--- pretty_name: CUE-Mem Benchmark license: mit tags: - multimodal - conversational-ai - memory - benchmark - personalization - arxiv:2609.32574 - text - image - audio language: - zh - en task_categories: - question-answering - image-text-to-text size_categories: - 1K # 🧠 CUE-Mem ### Benchmarking Long-Term User Memory via Implicit Cues in Multimodal Conversations **多模态对话隐式线索驱动的长期用户记忆评测基准** [![Paper](https://img.shields.io/badge/arXiv-2609.32574-B31B1B)](https://arxiv.org/abs/2609.32574) [![Code](https://img.shields.io/badge/GitHub-CUE--MEM-181717?logo=github)](https://github.com/yulinlp/CUE-MEM) [![Demo](https://img.shields.io/badge/Online_Demo-Explore-2563EB)](https://reichenbach1854-hash.github.io/CUE-Mem/) **Text · Image · Audio | Explicit & Implicit Cues | Long-Term User Memory** [English](#english) · [中文说明](#中文说明) · [Quick Start / 快速开始](#quick-start) · [Citation / 引用](#citation) ![Explicit and implicit user memory across multimodal conversations / 多模态对话中的显式与隐式用户记忆](figure/intro.png) *From visible routines to subtle clues: a cat bowl, a cat tree, and a background meow jointly suggest that the user has a cat.* *从日常活动到隐式线索：猫碗、猫爬架与背景中的猫叫声，共同指向用户养猫这一信息。* --- ## English ### Overview What should an agent remember when a user never says it directly? A recurring object in a photograph or a background sound in a voice message may carry information that matters in a later conversation. **CUE-Mem** evaluates long-term user memory across text, images, and audio, covering both explicit statements and implicit cues. See the [paper](https://arxiv.org/abs/2609.32574) for benchmark design and evaluation protocols. This repository distributes **conversation histories, questions, media, and metadata**. Construction and evaluation code lives on [GitHub](https://github.com/yulinlp/CUE-MEM); the [interactive demo](https://reichenbach1854-hash.github.io/CUE-Mem/) presents selected examples. ### Dataset at a glance Counts below are computed from this release's metadata. A conversation record means one row in `dialogue_turns.jsonl`, not an individual speaker message. Sessions are counted using profile and session ID together. | Component | Count | | :--- | ---: | | User profiles | 20 | | Conversation sessions | 648 | | Conversation records | 5,798 | | Evaluation questions | 2,674 | | Images | 4,876 | | Audio files | 2,228 | | Total media files | 7,104 | Dialogue and question text is primarily Chinese; some descriptive fields are in English. Bilingual documentation does not imply parallel translations of every example.

Question distribution by task, modality, and evidence type / 按任务、模态与证据类型划分的问题分布

*Question distribution in the paper. T / I / A denote text, image, and audio; Exp. / Imp. denote explicit and implicit evidence.* *论文中的问题分布：T / I / A 分别表示文本、图像和音频，Exp. / Imp. 分别表示显式和隐式证据。[View PDF / 查看原始 PDF](figure/dataset_statistics.pdf).* ### Tasks The release's `point` field distinguishes question categories. Image-based questions can supply visual answer options through `option_images`. | Task | Focus | `point` values | Questions | | :--- | :--- | :--- | ---: | | Entity Recall | User-related entities and attributes | `entity_text`, `entity_img` | 1,128 | | Long Pattern | Persistent preferences and habits | `pref_text`, `pref_img` | 363 | | Personalized Recommendation | Apply user memory to recommendations | `rec_text`, `rec_img` | 364 | | Answer Refusal | Insufficient evidence or misleading premises | `adversarial_text` | 819 | [![Examples of the four memory tasks / 四类记忆任务示例](figure/task_example.png)](figure/task_example.pdf) *Four tasks, illustrated with questions, answer choices, and supporting memory clues. Open the PDF to inspect the details.* *四类任务的具体示例，包括问题、候选答案与记忆线索。[View PDF / 查看原始 PDF](figure/task_example.pdf).* **Label detail:** `qa_type` is heterogeneous: some values are `explicit` / `implicit`, others are entity categories (`Relationship`, `Pets`, `Items`), and a few are empty. Do not use it as a universal binary evidence label. Preserve optional fields such as `entity_explicitness` and follow the evaluation code's interpretation. ### Construction Pipeline / 数据构建流程 ![Benchmark construction: script synthesis, multimodal dialogue synthesis, and data review / 基准构建：脚本合成、多模态对话合成与数据审核](figure/benchmark.png) *Construction pipeline: profile and event planning → multimodal dialogue synthesis → automatic checks and human review.