Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

频道: arXiv
原文: https://arxiv.org/abs/2609.23986
原文语言: en
统计: 共 75 轮

[¶1]

Dongming Jiang, Yi Li, Bingzhe Li†

Department of Computer Science, The University of Texas at Dallas († Corresponding author: {dongming.jiang, yi.li3, bingzhe.li}@utdallas.edu)

Dongming Jiang, Yi Li, Bingzhe Li†

德州大学达拉斯分校（The University of Texas at Dallas）计算机科学系（† 通讯作者：{dongming.jiang, yi.li3, bingzhe.li}@utdallas.edu）

[¶2]

Abstract

Agentic memory is becoming essential for long-horizon AI agents, yet many existing systems rely on autoregressive LLMs to control how memories are organized, retrieved, and used, placing expensive generation on the critical path of memory operations. We introduce Jev-Mem, a new agentic memory architecture inspired by System-One/System-Two cognition. System One captures fast, lightweight decision-making, whereas System Two performs slower, deliberative reasoning. Jev-Mem brings this division of labor to agentic memory through a dedicated System-One control plane, a structured multi-relational memory plane, and a System-Two reasoning plane. The System-One controller governs memory typing and relational organization during construction, and dynamically performs query routing, retrieval-budget allocation, graph traversal, candidate scoring, and adaptive stopping during retrieval. System Two is invoked only for complex reasoning and answer synthesis. This design improves both memory effectiveness and system efficiency: on LoCoMo Jev-Mem achieves an overall LLM-as-a-Judge score of 0.777, an 11.0% relative improvement over the strongest baseline, while reducing memory construction time to 158 s, a 6.6× speedup over the fastest competing memory system, and lowering average query latency to 0.93 s, a 36.7% reduction. The code of Jev-Mem is publicly available. (Footnote: 1 https://github.com/libingzheren/Jev-Mem)

摘要

对长程（long-horizon）AI 智能体而言，智能体记忆（agentic memory）正变得不可或缺；然而，许多现有系统依赖自回归（autoregressive）LLM 来控制记忆的组织、检索与使用方式，从而把代价高昂的生成过程置于记忆操作的关键路径（critical path）上。本文提出 Jev-Mem，这是一种新的智能体记忆架构，其灵感来自系统一/系统二认知（System One / System Two，源自双过程理论：快思考 / 慢思考）。系统一负责快速、轻量的决策，系统二则执行较慢的审慎推理。Jev-Mem 通过三个部分将这种分工引入智能体记忆：一个专用的系统一控制平面（control plane）、一个结构化的多关系（multi-relational）记忆平面（memory plane），以及一个系统二推理平面（reasoning plane）。在记忆构建（memory construction）阶段，系统一控制器负责记忆类型标注与关系组织；在检索阶段，它动态执行查询路由（query routing）、检索预算分配（retrieval-budget allocation）、图遍历（graph traversal）、候选打分（candidate scoring）与自适应停止（adaptive stopping）。系统二仅在复杂推理与答案综合时才被调用。这一设计同时提升了记忆效果与系统效率：在 LoCoMo 基准上，Jev-Mem 的总体 LLM 裁判（LLM-as-a-Judge）得分达到 0.777，相较最强基线（baseline）相对提升 11.0%；记忆构建时间缩短至 158 秒，相比最快的竞争记忆系统提速 6.6 倍；平均查询延迟降至 0.93 秒，降幅为 36.7%。Jev-Mem 的代码已公开。（脚注：1 https://github.com/libingzheren/Jev-Mem）

[¶3]

1 Introduction

1 引言

[¶4]

Large language model (LLM) agents are increasingly expected to operate as persistent systems over long interaction horizons, supporting applications such as coding assistants, personal agents, research agents, and autonomous workflows (Brown et al., 2020; Achiam et al., 2023; Wei et al., 2022). Such agents typically interleave reasoning with tool invocation and environment actions (Yao et al., 2023b; Schick et al., 2023), and operate over multi-step environments such as interactive web platforms and software repositories (Zhou et al., 2024; Yang et al., 2024). During these interactions, agents continuously accumulate user preferences, task history, and environment knowledge, quickly exceeding what can be maintained within a fixed context window (Beltagy et al., 2020; Liu et al., 2024; Press et al., 2021). Enlarging the nominal context length does not by itself resolve this problem, since models do not reliably exploit all positions of a long input (Hsieh et al., 2024), and long-term interactive settings introduce additional indexing, retrieval, and reading challenges (Lee et al., 2024; Wu et al., 2024; Hu et al., 2026). To remain effective over time, agents therefore need mechanisms that can retain useful experience beyond the prompt and recover it when needed. This requirement has made agentic memory which has the ability to preserve, organize, update, and retrieve past experience (Xu et al., 2025; Nan et al., 2025; Chhikara et al., 2025; Jiang et al., 2026a; Liu et al., 2026; Jiang et al., 2026c).

人们日益期望大语言模型（LLM）智能体能够作为持久运行的系统，在长程交互中持续工作，支撑编程助手、个人智能体、研究智能体以及自主工作流等应用 (Brown et al., 2020; Achiam et al., 2023; Wei et al., 2022)。这类智能体通常将推理与工具调用、环境动作交替进行 (Yao et al., 2023b; Schick et al., 2023)，并在交互式网页平台、软件代码仓库等多步环境中运行 (Zhou et al., 2024; Yang et al., 2024)。在这些交互过程中，智能体不断积累用户偏好、任务历史与环境知识，其规模很快便会超出固定上下文窗口所能容纳的范围 (Beltagy et al., 2020; Liu et al., 2024; Press et al., 2021)。单纯扩大名义上下文长度本身并不能解决这一问题：模型并不能可靠地利用长输入中的所有位置 (Hsieh et al., 2024)，而长期交互场景还会带来额外的索引、检索与阅读难题 (Lee et al., 2024; Wu et al., 2024; Hu et al., 2026)。因此，为了长期保持有效，智能体需要相应的机制，能够在提示词之外保留有用的经验，并在需要时将其取回。正是这一需求，使具备保存、组织、更新与检索过往经验能力的智能体记忆应运而生 (Xu et al., 2025; Nan et al., 2025; Chhikara et al., 2025; Jiang et al., 2026a; Liu et al., 2026; Jiang et al., 2026c)。

[¶5]

Recent work has transformed agentic memory from passive storage into an active and structured component of LLM agents. Early systems mainly store past interactions and retrieve them through semantic similarity, while newer approaches selectively retain salient information, consolidate repeated observations, and organize memory across hierarchical stores. Procedural memory further captures reusable skills, distilled experience, and recurring workflows (Wang et al., 2023; Zhao et al., 2024; Wang et al., 2025). More recent relational approaches represent memories through knowledge graphs or multiple graph views, enabling retrieval over semantic, temporal, causal, and entity relationships rather than isolated memory fragments. As memory becomes richer, however, controlling it becomes increasingly expensive. Persistent agents must repeatedly decide what to store, update, connect, retrieve, and when to stop searching. Existing systems typically rely on either fixed heuristics or general-purpose autoregressive LLMs: the former are efficient but inflexible, while the latter provide semantic flexibility at the cost of repeated token generation. As these decisions appear throughout memory construction and retrieval, memory control itself can become a major source of latency and inference overhead.

近期研究已将智能体记忆从被动存储转变为 LLM 智能体中一个主动的、结构化的组件。早期系统主要存储过往交互，并通过语义相似度对其进行检索；较新的方法则有选择地保留显著信息，整合（consolidate）重复出现的观测，并在层级化的存储结构中组织记忆。程序性记忆（procedural memory）进一步捕获可复用的技能、提炼出的经验以及反复出现的工作流 (Wang et al., 2023; Zhao et al., 2024; Wang et al., 2025)。更新近的关系型方法借助知识图谱或多种图视图来表示记忆，使检索能够基于语义、时间、因果与实体关系展开，而不再局限于孤立的记忆片段。然而，随着记忆日益丰富，对其进行控制的代价也越来越高。持久运行的智能体必须反复决定存储什么、更新什么、关联什么、检索什么，以及何时停止搜索。现有系统通常要么依赖固定的启发式规则，要么依赖通用的自回归 LLM：前者高效但缺乏灵活性，后者虽能提供语义层面的灵活性，却要以反复生成 token 为代价。由于这些决策贯穿记忆构建与检索的全过程，记忆控制本身就可能成为延迟与推理开销的主要来源。

[¶6]

This observation leads us to reconsider agentic memory through the lens of System One and System Two, a distinction drawn from dual-process accounts of human reasoning that separate fast, automatic processing from slower and more deliberative thought (Evans, 2008). Many memory operations including memory typing, relation judgment, query routing, candidate scoring, and stopping are structured, high-frequency decisions that naturally fit lightweight System-One computation, while evidence synthesis and final answer generation remain better suited to System Two. Jev TypeSafe AI (2026) makes such a separation practical by producing typed probabilistic decisions without autoregressive generation. Yet several questions remain: How should System One be integrated across the memory lifecycle? How can a weaker reasoner control memory without hurting downstream accuracy? What memory structure and retrieval process best support lightweight control?

这一观察促使我们从系统一与系统二的视角重新审视智能体记忆。这一区分源自关于人类推理的双过程（dual-process）理论，该理论将快速、自动的加工与较慢、更为审慎的思考区分开来 (Evans, 2008)。许多记忆操作——包括记忆类型标注、关系判断、查询路由、候选打分与停止判定——都是结构化的高频决策，天然适合由轻量的系统一计算来承担；而证据综合与最终答案生成则仍更适合交给系统二。Jev TypeSafe AI (2026) 无需自回归生成即可产出类型化（typed）的概率决策，从而使这种分离在实践中变得可行。然而，仍有若干问题有待解答：系统一应如何融入记忆的整个生命周期？能力较弱的推理器如何在不损害下游准确率的前提下控制记忆？何种记忆结构与检索流程最能支撑轻量级控制？

[¶7]

Inspired by this observation, we propose Jev-Mem, a new System-One/System-Two architecture that treats memory control itself as a first-class systems layer. Unlike existing agentic memory systems that rely on heuristics or repeatedly invoke autoregressive LLMs throughout memory construction and retrieval, Jev-Mem introduces a dedicated System-One control plane that handles high-frequency structured decisions, a shared structured memory data plane, and a System-Two reasoning plane reserved for complex synthesis and answer generation. This separation redesigns the full memory lifecycle: on the write path, System One performs memory typing, redundancy filtering, and semantic, temporal, causal, and entity relation construction; on the read path, it performs query routing, retrieval-budget allocation, graph traversal, candidate scoring, evidence assessment, and adaptive stopping. By moving these frequent decisions out of the generative reasoning loop, Jev-Mem reduces unnecessary autoregressive inference while still preserving strong reasoning capability through System Two. Jev provides one concrete realization of the System-One controller through typed probabilistic decisions, but the key novelty of Jev-Mem is the architectural separation of lightweight memory control from deliberative reasoning across both memory construction and retrieval.

