Why MobileMem
MobileMem：面向持续演进智能体的端侧记忆
MobileMem: On-Device Memory for Continually Evolving Agents
MobileMem 作为用户的个人记忆智能体，能够自动理解用户意图，执行跨时间的人脸识别以关联多年间女儿的身份，从海量照片档案中定位 2016 年的生日场景，并返回一家三口围坐在生日蛋糕旁的照片。
MobileMem acts as the user's personal memory agent, automatically understanding user intent, performing cross-time facial recognition to associate the daughter's identity across years, locating the 2016 birthday scene from a vast photo archive, and returning the photo of the family of three gathered around the birthday cake.
Could you find the photo from my daughter's 7th birthday?
User
Daughter
User
Daughter
Got it. Is this the one?
That's it, thank you.
MobileMem Features
真实的移动环境不仅包括人与助手的对话，也包括复杂的第三方应用生态系统。用户在这些应用中的交互在格式和语义上本身就是异构的。MobileMem 被设计为反映这一现实，要求记忆系统处理来自多样应用源的实时消息流，而不是依赖单一同质的交互流。
Real-world mobile environments include not only human-assistant dialogues but also a complex ecosystem of third-party applications. User interactions across these apps are inherently heterogeneous in both format and semantics. MobileMem is designed to reflect this reality by requiring memory systems to process real-time message streams originating from diverse application sources, rather than relying on a single homogeneous interaction stream.
在真实移动场景中，用户大部分时间都在与各类应用交互，而不是直接与 AI 助手互动。为确保无缝的用户体验，配备长期记忆系统的助手必须以非侵入方式运行，通过被动观察学习用户活动，而不是打断用户工作流。
In real-world mobile scenarios, users spend the majority of their time interacting with various applications rather than directly engaging with an AI assistant. To ensure a seamless user experience, an assistant equipped with a long-term memory system must operate in a non-intrusive manner, learning from the user's activities through passive observation rather than interrupting their workflow.
在真实移动环境中，用户交互天然包含大量敏感个人信息。从多个应用收集原始、未处理日志会带来隐私风险，包括潜在的重新识别和机密个人属性暴露。为缓解这些问题，MobileMem 采用现实引导的构建范式，充分保护用户隐私和安全。
In real-world mobile environments, user interactions naturally contain a wide range of sensitive personal information. Collecting raw, unprocessed logs from multiple apps poses privacy risks, including potential re-identification and the exposure of confidential personal attributes. To mitigate these concerns, MobileMem adopts a realism-guided construction paradigm fully protecting user privacy and security.
Dataset Construction
如下所示，MobileMem 具有真实用户参与、多模态交互和多源内容，并通过多阶段管线构建。
As illustrated below, MobileMem, featuring real-user participation, multimodal interactions, and multi-source content, is established through a multi-stage pipeline.
KEME 将零散的用户-App 会话作为基础知识锚点，反映已经发生且必须保留的内容。在用户画像知识和时间约束的引导下，KEME 将这些锚定会话分层组织成统一的交互流。
KEME treats the fragmented user-app sessions as foundational knowledge anchors that reflect what has already occurred and must be preserved. Guided by user persona knowledge and temporal constraints, KEME hierarchically organizes these anchored sessions into a unified interaction stream.
从人物画像根节点构建时序事件图，并递归展开：粗粒度生活阶段 -> 更细的子事件图 -> 叶级会话，并按拓扑顺序展开。
Builds a temporal event graph from the persona root and recursively expands it: coarse life phases →finer sub-event graphs →leaf-level sessions, expanded in topological order.
将每个锚定的 App 会话分配到时间兼容的事件节点；若没有节点适配，则修订事件图，使每个锚点都可落地，并向下传播非矛盾约束。
Assigns each anchored app session to a time-compatible event node; when none fits, it revises the graph so every anchor is groundable, propagating a non-contradiction constraint downward.
在叶节点采纳或合并锚定内容，或合成上下文一致的人机对话，并通过版本跟踪属性和证据链接驱动人物画像演化。
At leaf nodes it adopts or merges anchored content, or synthesizes context-consistent human-assistant dialogue, and drives persona evolution with version-tracked attributes and evidence links.
每次展开后，它会更新剩余图，增添、移除或调整未来事件与边，以解决不一致并丰富结构，形成自底向上的反馈循环。
After each expansion it updates the remaining graph —adding, removing or adjusting future events and edges to resolve inconsistencies and enrich structure, forming a bottom-up feedback loop.
为平衡真实性、多样性与可扩展性，MobileMem 采用混合方式构建人物画像：一半来自代表性受雇参与者并参考真实参与者数据，另一半由这些真实样例生成虚拟画像。统一 Schema 进一步组织基础属性、上一年状态和关键里程碑，形成核心记忆结构，共 8 个真实画像和 8 个虚拟画像。
To balance authenticity, diversity, and scalability, MobileMem builds personas through a hybrid approach: half are derived from representative hired participants that reference real participant data, and half are virtual personas generated from those real examples. A unified Schema then organizes basic attributes, previous-year status, and key milestones into the core memory structure —8 real and 8 virtual personas in total.