* *构建流程：用户档案与事件规划 → 多模态对话合成 → 自动检查与人工审核。* ### Files and annotations Files are published directly at the repository root, without an extra `CUE-Mem-Benchmark/` wrapper or archive extraction step. See the shared directory tree below. Each `history_with_qa_p*.json` contains `character_profile`, `multi_session_dialogues`, and `human-annotated QAs`. The flattened JSONL files are convenient entry points for analysis. | Metadata file | Contents | Useful fields | | :--- | :--- | :--- | | `dialogue_turns.jsonl` | One conversation record per line | `p_id`, `session_id`, `round`, `date`, `task_id`, `user`, `assistant`, `input_image`, `input_voice_message` | | `questions.jsonl` | One question per line | `qa_id`, `p_id`, `question`, `answer`, `point`, `qa_type`, `clue`, `option_images`, `source_file` | | `media_manifest.jsonl` | Media inventory | `path`, `media_type`, `format`, `size_bytes` | | `checksums.sha256` | SHA-256 checksums for `data/` files | Hash and relative file path | Fields vary by modality and task. Audio-only turns may have no `user` text field. Captions are available in `image_caption`, `voice_caption`, and `user_voice_message_caption` where present. Media paths, `source_file`, and manifest `path` use `/` separators and are relative to the dataset root. However, `clue` can contain evidence identifiers such as `D09:02` or `D09-001.wav`, which are not necessarily repository paths. Include `p_id` when joining session or turn references across profiles. ### Evaluation notes Keep reference answers, rationales, and evidence annotations outside the evaluated system's memory unless the protocol explicitly provides them, such as an oracle-evidence baseline. State whether the system receives raw media, captions, or both. Preserve profile/session boundaries and conversation order; report the dataset revision, question selection, model, and memory configuration. Generated scenarios and profile media should not be interpreted as verified observations about real people. Consult the [paper](https://arxiv.org/abs/2609.32574) for construction details and limitations, and the [GitHub README](https://github.com/yulinlp/CUE-MEM#readme) for environment setup, dataset preparation, and RQ1–RQ3 evaluation instructions. --- ## 中文说明 ### 基准简介 用户没有直接说出口的信息，智能体能否在之后的对话中记住并使用？照片中反复出现的物品、语音里的环境声音，都可能是理解用户的重要线索。 **CUE-Mem** 面向文本、图像和音频交织的长期对话，评测记忆系统对显式信息与隐式线索的保留、检索和利用能力。基准设计与评测协议见[论文](https://arxiv.org/abs/2609.32574)。 本仓库提供**对话历史、问答、媒体文件及元数据**；构建与评测代码位于 [GitHub](https://github.com/yulinlp/CUE-MEM)，可通过[在线 Demo](https://reichenbach1854-hash.github.io/CUE-Mem/)浏览精选示例。 ### 数据规模与任务 当前发布数据包含 **20 个用户档案、648 个会话、5,798 条对话记录、2,674 道问题**，以及 **4,876 张图像、2,228 个音频文件**。会话按用户和会话 ID 联合计数；对话记录数对应 `dialogue_turns.jsonl` 的行数，并非将双方消息分别计数。 | 任务 | 评测目标 | `point` 字段 | 题数 | | :--- | :--- | :--- | ---: | | 实体回忆（Entity Recall） | 记住与用户相关的实体及属性 | `entity_text`、`entity_img` | 1,128 | | 长期模式（Long Pattern） | 从多次交互中归纳偏好和习惯 | `pref_text`、`pref_img` | 363 | | 个性化推荐（Personalized Recommendation） | 将用户记忆用于新的推荐问题 | `rec_text`、`rec_img` | 364 | | 拒答（Answer Refusal） | 识别证据不足或问题前提不可靠的情况 | `adversarial_text` | 819 | 对话和问题主要为中文，部分描述字段为英文；双语说明并不代表全部样本都有平行翻译。图像类问题可通过 `option_images` 提供图片选项。 ### 文件与字段 仓库根目录直接包含 `README.md`、`LICENSE`、`data/` 和 `metadata/`，没有额外套一层目录，下载后无需拼接或解压分卷。完整目录树见下方。 `data/dialog/base/` 中每个用户对应一个 JSON，包含 `character_profile`（用户档案）、`multi_session_dialogues`（多会话对话）和 `human-annotated QAs`（问答）。`data/event/` 保存对话图像与音频，`data/qa/` 保存问题图片，`data/profile/generated_portraits/` 保存生成的用户／实体参考图像。 | 元数据文件 | 内容与读取要点 | | :--- | :--- | | `dialogue_turns.jsonl` | 每行一条对话记录；用 `p_id`、`session_id`、`round` 定位，媒体见 `input_image`、`input_voice_message` | | `questions.jsonl` | 每行一道问题；包含问题、答案、任务类型、线索及部分题目的图片选项 | | `media_manifest.jsonl` | 媒体清单：`path`、`media_type`、`format`、`size_bytes` | | `checksums.sha256` | `data/` 文件的 SHA-256 校验值，用于检查下载完整性 | 不同模态和题型的字段不完全一致。