受这一观察启发，我们提出 Jev-Mem：一种新的系统一/系统二架构，它将记忆控制本身视为一个一等（first-class）系统层。现有智能体记忆系统要么依赖启发式规则，要么在记忆构建与检索的全过程中反复调用自回归 LLM；与之不同，Jev-Mem 引入了一个负责处理高频结构化决策的专用系统一控制平面、一个共享的结构化记忆数据平面，以及一个专门用于复杂综合与答案生成的系统二推理平面。这种分离重新设计了完整的记忆生命周期：在写路径上，系统一执行记忆类型标注、冗余过滤，以及语义、时间、因果与实体关系的构建；在读路径上，它执行查询路由、检索预算分配、图遍历、候选打分、证据评估与自适应停止。通过把这些频繁的决策移出生成式推理循环，Jev-Mem 减少了不必要的自回归推理，同时仍借助系统二保留了强大的推理能力。Jev 通过类型化的概率决策为系统一控制器提供了一种具体实现，但 Jev-Mem 的关键创新在于架构层面的分离：在记忆构建与检索两个环节中，都将轻量的记忆控制与审慎推理分离开来。

[¶8]

We make four main contributions.

• We identify an opportunity to leverage System-One-style lightweight decision-making to enable more efficient agentic memory.

• We introduce Jev-Mem, a System-One/System-Two architecture that separates lightweight structured memory control from expensive generative reasoning.

• We develop a unified System-One control plane for both memory construction and retrieval, including multi-relational organization, query routing, adaptive graph traversal, candidate scoring, and stopping.

• Jev-Mem achieves the best overall accuracy while also delivering the highest efficiency among state-of-the-art baselines.

[Figure 1: Workflow of agentic memory.]

本文的主要贡献有以下四点。

• 我们发现了一个机会：可以利用系统一式的轻量决策，实现更高效的智能体记忆。

• 我们提出 Jev-Mem，一种将轻量的结构化记忆控制与昂贵的生成式推理相分离的系统一/系统二架构。

• 我们构建了一个同时服务于记忆构建与检索的统一系统一控制平面，涵盖多关系组织、查询路由、自适应图遍历、候选打分与停止判定。

• 在与当前最先进基线的比较中，Jev-Mem 在取得最佳总体准确率的同时，也实现了最高的效率。

[图 1：智能体记忆的工作流程。]

[¶9]

2 Background and Motivation

2 背景与动机

[¶10]

2.1 Agentic Memory

Figure 1 demonstrates the workflow of agnetic memory. Let an agent maintain an evolving memory M_{t}. At interaction step t, a query q_{t} retrieves relevant evidence

E_{t}=R(q_{t},M_{t})   (1)

which is then provided to a language model for reasoning and response generation:

o_{t}=L(q_{t},E_{t})   (2)

The resulting interaction may subsequently update the memory:

M_{t+1}=U(M_{t},q_{t},o_{t})   (3)

Agentic memory has evolved from simple retrieval over stored interaction histories toward increasingly active memory management. Modern systems may selectively organize observations, infer relationships among memories, consolidate information, route queries across different memory structures, and adapt retrieval based on the current query. As a result, memory is no longer only a passive store accessed by a retrieval function. It increasingly behaves as a dynamic subsystem that continuously makes decisions about how information should be organized and accessed.

2.1 智能体记忆

图 1 展示了智能体记忆的工作流程。设智能体维护一个不断演化的记忆 M_{t}。在第 t 个交互步，查询 q_{t} 检索出相关证据

E_{t}=R(q_{t},M_{t})   (1)

随后，这些证据被提供给语言模型，用于推理与生成回复：

o_{t}=L(q_{t},E_{t})   (2)

由此产生的交互随后可能会更新记忆：

M_{t+1}=U(M_{t},q_{t},o_{t})   (3)

智能体记忆已从对已存储交互历史的简单检索，逐步演进为日益主动的记忆管理。现代系统可以有选择地组织观测、推断记忆之间的关系、整合信息、在不同的记忆结构之间路由查询，并根据当前查询自适应地调整检索。因此，记忆不再只是一个由检索函数访问的被动存储；它越来越像一个动态子系统，持续地就信息应如何组织与访问做出决策。

[¶11]

This shift introduces an important systems question that has received comparatively less attention: what computational mechanism should execute these memory-management decisions? Both memory update U and retrieval R contain frequent semantic decisions. During memory construction, the system must determine how new information should be characterized and connected to existing knowledge. During retrieval, it must determine where to search, which candidates are useful, how much additional search is warranted, and when sufficient evidence has been collected.

External retrieval has long served as a mechanism for augmenting parametric language-model knowledge with non-parametric stores, through dense retrievers, retrieval-augmented pre-training, and retrieval-augmented generation (Karpukhin et al., 2020; Guu et al., 2020; Lewis et al., 2020; Borgeaud et al., 2022; Izacard et al., 2022). Agentic memory inherits much of this machinery, but differs in that its store is written by the agent’s own interaction history and must be maintained and reorganized over time rather than fixed in advance.

这一转变引出了一个重要但相对较少受到关注的系统问题：应当由何种计算机制来执行这些记忆管理决策？记忆更新 U 与检索 R 都包含频繁的语义决策。在记忆构建阶段，系统必须确定应如何刻画新信息，以及如何将其与已有知识相关联；在检索阶段，系统必须确定在哪里搜索、哪些候选有用、还值得进行多少额外搜索，以及何时已收集到足够的证据。

长期以来，外部检索一直是用非参数化存储来增强语言模型参数化知识的一种机制，其实现途径包括稠密检索器、检索增强预训练以及检索增强生成（RAG）等 (Karpukhin et al., 2020; Guu et al., 2020; Lewis et al., 2020; Borgeaud et al., 2022; Izacard et al., 2022)。智能体记忆继承了其中大部分机制，但不同之处在于：其存储内容由智能体自身的交互历史写入，并且必须随时间推移不断维护与重组，而非预先固定。

[¶12]

A closely related line of work studies adaptive control in retrieval-augmented generation. Rather than applying the same retrieval procedure to every query, active and adaptive methods dynamically determine whether retrieval is needed, when additional evidence should be acquired, or which retrieval strategy a given query warrants (Trivedi et al., 2023; Jiang et al., 2023; Asai et al., 2023; Jeong et al., 2024). These results motivate viewing retrieval as a controlled decision process rather than a fixed top-k operation. Jev-Mem builds on this view, and extends the same principle from evidence acquisition to the broader memory lifecycle, including memory construction, query routing, budget allocation, graph traversal, and stopping.

与此密切相关的一类工作研究的是检索增强生成中的自适应控制。主动式与自适应方法并不对每个查询套用相同的检索流程，而是动态地判断是否需要检索、何时应获取更多证据，或某一查询适合采用何种检索策略 (Trivedi et al., 2023; Jiang et al., 2023; Asai et al., 2023; Jeong et al., 2024)。这些成果启发人们将检索视为一个受控的决策过程，而非固定的 top-k 操作。Jev-Mem 以这一观点为基础，并将同一原则从证据获取扩展到更广泛的记忆生命周期，涵盖记忆构建、查询路由、预算分配、图遍历与停止判定。

[¶13]

Jev-Mem focuses on this memory-control layer. Rather than treating these decisions as incidental components embedded inside prompts or fixed heuristics, we make them an explicit part of the memory architecture.

Jev-Mem 聚焦于这一记忆控制层。我们不再把这些决策视为嵌在提示词或固定启发式规则中的附带组件，而是将其作为记忆架构中显式的组成部分。

[¶14]

2.2 The Opportunity for System-One Memory Control

A key observation is that many memory-control operations are semantic but not generative. During memory construction, the system may need to classify a memory or infer its relation to existing information; during retrieval, it may need to route a query, score candidates, or decide when to stop. These decisions typically produce bounded outputs such as labels, probabilities, or scores rather than free-form text.

Using an autoregressive LLM for such high-frequency decisions is therefore unnecessarily expensive: even simple judgments require token-by-token generation, formatting, and parsing. Because these operations repeatedly appear on the memory critical path, their overhead can accumulate quickly.

2.2 系统一记忆控制的机会

一个关键观察是：许多记忆控制操作是语义性的，而非生成性的。在记忆构建阶段，系统可能需要对一条记忆进行分类，或推断它与已有信息之间的关系；在检索阶段，系统可能需要路由查询、为候选打分，或决定何时停止。这些决策通常产生的是有界输出，例如标签、概率或分数，而非自由形式的文本。

因此，用自回归 LLM 来做这类高频决策，会带来不必要的高昂代价：即便是简单的判断，也需要逐 token 生成、格式化与解析。由于这些操作反复出现在记忆的关键路径上，其开销会迅速累积。

[¶15]

More broadly, cost-aware LLM inference systems have shown that different requests need not receive identical amounts of model computation. Cascading and routing approaches dynamically allocate requests across models of differing cost to improve the cost–quality tradeoff (Chen et al., 2024; Ong et al., 2025), and speculative decoding uses a small model to draft tokens that a larger model only verifies (Leviathan et al., 2023). Jev-Mem applies a related systems principle at a finer granularity: rather than routing only complete user requests between models, it separates frequent, bounded memory-control decisions from open-ended generative and deliberative reasoning.

更广泛地看，成本感知的 LLM 推理系统已经表明，不同请求无需获得同等的模型计算量。级联与路由方法在成本各异的模型之间动态分配请求，以改善成本与质量之间的权衡 (Chen et al., 2024; Ong et al., 2025)；推测解码（speculative decoding）则用一个小模型起草 token，再由大模型仅做验证 (Leviathan et al., 2023)。Jev-Mem 在更细的粒度上应用了一条相关的系统原则：它不只是在模型之间路由完整的用户请求，而是将频繁出现的、有界的记忆控制决策，与开放式的生成式推理和审慎推理分离开来。

[¶16]

This creates a natural opportunity for a System-One/System-Two design: use lightweight structured prediction for frequent memory-control decisions, while reserving System Two for complex reasoning and answer synthesis.

[Figure 2: Overview of Jev-Mem. A lightweight System-One control plane manages memory construction and adaptive retrieval, while System Two is reserved for complex reasoning and answer synthesis.]

这为系统一/系统二设计创造了一个天然的机会：对频繁的记忆控制决策采用轻量的结构化预测，而将系统二留给复杂推理与答案综合。

[图 2：Jev-Mem 概览。轻量的系统一控制平面负责管理记忆构建与自适应检索，而系统二则专门用于复杂推理与答案综合。]

[¶17]

3 Jev-Mem Algorithm Design

Jev-Mem redesigns agentic memory around a clear separation between fast memory control and deliberative reasoning. As illustrated in Figure 2, the architecture consists of three components: a System-One controller, a shared memory data plane, and a System-Two reasoning model. Rather than relying on an autoregressive LLM to manage every memory operation, the System-One controller makes lightweight, structured decisions throughout both writing and retrieval. The memory plane maintains canonical observations together with semantic, temporal, causal, and entity relations, while System Two is reserved for answer synthesis and other operations that require open-ended generation or deeper reasoning.

3 Jev-Mem 算法设计

Jev-Mem 围绕快速记忆控制与审慎推理之间的清晰分离，对智能体记忆进行了重新设计。如图 2 所示，该架构由三个组件构成：系统一控制器、共享的记忆数据平面，以及系统二推理模型。该架构并不依赖自回归 LLM 来管理每一项记忆操作，而是由系统一控制器在写入与检索的全过程中做出轻量的结构化决策。记忆平面维护规范观测及其语义、时间、因果与实体关系；系统二则专门用于答案综合，以及其他需要开放式生成或更深层推理的操作。

[¶18]

This architecture unifies memory construction and retrieval under the same control mechanism. On the write path, Jev-Mem preserves observations, assigns memory types, selects candidate memories, and determines how new observations should connect to the existing multi-relational memory. On the read path, it routes each query to the most relevant relational views, allocates retrieval effort, checks evidence sufficiency, expands the graph when necessary, and scores newly discovered candidates in an iterative feedback loop. By treating both workflows as coordinated System-One control processes, Jev-Mem replaces a collection of isolated heuristics and repeated LLM calls with a single, adaptive memory-management architecture.