每个人物画像由三部分表示：基础信息向量 B 捕捉核心身份与背景，初始状态向量 S 捕捉交互前的状态和日常模式，个人知识图谱 G 则由社会关系构建。G 以人物为中心，以频繁且重要的联系人作为周围节点；每个节点描述由 GPT-5.1 丰富，并通过 Seedream 文生图模型转化为参考照片。
Each persona is represented by three components: a basic information vector B capturing core identity and background, an initial state vector S capturing status and daily patterns before interaction, and a personal knowledge graph G built from social relationships. G centers on the persona, with frequent and significant contacts as surrounding nodes; every node description is enriched by GPT-5.1 and turned into reference photos via the Seedream text-to-image model.
在人物知识图谱驱动下，MobileMem 分三阶段生成构成每条轨迹主干的事件：重要日期收集锚定个人和文化上重要的日期；事件构建将其扩展为个性化的一年事件序列；步骤拆解把每个事件分解为按时间排序的子事件步骤，并为每一步标注涉及的移动应用和图像类型。
Driven by the persona knowledge graph, MobileMem generates the events that form the backbone of each trajectory in three stages: Important Date Collection anchors personally and culturally significant dates; Event Building expands them into a personalized one-year event sequence; and Step Breakdown decomposes every event into temporally ordered sub-event steps, each annotated with the mobile apps and image types involved.
事件步骤通过两个阶段变成具体的多模态对话。记忆点拆解将每个步骤分解为细粒度文本和视觉记忆点；随后会话生成让 GPT-5.1 作为 agent，将它们编织成自然对话轮次和匹配图像，这些图像通过 HTML 渲染、文生图和图像编辑工具生成，并用参考照片保持人物一致性。会话按时间顺序拼接成完整轨迹。
Event steps become concrete multimodal dialogues in two phases. Memory Point Decomposition breaks each step into fine-grained textual and visual memory points; Session Generation then lets GPT-5.1 act as an agent to weave them into natural dialogue turns and matching images —produced via HTML rendering, text-to-image, and image-editing tools, with reference photos keeping characters consistent. Sessions are concatenated in temporal order into the full trajectory.
为系统评估记忆能力，评测问题基于生成的记忆点和步骤信息构建，覆盖七类：单跳、多跳、知识更新、时间推理、隐式偏好、拒答和视觉推理。
To systematically assess memory systems, evaluation questions are built from the generated memory points and step information. They span seven categories: Single-Hop, Multi-Hop, Knowledge Update, Temporal Reasoning, Implicit Preference, Abstention, and Visual Reasoning.
Application
Application
MobileMem 提供统一的基准测试框架，能够与现有移动智能体生态系统无缝集成，包括任务理解、记忆检索问答、网页搜索、响应安全和执行模块。
MobileMem provides a unified benchmarking framework that seamlessly integrates with the existing mobile agent ecosystem, including task understanding, memory retrieval QA, web search, response safety, and execution modules.
通过与记忆中间件和持续学习框架对齐，它自然适配真实产品管线，并支持 AI 移动体验的快速迭代周期。
By aligning with memory middleware and continuous learning frameworks, it naturally fits into real product pipelines and enables rapid iteration cycles for AI-powered mobile experiences.
通过将长期记忆评估从能力展示转向可测量、可比较、可迭代的评估，MobileMem 直接帮助手机厂商定义端侧智能，并评估个人助手体验。
By shifting long-term memory evaluation from capability demonstration to measurable, comparable, and iterative assessment, MobileMem directly informs how mobile manufacturers define on-device intelligence and evaluate personal assistant experiences.
从核心上看，该基准将用户的长期多模态记忆作为个性化的中心单元。
At its core, the benchmark prioritizes the user's long-term multimodal memory as the central unit of personalization.
该基准已经在单一框架中统一了基础记忆、认知记忆、偏好记忆和视觉推理，同时将评估建立在真实用户数据、一年时间跨度和多样化多模态事件流之上。
The benchmark already unifies foundational memory, cognitive memory, preference memory, and visual reasoning within a single framework, while grounding evaluation in real user data, one-year temporal spans, and diverse multimodal event streams.
MobileMem 已集成到 OPPO 的 AI 助手开发流程中，用于系统评估并增强生产级移动场景中的长期记忆能力。
MobileMem has been integrated into OPPO's AI assistant development pipeline to systematically evaluate and enhance long-term memory capabilities in production-grade mobile scenarios.
详细下载和使用说明维护在项目仓库中。
Detailed download and usage instructions are maintained in the project repository.
本工作由以下机构联合完成。
This work is jointly conducted by the following institutions.
由 OPPO 和 OpenKG 开发。
Developed by OPPO and OpenKG.