音频轮次可能没有 `user` 文本字段；描述可见 `image_caption`、`voice_caption`、`user_voice_message_caption` 等可选字段。 读取时请注意： - 媒体路径、`source_file` 和清单中的 `path` 相对于数据集根目录；`clue` 中的 `D09:02`、`D09-001.wav` 等可能是线索 ID，不能一律当作磁盘路径。跨用户关联时应同时使用 `p_id`。 - `qa_type` 并非统一的显式／隐式二分类：部分为 `explicit` 或 `implicit`，实体题也使用 `Relationship`、`Pets`、`Items`，另有少量空值。请结合 `entity_explicitness` 等字段及评测代码解释标签。 ### 评测建议 请保持用户、会话边界和原始对话顺序，并记录数据版本、抽样方式、模型及记忆系统配置。除协议明确允许的情况（如 oracle-evidence 基线）外，不应将标准答案、解释或标注线索提前写入被评测系统的记忆。使用原始媒体、文本描述或二者组合时，也应在结果中说明。 生成场景及用户媒体不应被视为真实人物的已核实经历。构建方法和局限性见[论文](https://arxiv.org/abs/2609.32574)；环境配置、数据准备及 RQ1–RQ3 实验步骤见 [GitHub README](https://github.com/yulinlp/CUE-MEM#readme)。 --- ## Repository Layout / 仓库结构 ```text . ├── README.md ├── LICENSE ├── data/ │ ├── dialog/base/ # Profile histories + QA / 用户历史与问答 │ │ ├── history_with_qa_p0.json │ │ └── ... │ ├── event/images/ # Event images / 对话事件图像 │ ├── event/voice_mixed/ # Event audio / 对话音频 │ ├── qa/pref_images/ # Preference / 偏好题图片 │ ├── qa/rec_images/ # Recommendation / 推荐题图片 │ ├── qa/entity_images/ # Entity / 实体题图片 │ └── profile/generated_portraits/ # Profile/entity references / 档案与实体参考图 └── metadata/ ├── dialogue_turns.jsonl ├── questions.jsonl ├── media_manifest.jsonl └── checksums.sha256 ``` ## Quick Start / 快速开始 ### 1. Download / 下载 Download the full release with the Hugging Face CLI. / 使用 Hugging Face CLI 下载完整数据： ```bash python -m pip install -U huggingface_hub hf download Kkryptonite/CUE-Mem --repo-type dataset --local-dir benchmark-data ``` For a lightweight first look, download only metadata and profile JSON files. Download media before multimodal evaluation. / 若只想先查看结构，可仅下载元数据与用户 JSON；多模态评测前需补齐媒体： ```python from huggingface_hub import snapshot_download snapshot_download( repo_id="Kkryptonite/CUE-Mem", repo_type="dataset", local_dir="benchmark-data", allow_patterns=["README.md", "LICENSE", "metadata/*", "data/dialog/base/*.json"], ) ``` For reproducibility, set `revision` to a commit hash. See the [download guide](https://huggingface.co/docs/huggingface_hub/guides/download). / 固定版本时可设置 `revision` 为 commit hash，详见[下载说明](https://huggingface.co/docs/huggingface_hub/guides/download)。 ### 2. Read records / 读取数据 ```python import json from pathlib import Path root = Path("benchmark-data") def read_jsonl(relative_path): with (root / relative_path).open(encoding="utf-8") as handle: for line in handle: if line.strip(): yield json.loads(line) question = next(read_jsonl("metadata/questions.jsonl")) print(question["qa_id"], question["question"]) for label, relative_path in question.get("option_images", {}).items(): print(label, root / relative_path) turn = next(read_jsonl("metadata/dialogue_turns.jsonl")) print(turn["p_id"], turn["session_id"], turn["round"]) for field in ("input_image", "input_voice_message"): for relative_path in turn.get(field, []): print(field, root / relative_path) ``` ### 3. Verify files / 校验完整性 Run after downloading the full `data/` directory. / 完整下载 `data/` 后执行： ```bash cd benchmark-data # Linux sha256sum -c metadata/checksums.sha256 # macOS (alternative / 替代命令) shasum -a 256 -c metadata/checksums.sha256 ``` --- ## License / 许可证 This repository includes an [MIT License](LICENSE). Refer to the file for the full terms. / 本仓库附有 [MIT 许可证](LICENSE)，完整条款以该文件为准。 ## Citation If you use CUE-Mem in your research, please cite the [paper](https://arxiv.org/abs/2609.32574). / 如在研究中使用 CUE-Mem，请引用以下论文： ```bibtex @misc{hu2026cuemem, title={CUE-Mem: Benchmarking Long-Term User Memory via Implicit Cues in Multimodal Conversations}, author={Yulin Hu and Yanyan Zhao and Zimo Long and Xing Fu and Mengtong Ji and Weixiang Zhao and Yutai Hou and Qianchao Wang and Dandan Tu}, year={2026}, eprint={2609.32574}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2609.32574} } ``` Code and evaluation questions: [GitHub Issues](https://github.com/yulinlp/CUE-MEM/issues). Dataset questions: [Hugging Face Discussions](https://huggingface.co/datasets/Kkryptonite/CUE-Mem/discussions). / 代码与评测问题请提交至 [GitHub Issues](https://github.com/yulinlp/CUE-MEM/issues)，数据集问题可在 [Hugging Face Discussions](https://huggingface.co/datasets/Kkryptonite/CUE-Mem/discussions) 反馈。