该架构将记忆构建与检索统一在同一套控制机制之下。在写路径上，Jev-Mem 保留观测、为其分配记忆类型、筛选候选记忆，并决定新观测应如何接入现有的多关系记忆。在读路径上，它将每个查询路由到最相关的关系视图、分配检索资源、检查证据是否充分、在必要时扩展图，并在迭代反馈回路中为新发现的候选打分。通过把这两条工作流都视为相互协调的系统一控制过程，Jev-Mem 以单一的自适应记忆管理架构，取代了一系列彼此孤立的启发式规则和反复的 LLM 调用。

[¶19]

Let an observation be

o_{t}=(x_{t},\tau_{t},\mu_{t})   (4)

where x_{t} denotes its content, \tau_{t} an optional timestamp, and \mu_{t} its provenance. Jev-Mem maintains

\mathcal{M}_{t}=\left(V_{t},\{E_{t}^{g}\}_{g\in\mathcal{G}},I_{t}^{\mathrm{vec}},I_{t}^{\mathrm{lex}}\right)   (5)

where

\mathcal{G}=\{\mathrm{semantic},\mathrm{temporal},\mathrm{causal},\mathrm{entity}\}   (6)

All relational views share the same canonical memory nodes V_{t}; multiple typed edges may connect the same pair of memories, while the vector and lexical indexes provide complementary entry points into the same memory space.

设一条观测为

o_{t}=(x_{t},\tau_{t},\mu_{t})   (4)

其中 x_{t} 表示其内容，\tau_{t} 为可选的时间戳，\mu_{t} 为其来源（provenance）。Jev-Mem 维护的记忆结构为

\mathcal{M}_{t}=\left(V_{t},\{E_{t}^{g}\}_{g\in\mathcal{G}},I_{t}^{\mathrm{vec}},I_{t}^{\mathrm{lex}}\right)   (5)

其中

\mathcal{G}=\{\mathrm{semantic},\mathrm{temporal},\mathrm{causal},\mathrm{entity}\}   (6)

所有关系视图共享同一组规范记忆节点 V_{t}；同一对记忆之间可以由多条类型化的边相连，而向量索引与词法索引则为同一记忆空间提供了互补的入口点。

[¶20]

3.1 Typed System-One Memory Control

The central abstraction of Jev-Mem is a typed System-One controller \mathcal{J}(S,\mathcal{Q}), where S is structured state and \mathcal{Q} is a batch of explicit decision questions. The controller produces either probabilities for independent propositions or a distribution over mutually exclusive alternatives. For example, it can estimate whether two memories are semantically related, whether a candidate is relevant to the current query, or whether the retrieved evidence is sufficient. When exactly one action is required, it selects among predefined alternatives such as temporal relations.

This interface is intentionally different from free-form LLM prompting. Each decision exposes a small, known output space, allowing Jev-Mem to represent memory control directly as probabilities rather than generating intermediate natural-language reasoning and subsequently parsing it. Moreover, decisions sharing the same state can be evaluated together in a single batched invocation.

3.1 类型化的系统一记忆控制

Jev-Mem 的核心抽象是一个类型化的系统一控制器 \mathcal{J}(S,\mathcal{Q})，其中 S 为结构化状态，\mathcal{Q} 为一批显式的决策问题。该控制器要么为相互独立的命题输出概率，要么在互斥的备选项上输出一个概率分布。例如，它可以估计两条记忆是否语义相关、某个候选是否与当前查询相关，或已检索到的证据是否充分。当需要恰好选定一个动作时，它会在预定义的备选项（例如各种时间关系）中进行选择。

这一接口有意区别于自由形式的 LLM 提示（prompting）。每个决策只暴露一个规模小且事先已知的输出空间，使 Jev-Mem 能够直接以概率的形式表示记忆控制，而不必先生成中间的自然语言推理、再对其加以解析。此外，共享同一状态的多个决策可以在单次批量调用中一并评估。

[¶21]

The same interface governs both the write and read paths. On the write path, it controls memory typing and relation construction. On the read path, it performs query routing, candidate evaluation, and stopping. This shared control plane is a key architectural property of Jev-Mem: memory is not constructed by one collection of heuristics and retrieved by another, but instead managed throughout its lifecycle through the same System-One decision abstraction.

同一接口同时统辖写路径与读路径。在写路径上，它控制记忆类型标注与关系构建；在读路径上，它执行查询路由、候选评估与停止判定。这一共享的控制平面是 Jev-Mem 的一项关键架构特性：记忆并非由一套启发式规则构建、再由另一套启发式规则检索，而是在其整个生命周期中都通过同一个系统一决策抽象加以管理。

[¶22]

3.2 System-One-Guided Memory Construction

For each valid observation, Jev-Mem creates one canonical memory node and selectively constructs relations around it. The current design preserves observations rather than making an irreversible learned store-or-discard decision at ingestion time. This prevents information that appears unimportant initially from being permanently lost before a future query reveals its relevance. Selectivity is instead introduced when memory structure is constructed and later when memory is retrieved.

The controller first predicts four overlapping memory characteristics:

\mathbf{t}(v)=(t_{\mathrm{episodic}},t_{\mathrm{semantic}},t_{\mathrm{procedural}},t_{\mathrm{preference}})   (7)

3.2 系统一引导的记忆构建

对于每条有效观测，Jev-Mem 都会为其创建一个规范记忆节点，并有选择地围绕该节点构建关系。当前设计会保留观测，而不是在摄入阶段就做出一个基于学习且不可逆的「存储或丢弃」决策。这可以防止那些起初看似不重要的信息，在未来某个查询揭示其相关性之前就被永久丢失。取而代之的是，选择性被引入到记忆结构的构建阶段，以及之后的记忆检索阶段。

控制器首先预测四种可以相互重叠的记忆特征：

\mathbf{t}(v)=(t_{\mathrm{episodic}},t_{\mathrm{semantic}},t_{\mathrm{procedural}},t_{\mathrm{preference}})   (7)

[¶23]

These scores annotate the node rather than assigning it to a single mutually exclusive category. The canonical node retains the original observation, provenance, timestamp, embedding, entities, and type scores.

A central challenge is determining how a new observation relates to an increasingly large memory. Comparing it against every existing node would make controller cost grow directly with memory size. Jev-Mem therefore separates candidate discovery from relation judgment. Deterministic retrieval first combines vector similarity, lexical overlap, shared entities, and temporal proximity to identify at most K_{w} candidates:

C(v)=\operatorname{TopK}_{u\in V_{t}}s_{\mathrm{cand}}(v,u)   (8)

这些分数只是对节点加以标注，而不是将其归入某个单一的互斥类别。规范节点保留了原始观测、来源、时间戳、嵌入向量、实体以及类型分数。

一个核心挑战在于如何确定新观测与日益庞大的记忆之间的关系。若将其与每个已有节点逐一比较，控制器的开销就会直接随记忆规模增长。因此，Jev-Mem 将候选发现（candidate discovery）与关系判定（relation judgment）分离开来。首先由确定性检索综合向量相似度、词法重叠、共享实体与时间邻近度，找出至多 K_{w} 个候选：

C(v)=\operatorname{TopK}_{u\in V_{t}}s_{\mathrm{cand}}(v,u)   (8)

[¶24]

The System-One controller then evaluates only these candidate pairs.

For every pair (v,u), Jev estimates semantic relatedness, directional causal influence, same-episode membership, and, when necessary, entity equivalence. Whenever reliable structured information is already available, Jev-Mem avoids unnecessary learned inference. Timestamp ordering directly creates temporal relations, while exact shared identifiers directly create entity relations. When temporal order is implicit, the controller instead chooses among before, after, during, contains, overlaps, same_time, and unknown.

An inferred edge of type g is inserted only when

P(g\mid v,u)\geq\theta_{\mathrm{rel}}   (9)

随后，系统一控制器只对这些候选对进行评估。

对每一对 (v,u)，Jev 会估计二者的语义相关性、有方向的因果影响、是否属于同一情景，并在必要时估计实体等价性。只要已有可靠的结构化信息可用，Jev-Mem 就会避免不必要的学习式推断。时间戳的先后顺序会直接生成时间关系，而完全一致的共享标识符则直接生成实体关系。若时间顺序是隐含的，控制器则转而在 before（之前）、after（之后）、during（期间）、contains（包含）、overlaps（重叠）、same_time（同时）与 unknown（未知）之中做出选择。

类型为 g 的推断边仅在满足下式时才会被插入：

P(g\mid v,u)\geq\theta_{\mathrm{rel}}   (9)

[¶25]

Because relations are independent views over a common memory space, the same pair of nodes may simultaneously exhibit semantic, temporal, causal, and entity relationships. This differs from assigning each memory to a single graph or duplicating the same observation across several memory stores. The write path is

o_{t}\rightarrow\mathrm{type}\rightarrow\mathrm{candidates}\rightarrow\mathrm{relations}\rightarrow\mathcal{M}_{t}   (10)

This design improves memory quality in two ways. First, preserving canonical observations avoids premature information loss. Second, selective relation construction suppresses unnecessary graph connectivity while retaining relations that can later support semantic, temporal, causal, or entity-based reasoning.

由于各类关系是建立在同一记忆空间之上的相互独立的视图，同一对节点可以同时具有语义、时间、因果和实体关系。这既不同于把每条记忆只归入某一张图，也不同于把同一条观测复制到多个记忆存储中。写路径为

o_{t}\rightarrow\mathrm{type}\rightarrow\mathrm{candidates}\rightarrow\mathrm{relations}\rightarrow\mathcal{M}_{t}   (10)

这一设计从两个方面提升了记忆质量。其一，保留规范观测可以避免信息过早丢失。其二，有选择的关系构建在抑制不必要的图连通性的同时，保留了日后能够支撑语义、时间、因果或基于实体的推理的关系。

[¶26]

Periodic maintenance further examines a bounded neighborhood for redundancy, contradiction, obsolescence, and useful additional links. These decisions enrich the memory structure without removing the original evidence. If a higher-level textual abstraction is required, the control plane may explicitly escalate an approved merge or promotion to System Two; generation is therefore an optional consequence of a structured control decision rather than the default mechanism for memory maintenance.

周期性维护还会在一个有界邻域内检查冗余、矛盾、过时的内容，以及值得补充的有用链接。这些决策在不删除原始证据的前提下丰富了记忆结构。若需要更高层次的文本抽象，控制平面可以将已获批准的合并或提升（promotion）操作显式地移交给系统二；因此，生成只是结构化控制决策的一种可选结果，而非记忆维护的默认机制。

[¶27]

3.3 Adaptive System-One Retrieval

Jev-Mem treats retrieval as a closed-loop control process rather than a single top-k search. Given a query q, the controller first predicts the relevance of each relational view,

\mathbf{p}(q)=\{p_{g}(q)\}_{g\in\mathcal{G}}   (11)

together with a multi-hop requirement h(q) and a recency importance score r(q). Because these probabilities are evaluated independently, a query may activate several graph views simultaneously instead of being assigned to one discrete retrieval intent.

