector-only RAG. Reuse evaluation metrics reported (precision@5, MRR, task-centered time metrics) and consider replicating the held-out query benchmark format for operational queries. However, extend their pipeline by extracting explicit S–P–O triples and instrumenting comparisons of entity-resolution strategies (embedding similarity vs. LLM-as-judge) to measure impact on multi-hop path coherence and QA accuracy.

中文

可将论文提出的图遍历与稠密向量检索混合方法作为与纯向量RAG比较时的有力基线。复用其报告的评估指标（precision@5、MRR、以任务为中心的时间度量），并考虑采用类似的留出查询基准以评估运行时检索效果。需要在其流水线基础上增加明确的主谓宾三元组抽取，并对实体消歧策略（向量相似度 vs. LLM判定）进行对照实验，以评估对多跳路径连贯性与QA准确率的影响。

Future directions

English

Integrate agentic LLMs to populate and query the IKG during conversational incident diagnosis and evaluate whether a query-aware graph memory improves multi-hop reasoning. Conduct controlled experiments on entity-resolution strategies and quantify their effects on multi-step retrieval and verifiability. Explore generalization of the IKG approach beyond SRE to other operational domains.

中文

将智能体驱动的LLM集成到IKG的构建与查询流程中，以评估查询感知图记忆对多跳推理的提升效果。对实体消歧策略进行受控实验，并量化其对多步检索与可验证性的影响。探索IKG方法在SRE之外其他运维或业务领域的泛化性。

---