01 · GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory

GraphMemix：用于长期多模态智能体记忆的查询感知证据森林

Research record	Details
Authors	Geng Li, Yuhao Wang, Dong Li, Jianye Hao, Yuxin Peng
Published	2026-08-27
Sources	arxiv
Focus	query-aware graph memory, long-term multimodal agent memory, combinatorial optimization for memory selection, evidence budgeting 查询感知图记忆, 长期多模态智能体记忆, 记忆选择的组合优化, 证据预算

English

The paper introduces a concrete, optimization-based approach to query-aware subgraph selection and evidence budgeting that is directly applicable to limiting lifecycle cost and recovering low-similarity evidence in a persistent graph-memory system; these techniques could materially inform the thesis’s design and evaluations.

中文

该文提出了面向查询的组合优化子图选择与证据预算方法，直接可用于在持久图记忆系统中控制生命周期成本并检索低相似度证据；这些技术可能对论文的系统设计和评估带来实质性启发，因此值得深入阅读。

Research synopsis

English

GraphMemix addresses long-term memory organization for multimodal agents by framing memory selection as a query-aware evidence-forest construction problem. The paper proposes three components: (1) candidate graph construction that expands seed memories via schema and semantic relations to obtain query-relevant context; (2) separating evidence utility from activation costs to reduce redundancy and conflict; and (3) a forest optimization that selects a forest-format subgraph under a maximum evidence budget while recovering reliable relational structure. Experiments on four long-term multimodal memory benchmarks are reported to improve accuracy and lifecycle cost, claiming a new Pareto frontier between those metrics.

中文

GraphMemix 通过将记忆选择建模为查询感知的证据森林构造，解决长期多模态智能体记忆的组织问题。文章提出三部分方法：(1) 候选图构建：通过模式和语义关系扩展种子记忆以获取查询相关上下文；(2) 证据效用与激活成本解耦，以抑制冗余或冲突信息；(3) 森林优化：在最大证据预算下联合选择森林格式的子图及其可靠的关系结构。作者在四个长期多模态记忆基准上报告了在准确性和生命周期成本上的改进，并声称在二者之间建立了新的帕累托前沿。

Thesis relevance

English

Overlap: GraphMemix tackles query-aware, long-term agent memory organized as graph-structured subcontexts, which directly intersects the thesis interest in query-aware graph memory and persistent agentic memory. Differences: GraphMemix emphasizes budgeted, query-time selection of evidence forests for multimodal agents rather than incremental, persistent knowledge-graph assembly from the agent’s own retrieval/inference activity; it does not foreground entity-resolution trade-offs or comparisons among agentic RAG paradigms. Complementarity: The forest-optimization and evidence-activation cost ideas could be integrated into the thesis system as a subgraph selection mechanism to limit lifecycle cost and to surface low-similarity complementary evidence.

中文

重合点：GraphMemix 关注将长期智能体记忆组织为图结构的查询相关子上下文，这与论文中对查询感知图记忆和持久智能体记忆的研究直接相关。差异：GraphMemix 强调在预算约束下的查询时证据森林选择，主要面向多模态智能体，而不是论文中从系统检索/推理活动增量构建的持久知识图谱；该工作也没有把实体解析权衡或不同 agentic RAG 范式的对比作为重点。互补性：其森林优化和证据激活成本建模可以作为论文系统的子图选择模块，用来控制生命周期成本并检索低相似度的互补证据。

Writing and related work

English

Position GraphMemix as a memory-selection/organization technique that complements persistent KG construction. When discussing related work, contrast its query-time, budgeted forest selection with approaches that build or accumulate structured memory incrementally. Cite it when motivating budgeted evidence selection and suppression of redundant/conflicting context.

中文

将 GraphMemix 定位为一种记忆选择/组织技术，可补充持久知识图谱的构建。在相关工作中，应将其的查询时、受预算约束的森林选择与增量构建结构化记忆的方法进行对比。可在论述预算化证据选择和抑制冗余/冲突上下文时引用该工作。

Method and evaluation

English

Consider adapting forest optimization to the thesis pipeline by treating multi-hop retrieval as a budgeted path-selection problem over the evolving graph. Incorporate evidence utility and activation-cost terms when selecting subgraphs to present to the local LLM (Qwen2.5-7B), and measure impacts on latency and retrieval accuracy. Evaluate on MultiHop-RAG by comparing forest-selected subgraphs against pure vector RAG and agentic RAG baselines, and report lifecycle cost (e.g., memory maintenance or retrieval overhead) alongside QA metrics.

中文

可将森林优化移植到论文的流水线中，将多跳检索视为在演化图上受预算约束的路径选择问题。在选择呈现给本地 LLM（如 Qwen2.5-7B）的子图时，加入证据效用和激活成本项，并衡量其对延迟和检索准确率的影响。在 MultiHop-RAG 上比较森林选择的子图与纯向量 RAG 和无持久图记忆的 agentic RAG 基线，并同时报告生命周期成本（如记忆维护或检索开销）与 QA 指标。

Future directions

English

Apply GraphMemix-style forest optimization to incrementally assembled, persistent graph memory and integrate explicit entity-resolution signals (embedding vs. LLM-judge) into candidate expansion and activation. Test the method inside different agentic paradigms (ReAct, Plan-and-Execute, Self-Ask, Reflexion) to measure effects on multi-hop retrieval, verifiability, and end-to-end latency.

中文

将 GraphMemix 风格的森林优化应用到逐步组装的持久图记忆中，并将显式实体解析信号（嵌入相似度 vs. LLM 评判）纳入候选扩展与激活流程。在不同的智能体范式（ReAct、Plan-and-Execute、Self-Ask、Reflexion）内部测试该方法，以衡量对多跳检索、可验证性和端到端延迟的影响。

---