A relation type is active when

p_{g}(q)\geq\theta_{\mathrm{act}}   (12)

Given a total graph-expansion budget B, Jev-Mem distributes search effort according to the predicted graph needs:

3.3 自适应的系统一检索

Jev-Mem 将检索视为一个闭环控制过程，而非单次的 top-k 搜索。给定查询 q，控制器首先预测每个关系视图的相关度，

\mathbf{p}(q)=\{p_{g}(q)\}_{g\in\mathcal{G}}   (11)

同时预测多跳需求 h(q) 与新近性重要度分数（recency importance score）r(q)。由于这些概率是各自独立评估的，一个查询可以同时激活多个图视图，而不必被归入某一个离散的检索意图。

当满足下式时，相应的关系类型即处于激活状态：

p_{g}(q)\geq\theta_{\mathrm{act}}   (12)

给定总的图扩展预算 B，Jev-Mem 按预测的各图需求分配搜索资源：

[¶28]

w_{g}(q)=\frac{p_{g}(q)^{\gamma}}{\sum_{j\in\mathcal{A}(q)}p_{j}(q)^{\gamma}}   (13)

where \mathcal{A}(q) denotes the active graphs and \gamma controls the concentration of the allocation. After assigning a feasible minimum budget m to active graphs, the remaining budget is distributed proportionally:

b_{g}=m+\operatorname{LRound}_{g}\left[(B-m|\mathcal{A}(q)|)w_{g}(q)\right]   (14)

The multi-hop prediction further determines the allowed traversal depth,

D(q)=\min\left\{D_{\max},\max\left(1,\left\lceil D_{\max}h(q)\right\rceil\right)\right\}   (15)

This probabilistic routing mechanism serves two purposes. It avoids spending equal retrieval effort on relations that are unlikely to help the current query, while still allowing multiple forms of evidence to be explored when a question requires them.

w_{g}(q)=\frac{p_{g}(q)^{\gamma}}{\sum_{j\in\mathcal{A}(q)}p_{j}(q)^{\gamma}}   (13)

其中 \mathcal{A}(q) 表示处于激活状态的图，\gamma 控制分配的集中程度。在为各激活图分配一个可行的最低预算 m 之后，剩余预算再按比例分配：

b_{g}=m+\operatorname{LRound}_{g}\left[(B-m|\mathcal{A}(q)|)w_{g}(q)\right]   (14)

多跳预测还进一步决定了允许的遍历深度，

D(q)=\min\left\{D_{\max},\max\left(1,\left\lceil D_{\max}h(q)\right\rceil\right)\right\}   (15)

这一概率路由机制有两重作用：一方面，它避免在不太可能对当前查询有帮助的关系上投入同等的检索资源；另一方面，当问题确实需要时，它仍允许探索多种形式的证据。

[¶29]

Anchor retrieval.

Before graph traversal, Jev-Mem identifies high-quality entry points using both semantic and lexical retrieval. Vector and keyword rankings are fused through reciprocal-rank fusion:

s_{\mathrm{RRF}}(v,q)=\sum_{L\in\{L_{\mathrm{vec}},L_{\mathrm{lex}}\}:v\in L}\frac{1}{\kappa+\operatorname{rank}_{L}(v)}   (16)

where \kappa=60. The highest-ranked nodes initialize the visited set and search frontier. Hybrid anchors provide robust starting points before the controller begins more expensive graph reasoning.

锚点检索

在图遍历之前，Jev-Mem 同时借助语义检索与词法检索来确定高质量的入口点。向量排名与关键词排名通过倒数排名融合（reciprocal-rank fusion，RRF）合并：

s_{\mathrm{RRF}}(v,q)=\sum_{L\in\{L_{\mathrm{vec}},L_{\mathrm{lex}}\}:v\in L}\frac{1}{\kappa+\operatorname{rank}_{L}(v)}   (16)

其中 \kappa=60。排名最高的若干节点用于初始化已访问集合与搜索前沿。在控制器开始代价更高的图推理之前，混合锚点为其提供了稳健的起点。

[¶30]

Evidence-guided expansion.

After each retrieval round d, Jev-Mem evaluates the current evidence set E_{d}. Instead of blindly continuing until a fixed graph depth or node count is reached, the controller estimates

\displaystyle s_{d}
\displaystyle:\text{evidence sufficiency},
\displaystyle u_{d}
\displaystyle:\text{expected utility of further retrieval},
\displaystyle m_{d}
\displaystyle:\text{missing required evidence},
\displaystyle c_{d}
\displaystyle:\text{unresolved contradiction}.   (17)

Retrieval terminates with sufficient evidence when

s_{d}\geq\theta_{\mathrm{suff}}\land m_{d}<\theta_{\mathrm{cont}}\land c_{d}<\theta_{\mathrm{cont}}   (21)

证据引导的扩展

在每轮检索 d 结束后，Jev-Mem 都会评估当前的证据集合 E_{d}。控制器不会盲目地一直检索到固定的图深度或节点数量为止，而是估计以下各量：

\displaystyle s_{d}
\displaystyle:\text{evidence sufficiency},
\displaystyle u_{d}
\displaystyle:\text{expected utility of further retrieval},
\displaystyle m_{d}
\displaystyle:\text{missing required evidence},
\displaystyle c_{d}
\displaystyle:\text{unresolved contradiction}.   (17)

当满足下式时，检索因证据充分而终止：

s_{d}\geq\theta_{\mathrm{suff}}\land m_{d}<\theta_{\mathrm{cont}}\land c_{d}<\theta_{\mathrm{cont}}   (21)

[¶31]

It may also terminate when

u_{d}<\theta_{\mathrm{cont}}   (22)

indicating that additional traversal is unlikely to improve the evidence even if the current evidence is not fully sufficient.

This stopping mechanism is important for both efficiency and retrieval quality. Stopping too early risks missing necessary evidence, whereas uncontrolled expansion introduces irrelevant memories that can distract downstream reasoning. Jev-Mem therefore explicitly models both evidence completeness and the expected benefit of further search.

When additional evidence is needed, Jev-Mem expands neighboring nodes under the relation-specific budgets b_{g} and global limits on nodes, edges, depth, controller calls, and latency. Each candidate v is then evaluated by the System-One controller along four complementary dimensions: query relevance a_{v}, relation usefulness \ell_{v}, information novelty n_{v}, and support for the current evidence c_{v}.

当满足下式时，检索也可以终止：

u_{d}<\theta_{\mathrm{cont}}   (22)

这表明即便当前证据尚不完全充分，继续遍历也不太可能改善证据。

这一停止机制对效率和检索质量都很重要。停止得过早可能会遗漏必要的证据，而不加控制的扩展则会引入无关记忆，干扰下游推理。因此，Jev-Mem 对证据的完整性与继续搜索的预期收益都进行了显式建模。

当需要更多证据时，Jev-Mem 会在按关系类型分配的预算 b_{g} 以及对节点数、边数、深度、控制器调用次数和延迟的全局限制下扩展相邻节点。随后，系统一控制器从四个互补的维度评估每个候选 v：查询相关性 a_{v}、关系有用性 \ell_{v}、信息新颖性 n_{v}，以及对当前证据的支持度 c_{v}。

[¶32]

These System-One predictions are combined with deterministic retrieval signals. For a candidate reached through graph type g, the transition score is

s(v\mid q,E_{d})=\frac{1}{\sum_{i=1}^{5}\lambda_{i}}\left[\begin{array}[]{l}\lambda_{1}z_{v}+\lambda_{2}a_{v}+\lambda_{3}p_{g}(q)\ell_{v}\\
+\lambda_{4}n_{v}+\lambda_{5}(\pi_{e}+c_{v})/2\end{array}\right]   (23)

where z_{v} denotes embedding similarity and \pi_{e} is the stored edge weight. If timestamps are available, the score is adjusted using the query-specific recency prediction:

\rho_{v}=\frac{1}{1+\max(0,\tau_{*}-\tau_{v})/\mathrm{day}}   (24)

\widetilde{s}(v)=\frac{s(v)+0.1r(q)\rho_{v}}{1+0.1r(q)}   (25)

这些系统一预测会与确定性检索信号相结合。对于经由图类型 g 到达的候选，其转移分数为

s(v\mid q,E_{d})=\frac{1}{\sum_{i=1}^{5}\lambda_{i}}\left[\begin{array}[]{l}\lambda_{1}z_{v}+\lambda_{2}a_{v}+\lambda_{3}p_{g}(q)\ell_{v}\\
+\lambda_{4}n_{v}+\lambda_{5}(\pi_{e}+c_{v})/2\end{array}\right]   (23)

其中 z_{v} 表示嵌入相似度，\pi_{e} 为存储的边权重。若时间戳可用，则利用针对该查询的新近性预测对分数进行调整：

\rho_{v}=\frac{1}{1+\max(0,\tau_{*}-\tau_{v})/\mathrm{day}}   (24)

\widetilde{s}(v)=\frac{s(v)+0.1r(q)\rho_{v}}{1+0.1r(q)}   (25)

[¶33]

The top W candidates form the next beam and are added to the accumulated evidence. Retrieval therefore follows an iterative feedback loop:

\begin{array}[]{c}\text{route}\rightarrow\text{retrieve}\rightarrow\text{assess}\\
\rightarrow\text{expand}\rightarrow\text{reassess}\end{array}   (26)

This process differs fundamentally from fixed top-k retrieval or static graph traversal. The relational views being explored, the allocation of search effort, the candidates retained at each round, and the decision to continue all depend on the query and on the evidence already collected. In this sense, retrieval itself becomes a System-One-controlled decision process.

得分最高的 W 个候选构成下一轮的束（beam），并被加入累积证据。因此，检索遵循一个迭代反馈回路：

\begin{array}[]{c}\text{route}\rightarrow\text{retrieve}\rightarrow\text{assess}\\
\rightarrow\text{expand}\rightarrow\text{reassess}\end{array}   (26)

这一过程与固定的 top-k 检索或静态图遍历有着根本区别。所探索的关系视图、搜索资源的分配、每一轮保留的候选，以及是否继续检索的决定，全都取决于查询本身和已经收集到的证据。从这个意义上说，检索本身成为了一个由系统一控制的决策过程。

[¶34]

Once retrieval terminates, the highest-scoring K memories are passed to the System-Two LLM:

y=\mathrm{SystemTwo}(q,E)   (27)

System Two performs the final synthesis over the selected evidence but is not involved in normal graph routing, candidate expansion, or stopping decisions. Jev-Mem therefore concentrates expensive autoregressive reasoning where it is most valuable while using fast structured control to improve the quality and efficiency of the evidence supplied to it.

The control overhead remains explicitly bounded. A normal write requires one batched typing request and, when candidates exist, one batched relation request over at most K_{w} memory pairs. A query requires one routing request and, for each retrieval round, at most one evidence assessment and one batched candidate-scoring request. Independent limits on graph expansions, inspected edges, visited nodes, depth, controller invocations, and elapsed time prevent the control process itself from growing without bound.

检索一旦终止，得分最高的 K 条记忆便被传给系统二 LLM：

y=\mathrm{SystemTwo}(q,E)   (27)

系统二基于选出的证据完成最终的综合，但不参与常规的图路由、候选扩展或停止决策。因此，Jev-Mem 把代价高昂的自回归推理集中用在最能发挥价值之处，同时借助快速的结构化控制，提升向其供给证据的质量与效率。

控制开销始终保持在明确的上界之内。一次常规写入只需一次批量的类型标注请求；若存在候选，再加上一次针对至多 K_{w} 个记忆对的批量关系请求。一次查询需要一次路由请求，并且在每一轮检索中至多需要一次证据评估和一次批量的候选打分请求。对图扩展次数、检查的边数、访问的节点数、深度、控制器调用次数和耗时分别设置的独立上限，可防止控制过程本身无限膨胀。

[¶35]

4 Experiments

We conduct comprehensive experiments to evaluate Jev-Mem along two dimensions: reasoning effectiveness and system efficiency. We examine whether System-One-controlled memory construction and retrieval improve long-horizon question answering, and whether separating fast memory control from System-Two reasoning reduces the overhead of building and accessing persistent memory. Accordingly, we report answer accuracy together with memory construction time and per-query latency.

4 实验

我们开展了全面的实验，从推理有效性与系统效率两个维度评估 Jev-Mem。我们考察两个问题：由系统一控制的记忆构建与检索能否提升长程问答的表现；将快速的记忆控制与系统二推理相分离，能否降低构建和访问持久记忆的开销。据此，我们报告答案准确率，以及记忆构建时间与单次查询延迟。

[¶36]

4.1 Experimental Setup

Datasets. We evaluate long-term conversational memory on two widely used benchmarks. LoCoMo (Maharana et al., 2024) contains ultra-long multi-session conversations and evaluates an agent’s ability to recover information requiring temporal, causal, and cross-session reasoning.

Baselines. We compare Jev-Mem with representative long-term memory approaches using the same backbone answer model whenever applicable.

• Full Context: provides the complete conversation history directly to the LLM without external memory.

• A-MEM (Xu et al., 2025): dynamically organizes and evolves memories based on agent experiences.

4.1 实验设置

数据集。我们在两个广泛使用的基准上评估长期对话记忆。LoCoMo (Maharana et al., 2024) 包含超长的多会话对话，用于评估智能体能否找回那些需借助时间推理、因果推理和跨会话推理才能获得的信息。

基线。我们将 Jev-Mem 与若干具有代表性的长期记忆方法进行比较；在适用的情况下，各方法均使用相同的骨干作答模型。

• Full Context：不使用外部记忆，直接将完整的对话历史提供给 LLM。

• A-MEM (Xu et al., 2025)：基于智能体的经验动态地组织并演化记忆。

[¶37]

• Nemori (Nan et al., 2025): structures conversational memory through episodic segmentation and graph-based retrieval.

• MemoryOS (Kang et al., 2025b): organizes persistent memory using a hierarchical multi-level architecture.

• MAGMA Jiang et al. (2026a): represents memories through multiple semantic, temporal, causal, and entity relations for structured retrieval.

• Jev-Mem (ours): uses a System-One controller to construct and adaptively retrieve from a multi-relational memory, while reserving the LLM for final reasoning and answer synthesis.

Metrics. We evaluate Jev-Mem from both accuracy and system-efficiency perspectives. For reasoning quality, we use the LLM-as-a-Judge score (Zheng et al., 2023), which measures whether the generated answer is correct with respect to the reference answer. For system efficiency, we report total memory construction time and average per-query latency. Memory construction time measures the end-to-end cost of building the persistent memory from the conversation history, while query latency measures the average time required to retrieve relevant evidence and produce an answer for each query. Together, these metrics capture both the effectiveness and runtime efficiency of the proposed memory architecture.

• Nemori (Nan et al., 2025)：通过情景分段与基于图的检索来组织对话记忆。

• MemoryOS (Kang et al., 2025b)：采用层次化的多级架构来组织持久记忆。

• MAGMA Jiang et al. (2026a)：通过语义、时间、因果与实体等多种关系来表示记忆，以支持结构化检索。

• Jev-Mem（本文方法）：利用系统一控制器构建多关系记忆，并从中自适应地检索，同时将 LLM 留作最终推理与答案综合之用。

评估指标。我们从准确率与系统效率两个角度评估 Jev-Mem。在推理质量方面，我们采用 LLM 裁判评分 (Zheng et al., 2023)，用以衡量生成的答案相对于参考答案是否正确。在系统效率方面，我们报告记忆构建总时间与平均单次查询延迟。记忆构建时间衡量的是依据对话历史构建持久记忆的端到端开销；查询延迟衡量的则是针对每个查询检索相关证据并生成答案所需的平均时间。这些指标共同刻画了所提出的记忆架构的有效性与运行时效率。

[¶38]

Table 1: Performance on the LoCoMo benchmark evaluated using the LLM-as-a-Judge metric. Higher scores indicate better performance. LLM model is based on gpt-4o-mini.

表 1：在 LoCoMo 基准上以 LLM 裁判指标评估的性能。分数越高表示性能越好。所用 LLM 基于 gpt-4o-mini。

[¶39]

Method	Multi-Hop	Temporal	Open-Domain	Single-Hop	Adversarial	Overall
Full Context	0.468	0.562	0.486	0.630	0.205	0.481
A-MEM	0.495	0.474	0.385	0.653	0.616	0.580
MemoryOS	0.552	0.422	0.504	0.674	0.428	0.553
Nemori	0.569	0.649	0.485	0.764	0.325	0.590
MAGMA	0.528	0.650	0.517	0.776	0.742	0.700
Jev-Mem	0.623	0.637	0.618	0.802	0.962	0.777

[¶40]

4.2 Overall Performance

Table 1 summarizes the performance of Jev-Mem and the baselines on LoCoMo. Jev-Mem achieves the highest overall LLM-as-a-Judge score of 0.777, compared with 0.700 for the strongest baseline, corresponding to an 11.0% relative improvement. The gains are also consistent across question types: Jev-Mem performs best in five of the six categories and matches the best result on temporal reasoning.

The improvement is especially clear on queries that require stronger retrieval and evidence selection. On Multi-Hop questions, Jev-Mem reaches 0.625, compared with 0.569 for the strongest baseline. On Open-Domain questions, it achieves 0.610, substantially higher than the best baseline score of 0.517. The largest gain appears in the Adversarial category, where Jev-Mem reaches 0.962, compared with 0.742 for the strongest baseline. Jev-Mem also achieves the highest Single-Hop score of 0.797 and matches the best Temporal score of 0.650.

4.2 整体性能

表 1 汇总了 Jev-Mem 与各基线在 LoCoMo 上的性能。Jev-Mem 取得了最高的总体 LLM 裁判评分 0.777，而最强基线为 0.700，相当于 11.0% 的相对提升。这种提升在不同问题类型上也保持一致：Jev-Mem 在六个类别中的五个上表现最佳，并在时间推理上与最佳结果持平。

在对检索与证据筛选要求更高的查询上，提升尤为明显。在多跳问题上，Jev-Mem 达到 0.625，而最强基线为 0.569。在开放域问题上，它取得 0.610，显著高于最佳基线的 0.517。提升幅度最大的是对抗类别：Jev-Mem 达到 0.962，而最强基线为 0.742。Jev-Mem 还取得了最高的单跳得分 0.797，并与最佳的时间类得分 0.650 持平。

[¶41]

These results indicate that Jev-Mem improves performance across a broad range of memory-intensive queries rather than specializing in a single reasoning pattern. The largest gains appear when the system must combine evidence across multiple memories or distinguish relevant information from plausible but misleading distractors. Jev-Mem addresses these cases through adaptive memory control: it estimates which relational views are useful for each query, allocates retrieval effort accordingly, evaluates candidate memories during traversal, and decides when the collected evidence is sufficient. As a result, retrieval is adapted to the needs of each query instead of following a fixed search strategy, allowing the reasoning model to operate on a more relevant and focused evidence set.

这些结果表明，Jev-Mem 在广泛的记忆密集型查询上都提升了性能，而不是只擅长某一种推理模式。当系统必须综合多条记忆中的证据，或需要把相关信息与看似合理却具有误导性的干扰项区分开来时，提升最为显著。Jev-Mem 通过自适应记忆控制来应对这些情形：它估计哪些关系视图对当前查询有用，据此分配检索投入，在遍历过程中评估候选记忆，并判断已收集的证据何时足够。因此，检索会依据每个查询的需要进行调整，而不是遵循固定的搜索策略，从而使推理模型能够基于更相关、更聚焦的证据集开展工作。

[¶42]

Table 2: System efficiency comparison with total memory build time (in seconds) and average query latency (in seconds).

表 2：系统效率对比，包括记忆构建总时间（秒）与平均查询延迟（秒）。

[¶43]

Method	Build Time (s)	Latency (s)
Full Context	N/A	1.74
A-MEM	3636	2.26
MemoryOS	3276	32.68
Nemori	1044	2.59
MAGMA	1404	1.47
Jev-Mem	158	0.93

[¶44]

4.3 Efficiency of System-One Memory Control

Table 2 compares the system efficiency of Jev-Mem with the baselines in terms of total memory construction time and average query latency. Jev-Mem achieves substantial improvements in both stages. Its total memory build time is only 158 seconds, compared with 1,044 seconds for the fastest competing memory system, corresponding to an 84.9% reduction, or a 6.6× speedup. The gap is even larger compared with A-MEM and MemoryOS, whose memory construction requires more than 3,000 seconds.

Jev-Mem also achieves the lowest query latency, requiring only 0.93 seconds per query on average. This is 36.7% lower than the fastest memory-based baseline at 1.47 seconds, and 46.6% lower than directly processing the full context at 1.74 seconds. In contrast, MemoryOS requires 32.68 seconds per query, highlighting the substantial runtime overhead that memory management can introduce when complex processing remains on the retrieval path.

4.3 系统一记忆控制的效率

表 2 从记忆构建总时间与平均查询延迟两方面，比较了 Jev-Mem 与各基线的系统效率。Jev-Mem 在这两个阶段都取得了大幅改进。其记忆构建总时间仅为 158 秒，而最快的对比记忆系统需要 1,044 秒，相当于缩短了 84.9%，即 6.6 倍的加速。与 A-MEM 和 MemoryOS 相比差距更大，二者的记忆构建均需 3,000 秒以上。

Jev-Mem 的查询延迟同样最低，平均每个查询仅需 0.93 秒。这比最快的基于记忆的基线（1.47 秒）低 36.7%，比直接处理完整上下文（1.74 秒）低 46.6%。相比之下，MemoryOS 每个查询需要 32.68 秒，这凸显出：当复杂处理仍留在检索路径上时，记忆管理可能引入可观的运行时开销。

[¶45]

The efficiency gains reflect the separation of memory control from autoregressive reasoning in Jev-Mem. During memory construction, lightweight typed decisions are batched over a bounded candidate set rather than repeatedly invoking a general-purpose LLM for free-form memory processing. During retrieval, System-One control performs query routing, candidate evaluation, and evidence checking, while graph exploration is constrained by explicit search budgets. Adaptive stopping further avoids unnecessary traversal once sufficient evidence has been collected. Together, these mechanisms reduce the computational overhead of both constructing and accessing long-term memory.

这些效率收益源于 Jev-Mem 将记忆控制与自回归推理相分离的设计。在记忆构建阶段，轻量的类型化决策在有界的候选集上批量执行，而不是反复调用通用 LLM 进行自由形式的记忆处理。在检索阶段，系统一控制负责查询路由、候选评估与证据核查，而图探索则受到显式搜索预算的约束。一旦收集到足够的证据，自适应停止机制还会进一步避免不必要的遍历。这些机制共同降低了长期记忆在构建与访问两方面的计算开销。

[¶46]

Notably, the efficiency improvement does not come at the expense of reasoning quality. As shown in Table 1, Jev-Mem simultaneously achieves the highest overall accuracy on LoCoMo. These results demonstrate that separating fast memory control from System-Two reasoning can improve both the effectiveness and the efficiency of agentic memory.

值得注意的是，效率的提升并未以牺牲推理质量为代价。如表 1 所示，Jev-Mem 同时在 LoCoMo 上取得了最高的总体准确率。这些结果表明，将快速的记忆控制与系统二推理相分离，能够同时提升智能体记忆的有效性与效率。

[¶47]

5 Conclusion

We present Jev-Mem, a System-One/System-Two architecture for efficient agentic memory. Jev-Mem uses lightweight structured prediction for high-frequency memory-control decisions and reserves System Two for complex reasoning and answer synthesis. On LoCoMo, Jev-Mem achieves the best overall accuracy and efficiency among the evaluated baselines, reaching an LLM-as-a-Judge score of 0.777, a memory construction time of 158 s, and an average query latency of 0.93 s. These results show that separating memory control from generative reasoning is a promising direction for building more effective and efficient long-horizon agents.

5 结论

本文提出 Jev-Mem，一种面向高效智能体记忆的系统一/系统二架构。Jev-Mem 使用轻量级的结构化预测来处理高频的记忆控制决策，而将系统二留给复杂推理与答案综合。在 LoCoMo 上，与所评估的各基线相比，Jev-Mem 取得了最佳的总体准确率与效率：LLM 裁判评分达到 0.777，记忆构建时间为 158 秒，平均查询延迟为 0.93 秒。这些结果表明，将记忆控制与生成式推理相分离，是构建更有效、更高效的长程智能体的一个有前景的方向。

[¶48]

Appendix A Appendix

附录 A 附录

[¶49]

A.1 Related Work

From memory storage to active memory management. Early work on agent memory established that long-term interaction requires more than retaining raw conversation history. Generative Agents (Park et al., 2023) maintain an experience stream and periodically synthesize higher-level reflections, while Reflexion (Shinn et al., 2023) stores linguistic feedback in episodic memory to improve future decisions. MemoryBank (Zhong et al., 2024) introduces long-term updating and forgetting, and MemGPT (Packer et al., 2023) manages information across different memory tiers through an operating-system-inspired virtual context. These systems shift memory from passive storage toward an active component that selects, transforms, and reuses past experience. A parallel line of work retains reusable behavior rather than conversational content: Voyager (Wang et al., 2023) accumulates a persistent skill library for an embodied agent, ExpeL (zhao2024expelllmagentsexperiential) distills transferable insights from past trajectories, and Agent Workflow Memory (wang2025workflowmemory) induces reusable workflows from prior action sequences. ReadAgent (Lee et al., 2024) takes a different route, compressing very long inputs into gist memories that are revisited on demand. Benchmarks such as LongMemEval (Wu et al., 2024), MemBench (Tan et al., 2025), and MemoryAgentBench (Hu et al., 2026) measure how well such systems retain and use information across long interactive horizons.

A.1 相关工作

从记忆存储到主动记忆管理。早期的智能体记忆研究已经确立：长期交互所需的远不止保留原始对话历史。Generative Agents (Park et al., 2023) 维护一条经验流，并定期综合出更高层次的反思；Reflexion (Shinn et al., 2023) 则将语言形式的反馈存入情景记忆，以改进后续决策。MemoryBank (Zhong et al., 2024) 引入了长期的更新与遗忘机制；MemGPT (Packer et al., 2023) 则借助受操作系统启发的虚拟上下文，在不同的记忆层级之间管理信息。这些系统使记忆从被动存储转变为一个主动组件，能够对过往经验进行选择、转换与复用。另一条平行的研究路线保留的是可复用的行为，而非对话内容：Voyager (Wang et al., 2023) 为具身智能体积累一个持久的技能库，ExpeL (zhao2024expelllmagentsexperiential) 从过往轨迹中提炼可迁移的洞见，Agent Workflow Memory (wang2025workflowmemory) 则从先前的动作序列中归纳出可复用的工作流。ReadAgent (Lee et al., 2024) 另辟蹊径，将超长输入压缩为要旨记忆（gist memory），并按需回看。LongMemEval (Wu et al., 2024)、MemBench (Tan et al., 2025) 与 MemoryAgentBench (Hu et al., 2026) 等基准则衡量此类系统在长程交互中保留并利用信息的能力。

[¶50]

Structured and evolving agentic memory. Recent work increasingly organizes memory into richer and more adaptive structures. A-MEM (Xu et al., 2025) constructs interconnected memory notes and allows their representations to evolve as new information arrives. Mem0 (Chhikara et al., 2025) dynamically extracts and consolidates salient information and further explores graph-based memory for relational structure. MemoryOS (Kang et al., 2025a) organizes information across short-, mid-, and long-term memory tiers, while Nemori (Nan et al., 2025) structures interactions into coherent episodes and distills higher-level knowledge from them. Graph-based approaches further capture relationships that cannot be represented by semantic similarity alone. MAGMA (Jiang et al., 2026a), for example, organizes memories through semantic, temporal, causal, and entity relations. HAGE (Jiang et al., 2026b) extends this direction with a weighted multi-relational graph in which query-conditioned routing and reinforcement learning jointly optimize relation features and traversal behavior, allowing retrieval paths to adapt to downstream reasoning objectives. Related structure also appears in retrieval systems that are not themselves agent-memory architectures: RAPTOR (Sarthi et al., 2024) builds recursive abstractive summary trees, GraphRAG (Edge et al., 2024) constructs entity graphs and community summaries to answer global queries, and LightRAG (Guo et al., 2025) pairs graph structure with incremental updates for efficiency. HippoRAG (Jimenez Gutierrez et al., 2024) and its continual-learning extension (Gutiérrez et al., 2025) make the link to memory more explicit, using graph-based association over a persistent non-parametric store. These systems share representational machinery with agentic memory, but target retrieval over a given corpus rather than the maintenance of an agent’s own accumulating experience, and they do not define the multi-relation semantics used here. Together, these systems show a broader shift from static memory stores toward structured, evolving, and query-adaptive memory representations.

结构化与持续演化的智能体记忆。近期工作越来越多地把记忆组织成更丰富、更具适应性的结构。A-MEM (Xu et al., 2025) 构建相互关联的记忆笔记，并允许其表示随新信息的到来而不断演化。Mem0 (Chhikara et al., 2025) 动态地抽取并整合关键信息，还进一步探索了用基于图的记忆来表达关系结构。MemoryOS (Kang et al., 2025a) 在短期、中期与长期三个记忆层级之间组织信息；Nemori (Nan et al., 2025) 则把交互组织成连贯的情景片段，并从中提炼更高层次的知识。基于图的方法进一步捕捉了仅凭语义相似度无法表示的关系。例如，MAGMA (Jiang et al., 2026a) 通过语义、时间、因果与实体关系来组织记忆。HAGE (Jiang et al., 2026b) 沿这一方向更进一步，采用加权的多关系图，由以查询为条件的路由与强化学习联合优化关系特征与遍历行为，使检索路径能够适应下游的推理目标。相关的结构也出现在一些本身并非智能体记忆架构的检索系统中：RAPTOR (Sarthi et al., 2024) 构建递归的抽象式摘要树；GraphRAG (Edge et al., 2024) 构建实体图与社区摘要，用以回答全局性查询；LightRAG (Guo et al., 2025) 则将图结构与增量更新相结合，以提升效率。HippoRAG (Jimenez Gutierrez et al., 2024) 及其持续学习扩展 (Gutiérrez et al., 2025) 与记忆的联系更为明确：它们在持久的非参数化存储之上进行基于图的联想。这些系统与智能体记忆共享表示层面的机制，但其目标是在给定语料上进行检索，而非维护智能体自身不断累积的经验，而且它们并未定义本文所用的多关系语义。总体而言，这些系统呈现出一个更广泛的转变：从静态的记忆存储，走向结构化、持续演化且随查询自适应的记忆表示。

[¶51]

Adaptive retrieval and memory access. Retrieval-augmented generation established external stores as a way to supplement parametric knowledge (Lewis et al., 2020; Guu et al., 2020; Izacard et al., 2022). Later work made the retrieval procedure itself adaptive: IRCoT (Trivedi et al., 2023) interleaves retrieval with chain-of-thought steps, FLARE (Jiang et al., 2023) triggers retrieval during generation when the model becomes uncertain, Self-RAG (Asai et al., 2023) learns when to retrieve and how to critique the evidence it obtains, and Adaptive-RAG (Jeong et al., 2024) selects among retrieval strategies according to question complexity. These methods control evidence acquisition within a single generation episode: whether, when, and how much to retrieve from a corpus that they do not modify. Jev-Mem shares the view of retrieval as a learned, query-dependent decision, but situates that control in the agent-memory lifecycle, where the store is itself written and reorganized by the controller and where the decisions also include what to admit, how new information relates to existing memory, which relational view to traverse, and when to stop.

自适应检索与记忆访问。检索增强生成（RAG）确立了以外部存储补充参数化知识的做法 (Lewis et al., 2020; Guu et al., 2020; Izacard et al., 2022)。后续工作使检索过程本身变得自适应：IRCoT (Trivedi et al., 2023) 将检索与思维链（chain-of-thought）步骤交替进行；FLARE (Jiang et al., 2023) 在生成过程中、当模型变得不确定时触发检索；Self-RAG (Asai et al., 2023) 学习何时检索，以及如何评判所获得的证据；Adaptive-RAG (Jeong et al., 2024) 则根据问题的复杂度在不同检索策略之间做出选择。这些方法是在单次生成过程之内控制证据的获取：是否检索、何时检索，以及从一个它们并不修改的语料中检索多少。Jev-Mem 同样将检索视为一种习得的、依赖于查询的决策，但把这种控制置于智能体记忆的生命周期之中：在这里，存储本身由控制器写入并重组，而决策还涵盖准入哪些内容、新信息与已有记忆如何关联、遍历哪一个关系视图，以及何时停止。

[¶52]

Efficiency of agent memory. As memory architectures become more sophisticated, the computation required to construct, maintain, and retrieve memory has emerged as an equally important concern. SimpleMem (Liu et al., 2026) improves efficiency through semantic compression, asynchronous consolidation, and query-adaptive retrieval. LightMem (Fang et al., 2026) similarly separates lightweight online processing from more expensive memory consolidation and reduces the amount of computation required during interaction. More aggressively, Zero-Mem (Xiao et al., 2026) removes LLM generation from intermediate memory operations and retrieves directly over structured representations. These works demonstrate that memory quality and memory cost must be considered jointly rather than treating memory management as negligible compared with final answer generation. A complementary body of systems work reduces LLM inference cost directly, through model cascades and learned routers that dispatch each request to a model of appropriate capability (Chen et al., 2024; Ong et al., 2025). These concerns are distinct: lowering the cost of the memory pipeline, choosing which model answers a request, and controlling the fine-grained decisions taken inside the memory lifecycle. Jev-Mem targets the last of these.

智能体记忆的效率。随着记忆架构日趋复杂，构建、维护与检索记忆所需的计算已成为一个同等重要的问题。SimpleMem (Liu et al., 2026) 通过语义压缩、异步整合与随查询自适应的检索来提升效率。LightMem (Fang et al., 2026) 同样把轻量的在线处理与代价更高的记忆整合分离开来，从而减少交互过程中所需的计算量。Zero-Mem (Xiao et al., 2026) 的做法更为激进：它将 LLM 生成从中间记忆操作中移除，直接在结构化表示上进行检索。这些工作表明，记忆质量与记忆成本必须统筹考虑，而不能把记忆管理视为相对于最终答案生成可以忽略不计的环节。另有一类与之互补的系统研究直接降低 LLM 的推理成本，其手段是模型级联（model cascade）以及学习型路由器——后者将每个请求分派给能力相称的模型 (Chen et al., 2024; Ong et al., 2025)。这几类关注点各不相同：降低记忆流水线的成本、选择由哪个模型来回答请求，以及控制记忆生命周期内部所做出的细粒度决策。Jev-Mem 针对的是其中最后一类。

[¶53]

System-One and System-Two computation. The terminology we adopt originates in dual-process accounts of human reasoning, which distinguish fast, automatic processing from slower and more effortful deliberation (Evans, 2008). We use this distinction as an analogy for how computation is allocated, not as a claim that the underlying mechanisms correspond. Analogous separations have been explored in LLM systems: System 2 Attention (Weston and Sukhbaatar, 2023) inserts an explicit reconsideration step before answering, while Tree of Thoughts (Yao et al., 2023a) and Graph of Thoughts (Besta et al., 2024) spend additional structured search at test time on problems that reward deliberation. These methods allocate deliberative computation within a single reasoning episode. Jev-Mem instead applies the distinction architecturally, to the memory subsystem itself.

系统一与系统二计算。本文采用的术语源自人类推理的双过程理论，该理论将快速、自动的加工与更缓慢、更费力的审慎思考区分开来 (Evans, 2008)。我们仅把这一区分用作计算分配方式的类比，并不主张二者的底层机制彼此对应。LLM 系统中也已有工作探索过类似的划分：System 2 Attention (Weston and Sukhbaatar, 2023) 在作答之前插入一个显式的重新审视步骤，而 Tree of Thoughts (Yao et al., 2023a) 与 Graph of Thoughts (Besta et al., 2024) 则在测试时针对那些审慎思考能带来收益的问题投入额外的结构化搜索。这些方法是在单次推理过程内部分配审慎计算；Jev-Mem 则在架构层面应用这一区分，并将其作用于记忆子系统本身。

[¶54]

Memory control as a systems abstraction. Jev-Mem builds on these developments but focuses on a different architectural question: how should the decisions that govern memory be executed? Existing work has primarily improved what is stored, how memories are organized, or how retrieval is performed. Jev-Mem instead treats memory control itself as a first-class systems plane. Inspired by the System-One/System-Two distinction, it separates fast, structured memory decisions from deliberative language generation. A System-One controller governs both sides of the memory lifecycle—including memory typing, relation construction, query routing, retrieval-budget allocation, candidate scoring, evidence assessment, and adaptive stopping—over a shared structured memory plane, while System Two is reserved for synthesis and complex reasoning. This design provides a common control abstraction across memory construction and retrieval, targeting both reasoning quality and system efficiency rather than optimizing either stage in isolation.

作为系统抽象的记忆控制。Jev-Mem 建立在上述进展的基础之上，但聚焦于一个不同的架构问题：支配记忆的各项决策应当以何种方式执行？现有工作主要改进的是存储什么、如何组织记忆，或如何执行检索；Jev-Mem 则把记忆控制本身视为一个一等的系统平面。受系统一与系统二之分的启发，它将快速、结构化的记忆决策与审慎的语言生成分离开来。系统一控制器在共享的结构化记忆平面之上统管记忆生命周期的两端——涵盖记忆类型标注、关系构建、查询路由、检索预算分配、候选打分、证据评估与自适应停止——而系统二则专门用于综合与复杂推理。这一设计为记忆构建与检索提供了统一的控制抽象，同时瞄准推理质量与系统效率，而非孤立地优化其中任一阶段。

[¶55]

Appendix B Jev Prompts for Memory Writing and Retrieval

This section describes the typed prompts used by Jev-Mem to construct and retrieve conversational memory. The quoted instructions and criteria are reproduced from the implementation, with candidate index 0 instantiated where applicable. The memories below are fictional and illustrate the request structure; no scores or retrieval outcomes are presented as measured Jev predictions.

附录 B 用于记忆写入与检索的 Jev 提示词

本节介绍 Jev-Mem 在构建与检索对话记忆时所使用的类型化提示词。文中引用的指令与判据均摘录自实际实现，并在适用之处将候选索引实例化为 0。下文中的记忆均为虚构，仅用于说明请求的结构；本节不将任何分数或检索结果呈现为实测的 Jev 预测。

[¶56]

B.1 Typed interface and request state

Jev-Mem submits a shared state object and a batch of typed questions through TypeSafeClient.system_one. Each Noul specifies a binary proposition using an instruction and explicit true and false criteria. Its returned value lies in [0, 1]. These model-reported values are not assumed to be calibrated probabilities. Multiple Nouls are evaluated as independent propositions rather than mutually exclusive labels; for example, an observation may be both episodic and semantic. Independence here concerns the question formulation, not statistical independence of the returned scores.

Question identifiers are bookkeeping keys and are not themselves model input. Consequently, every instruction names the state fields it uses. For candidate index i, relation and traversal keys are prefixed with pair_i_ and candidate_i_, respectively. Questions in a batch do not consume one another’s answers. Choice is used when alternatives are mutually exclusive, as in the consolidation representation decision below; it returns a selected label and an option distribution.

B.1 类型化接口与请求状态

Jev-Mem 通过 TypeSafeClient.system_one 提交一个共享的 state 对象和一批类型化的 questions。每个 Noul（Jev 的概率型真/假输出，取值 0–1）借助一条指令以及显式的 true 与 false 判据来界定一个二元命题，其返回值位于 [0, 1] 区间内。我们并不假定这些由模型报告的数值是经过校准的概率。多个 Noul 被当作相互独立的命题分别评估，而非互斥的标签；例如，一条观测可以同时属于情景型与语义型。此处的「独立」针对的是问题的表述方式，而非返回分数在统计意义上的独立性。

问题标识符只是用于簿记的键，其本身并不作为模型输入。因此，每条指令都会写明其所使用的状态字段。对于候选索引 i，关系键与遍历键分别以 pair_i_ 和 candidate_i_ 为前缀。同一批次中的问题不会使用彼此的答案。当备选项相互排斥时（如下文整合环节中的表示形式决策），则使用 Choice；它返回所选中的标签以及各选项上的分布。

[¶57]

Our running example contains memory m1, observed on 14 May 2024 at 10:00: "Mira: My old bicycle broke.", and memory m2, observed on 16 May 2024 at 10:00: "Mira: I bought a new bicycle yesterday because my old one broke." Both have entity identifiers Mira and bicycle. The query is "When did Mira buy a new bicycle, and why?" Memory objects passed to Jev contain id, content, timestamp, and entities; the timestamp records when the statement was observed, rather than necessarily when the event occurred.

我们的贯穿示例包含两条记忆：记忆 m1 于 2024 年 5 月 14 日 10:00 被观测到，内容为「Mira：我的旧自行车坏了」；记忆 m2 于 2024 年 5 月 16 日 10:00 被观测到，内容为「Mira：我昨天买了一辆新自行车，因为旧的那辆坏了」。两者都带有实体标识符 Mira 和 bicycle。所用查询为：「Mira 是什么时候买的新自行车？为什么？」传给 Jev 的记忆对象包含 id、content、timestamp 和 entities；其中时间戳记录的是该陈述被观测到的时间，而不一定是事件实际发生的时间。

[¶58]

B.2 Memory writing

The active profile disables admission filtering (admission_enabled=false). Every valid nonempty observation therefore enters the write path; memory-type scores do not decide whether it is retained. The typing request contains state = {"observation": text} and four Nouls: episodic, semantic, procedural, and preference. The following two templates illustrate event and preference typing.

B.2 记忆写入

当前启用的配置关闭了准入过滤（admission filtering；admission_enabled=false）。因此，每条有效的非空观测都会进入写路径；记忆类型分数并不决定该观测是否被保留。类型标注请求包含 state = {"observation": text} 以及四个 Noul：episodic、semantic、procedural 和 preference。以下两个模板分别示例事件类型标注与偏好类型标注。

[¶59]

Event typing

episodic (Noul).

Instruction: Does observation describe a particular experience or event involving a participant?

true: A specific past, current or planned event, even if its exact time is unstated.

false: Only a general fact, procedure or preference with no particular event.

事件类型标注

episodic（Noul）。

指令： observation 是否描述了涉及某位参与者的一次特定经历或事件？

true： 一个具体的过去、当前或计划中的事件，即便其确切时间未被说明。

false： 仅为一般性的事实、流程或偏好，不涉及任何特定事件。

[¶60]

Preference typing

preference (Noul).

Instruction: Does observation express a participant’s preference, aversion or habitual choice?

true: An attributable like, dislike, preferred option or habitual choice.

false: An isolated action alone, another person’s unattributed preference, or no preference evidence.

After typing, deterministic vector, keyword, entity and timestamp signals select at most 10 existing candidate memories. A relation request contains new_memory and a candidates list. For the running example, new_memory is m2 and candidates[0] is m1. The request batches semantic association and both causal directions; entity alias resolution is added only when exact entity identifiers do not already match.

偏好类型标注

preference（Noul）。

指令： observation 是否表达了某位参与者的偏好、反感或习惯性选择？

true： 可归属于具体参与者的喜好、厌恶、偏爱的选项或习惯性选择。

false： 仅有一次孤立的行为、未明确归属的他人偏好，或不存在任何偏好证据。

完成类型标注后，由确定性的向量、关键词、实体与时间戳信号选出至多 10 条已有的候选记忆。关系请求包含 new_memory 和一个 candidates 列表。在贯穿示例中，new_memory 为 m2，candidates[0] 为 m1。该请求将语义关联与两个方向的因果关系合并为一批进行判断；只有当精确的实体标识符尚未匹配时，才会额外加入实体别名消解。

[¶61]

Semantic relation

semantic (Noul).

Instruction: Compare new_memory.content with candidates[0].content. Would a semantic link between these observations help retrieve a shared specific topic or fact?

true: A specific shared topic, fact or event makes the connection useful.

false: Only generic conversational vocabulary or no meaningful semantic connection.

语义关系

semantic（Noul）。

指令： 比较 new_memory.content 与 candidates[0].content。在这两条观测之间建立语义链接，是否有助于检索二者共有的某个具体话题或事实？

true： 某个共有的具体话题、事实或事件使这一连接有用。

false： 仅有泛泛的对话用语，或不存在有意义的语义联系。

[¶62]

Directed causal relation

caused_by (Noul).

Instruction: Compare new_memory.content with candidates[0].content. Does the candidate event cause, enable or explain the event in new_memory.content?

true: The supplied accounts support this direction of causal influence.

false: Only similarity, chronology, a shared entity, or insufficient causal evidence.

A returned relation score of at least 0.60 creates the corresponding typed edge. For caused_by, the direction is candidate to new memory; causes tests the reverse direction. The causal criteria explicitly distinguish causal support from shared topics or mere chronology. Exact entity intersections are linked deterministically. In the example, the shared entity identifiers therefore make an alias-resolution prompt unnecessary.

有向因果关系

caused_by（Noul）。

指令： 比较 new_memory.content 与 candidates[0].content。候选中的事件是否导致、促成或解释了 new_memory.content 中的事件？

true： 所提供的记述支持这一方向上的因果影响。

false： 仅存在相似性、时间先后、共同实体，或因果证据不足。

返回的关系分数不低于 0.60 时，即创建相应的类型化边。对于 caused_by，方向为从候选记忆指向新记忆；causes 则检验相反方向。因果判据明确地把因果支持与共同话题或单纯的时间先后区分开来。精确的实体交集以确定性方式建立链接。因此在本例中，由于二者共享实体标识符，便无需再调用别名消解提示词。

[¶63]

B.3 Periodic consolidation

Every 20 successful Jev writes, the active profile evaluates candidate pairs for consolidation. Four Nouls assess redundancy, contradiction, obsolescence and link usefulness; the following Choice selects a representation. This is a separate post-insertion decision, not an admission filter.

B.3 周期性整合

当前启用的配置每完成 20 次成功的 Jev 写入，就会对候选记忆对进行一次整合评估。四个 Noul 分别评估冗余性、矛盾、过时性与链接有用性；随后由下面的 Choice 选择一种表示形式。这是一个在插入之后单独做出的决策，而非准入过滤器。

[¶64]

Representation choice

representation (Choice).

Instruction: Compare new_memory.content with candidates[0].content. Which representation best fits the relationship between these two observations? Judge from the supplied accounts; do not assume answers to other questions.

keep_separate: Contradictory accounts, unique details that a combined representation would lose, or distinct facts/events without a supported general pattern.

merge: Compatible accounts of the same fact or event can be combined without losing unique details.

promote: Distinct repeated episodes explicitly support a stable general pattern suitable for semantic abstraction; prefer this over merge for repeated events.

表示形式选择

representation（Choice）。

指令： 比较 new_memory.content 与 candidates[0].content。哪种表示形式最契合这两条观测之间的关系？请依据所提供的记述作出判断，不要预设其他问题的答案。

keep_separate： 记述相互矛盾；存在合并表示会丢失的独特细节；或者属于彼此不同的事实/事件，且不存在有证据支撑的一般模式。

merge： 关于同一事实或事件的相容记述，可以在不丢失独特细节的前提下合并。

promote： 多个不同的重复情景明确支持一个稳定的一般模式，适合进行语义抽象；对于重复发生的事件，应优先选择此项而非 merge。

[¶65]

uncertain: Insufficient evidence to choose a safe combined or separate representation.

Consolidation preserves raw observations. The normal periodic path records decisions and links; it does not automatically replace source memories. A caller-supplied System-Two summarizer can create a new representation only when merge or promote is selected with probability at least 0.85 and the contradiction score is below 0.85. The selected-option probability, rather than a separate confidence summary, controls this threshold.

uncertain： 证据不足，无法稳妥地选择合并或分开的表示形式。

整合会保留原始观测。常规的周期性路径只记录决策与链接，不会自动替换源记忆。仅当 merge 或 promote 以不低于 0.85 的概率被选中、且矛盾分数低于 0.85 时，由调用方提供的系统二摘要器才能生成新的表示。控制这一阈值的是所选选项的概率，而不是另行给出的置信度汇总。

[¶66]

B.4 Retrieval routing and candidate scoring

The routing request contains only query. Six Nouls estimate semantic, temporal, causal and entity needs, together with multi_hop_need and recency_importance. The following templates illustrate time-sensitive and explanatory retrieval. A query may activate several graph types simultaneously.

B.4 检索路由与候选打分

路由请求只包含 query。六个 Noul 分别估计语义、时间、因果与实体方面的需求，以及 multi_hop_need 和 recency_importance。以下模板分别示例时间敏感型检索与解释型检索。一个查询可以同时激活多种图类型。

[¶67]

Temporal routing

temporal (Noul).

Instruction: Does answering query require event dates, durations, ordering or changes over time?

true: A time relation is needed to answer correctly.

false: Dates or ordering are incidental to the answer.

时间路由

temporal（Noul）。

指令： 回答 query 是否需要事件的日期、持续时长、先后顺序或随时间发生的变化？

true： 需要借助某种时间关系才能正确作答。

false： 日期或先后顺序对答案而言只是附带信息。

[¶68]

Causal routing

causal (Noul).

Instruction: Does answering query require explaining a cause, motivation, enabling condition or effect?

true: Causal or explanatory evidence is needed.

false: Only factual association or chronology is requested.

The active profile uses a total graph-expansion budget of 80. Graphs with positive need scores of at least 0.10 receive a minimum allocation of one; the remaining budget is distributed in proportion to the active need scores, with largest-remainder rounding. The probability exponent is 1.0. Vector and keyword search provide anchors before graph expansion; the budget does not imply one provider request per edge.

因果路由

causal（Noul）。

指令： 回答 query 是否需要解释某个原因、动机、促成条件或结果？

true： 需要因果性或解释性的证据。

false： 仅要求事实关联或时间先后。

当前启用的配置所使用的图扩展总预算为 80。需求分数为正且不低于 0.10 的图获得最低配额 1；剩余预算按各活跃图的需求分数成比例分配，并采用最大余数法取整。概率指数取 1.0。在图扩展之前，由向量检索与关键词检索提供锚点；该预算并不意味着每条边都要向服务提供方发起一次请求。

[¶69]

A traversal request contains query, currently selected evidence, and proposed candidates. Each candidate includes its memory fields, graph type, relation properties, and source/target identifiers so that direction is explicit. For illustration, evidence can contain m1 and the candidate can be m2, reached through a causal edge from m1 to m2. This is a schematic traversal state, not a claim that this tiny example necessarily requires expansion in an actual run.

遍历请求包含 query、当前已选中的 evidence 以及拟议的 candidates。每个候选都附带其记忆字段、图类型、关系属性以及源/目标标识符，从而使方向明确。举例而言，证据中可以包含 m1，候选则可以是 m2——它经由一条从 m1 指向 m2 的因果边到达。这只是一个示意性的遍历状态，并不表示这个小例子在实际运行中必然需要扩展。

[¶70]

Candidate relevance

relevance (Noul).

Instruction: Does candidates[0].content contain a fact needed to answer query?

true: Direct answer evidence or a necessary intermediate fact.

false: Only topic overlap or unrelated content.

候选相关性

relevance（Noul）。

指令： candidates[0].content 是否包含回答 query 所需的事实？

true： 可直接支撑答案的证据，或必要的中间事实。

false： 仅有话题上的重叠，或内容无关。

[¶71]

Additional evidence

new_information (Noul).

Instruction: Does candidates[0].content add an answer-relevant detail absent from evidence?

true: A distinct relevant detail or missing reasoning link.

false: Only duplicated evidence or irrelevant new details.

Two further Nouls assess relation_usefulness and supports_current_evidence. The implementation combines these four scores with query-to-memory cosine similarity, the need score of the traversed graph, and the stored edge probability. A recency adjustment uses recency_importance. Candidates are ranked with this combined score and the beam width is 10; a high relevance score alone is not an unconditional admission to the retrieved evidence.

新增证据

new_information（Noul）。

指令： candidates[0].content 是否补充了 evidence 中缺失的、与答案相关的细节？

true： 一个有别于现有证据的相关细节，或一个缺失的推理环节。

false： 仅为重复的证据，或无关的新细节。

另有两个 Noul 分别评估 relation_usefulness 与 supports_current_evidence。实现中将这四个分数与查询到记忆的余弦相似度、所遍历图的需求分数以及已存储的边概率组合在一起，时近性调整则使用 recency_importance。候选按这一综合分数排序，束宽（beam width）为 10；仅凭较高的相关性分数，候选并不能无条件地进入检索证据。

[¶72]

B.5 Evidence-based stopping

The stopping request contains query, the currently selected top-k evidence, and retrieval depth. It evaluates evidence_sufficient, continue_useful, missing_evidence, and contradiction. The following templates distinguish answer support from the prospective usefulness of another retrieval round.

B.5 基于证据的停止

停止请求包含 query、当前选中的 top-k evidence 以及检索深度 depth。它评估 evidence_sufficient、continue_useful、missing_evidence 与 contradiction。以下模板将「对答案的支撑」与「再进行一轮检索的预期效用」区分开来。

[¶73]

Answer sufficiency

evidence_sufficient (Noul).

Instruction: Does evidence contain support for every factual part of an answer to query?

true: A grounded answer can be given from these memories without inventing missing facts.

false: Any required fact or reasoning link is unsupported; related topics alone are insufficient.

答案充分性

evidence_sufficient（Noul）。

指令： evidence 是否为 query 的答案中每一个事实性部分都提供了支撑？

true： 无需编造缺失的事实，仅凭这些记忆即可给出有依据的答案。

false： 存在任何一项缺乏支撑的必需事实或推理环节；仅有相关话题并不充分。

[¶74]

Value of continued retrieval

continue_useful (Noul).

Instruction: Given query and evidence, is another retrieval round likely to fill a specific gap or resolve a conflict?

true: An identifiable missing fact or conflict could benefit from more memory retrieval.

false: No identifiable retrieval need remains or more memories are unlikely to help.

For temporal queries, evidence and candidate objects also include timestamp_role and temporal_references. These fields identify the timestamp as observation time and attach grounded expressions with their precision. Thus, sufficiency is judged with the same temporal grounding used to present the selected evidence to the answerer, rather than by treating every conversation date as an event date.

继续检索的价值

continue_useful（Noul）。

指令： 给定 query 与 evidence，再进行一轮检索是否有可能填补某个具体缺口或化解某个冲突？

true： 存在可识别的缺失事实或冲突，能从更多的记忆检索中获益。

false： 已不存在可识别的检索需求，或更多的记忆不太可能有所帮助。

对于时间类查询，证据对象与候选对象还会包含 timestamp_role 和 temporal_references。这些字段把时间戳标明为观测时间，并附上经过时间锚定的表达式及其精度。因此，充分性判断所采用的时间锚定，与向作答方呈现所选证据时所用的时间锚定一致，而不是把每个对话日期都当作事件日期。

[¶75]

Evidence-based stopping requires evidence_sufficient to be at least 0.95 and both missing_evidence and contradiction to be below 0.15. Retrieval can also terminate when continue_useful is below 0.15. Independent limits bound depth (8), visited nodes (60), examined edges (2400), Jev attempts (16), and the retrieval time budget (15 seconds). These limits can terminate retrieval even if evidence is incomplete; the time budget is checked between operations and is not a strict preemption guarantee for local computation.

Jev controls evidence selection but does not generate the final free-form answer. The selected memories are passed to a separate System-Two language model. This separation makes the prompt examples above memory-control decisions, rather than answer-generation or evaluator prompts. Gold answers and benchmark evidence annotations are excluded from these Jev states.

基于证据的停止要求 evidence_sufficient 不低于 0.95，且 missing_evidence 与 contradiction 均低于 0.15。当 continue_useful 低于 0.15 时，检索同样可以终止。此外还有几项相互独立的上限，分别约束深度（8）、已访问节点数（60）、已检查边数（2400）、Jev 尝试次数（16）以及检索时间预算（15 秒）。即使证据尚不完整，这些上限也可能终止检索；时间预算只在各操作之间检查，并不能对本地计算提供严格的抢占保证。

Jev 负责控制证据选择，但并不生成最终的自由形式答案。选出的记忆会被传给一个独立的系统二语言模型。这种分离使得上文的提示词示例属于记忆控制决策，而非答案生成提示词或评估器提示词。这些 Jev 状态中不包含标准答案与基准的证据标注。