ENPMR-Bench: Benchmarking Proactive Memory Retrieval for Emotional Support Agents
Introduces ENPMR-Bench, a benchmark for proactive memory retrieval based on Maslow’s hierarchy, revealing significant deficiencies with current models, and outlining future directions.
Key Findings
Methodology
This work proposes a structured framework grounded in Maslow’s hierarchy to evaluate proactive memory retrieval for emotional needs. It constructs over 1800 multi-modal dialogues, mapping unmet needs to specific memory types. Embedding models (e.g., Qwen3-Embedding-8B) and large language models (GPT-4o, DeepSeek-v3) are assessed using metrics like Recall@10 and nDCG. Chain-of-Thought (CoT) prompting is employed to improve user need inference. The evaluation systematically compares models’ ability to identify, retrieve, and utilize emotionally salient memories in long-term emotional support scenarios, highlighting deficiencies in current approaches.
Key Results
- The best models achieve only 46.41% Recall@10, with Top-1 accuracy below 10%, indicating semantic similarity alone is insufficient for emotional memory retrieval.
- Even under Golden memory conditions, empathy scores only reach 4.91, showing a significant gap between current retrieval quality and ideal performance.
- Chain-of-Thought prompting offers moderate improvements, but a persistent performance gap remains, revealing systematic biases in memory selection and emotional understanding.
Significance
This research exposes critical limitations in current affective dialogue systems’ proactive memory retrieval capabilities. It emphasizes that accurate identification and matching of users’ latent emotional needs are essential for genuine empathy. The benchmark provides a systematic evaluation platform, guiding future innovations to develop more emotionally intelligent agents, with broad implications for mental health, companionship robots, and personalized support systems.
Technical Contribution
The paper introduces a novel structured mapping based on Maslow’s hierarchy, integrating multi-modal data and multi-metric evaluation. It combines demand inference with memory retrieval, employing chain-of-thought reasoning to enhance accuracy. The creation of a large, annotated dialogue dataset fills a significant gap in systematic evaluation of proactive emotional memory retrieval, advancing the state-of-the-art in affective AI.
Novelty
This is the first work to systematically incorporate Maslow’s hierarchy into a structured memory retrieval framework for emotional support. It combines multi-modal data, multi-metric evaluation, and demand-aware retrieval, setting a new standard beyond static, single-dimensional benchmarks. The approach bridges psychological theory and AI engineering, offering a comprehensive assessment of emotional intelligence in dialogue agents.
Limitations
- Models struggle with complex, multi-demand scenarios, often over-relying on surface cues, which limits their nuanced understanding of emotional states.
- The benchmark relies on static dialogue data, lacking real long-term interaction validation, which restricts practical deployment.
- Memory bias towards certain types (e.g., highlight memories) reduces emotional resonance, indicating need for more balanced retrieval strategies.
Future Work
Future research will focus on integrating multi-modal signals (visual, auditory) for richer emotional understanding, developing dynamic memory management for long-term interactions, and expanding datasets to include real-world, long-term user interactions. These efforts aim to create more adaptive, empathetic AI systems capable of genuine emotional support.
AI Executive Summary
In recent years, affective AI has gained prominence, especially in applications like mental health support and companionship. Despite advances, current systems often rely on static knowledge retrieval, which limits their ability to understand and respond to users’ latent emotional needs. This gap hampers genuine empathy and long-term engagement. To address this, we introduce ENPMR-Bench, a comprehensive benchmark grounded in Maslow’s hierarchy, designed to evaluate proactive memory retrieval in emotional support scenarios.
The benchmark constructs over 1800 multi-modal dialogues, systematically mapping unmet needs—such as physiological, belonging, esteem, and self-actualization—to specific memory types. Using models like Qwen3-Embedding-8B and GPT-4o, we assess their ability to identify, retrieve, and utilize emotionally salient memories. Metrics like Recall@10 and nDCG reveal that even the best models only reach 46.41% in retrieval recall, with Top-1 accuracy below 10%. Chain-of-Thought prompting improves need inference but does not close the gap, indicating persistent biases in memory selection.
These findings highlight that current models lack the nuanced understanding required for emotionally appropriate memory retrieval, which is vital for genuine empathy. The study underscores the importance of demand-aware retrieval strategies and multi-modal integration. Its implications extend to designing more emotionally intelligent agents capable of long-term, personalized support, with applications spanning mental health, social robotics, and beyond.
While promising, the work faces limitations such as reliance on static datasets and biases in memory preference. Future directions include incorporating visual and auditory cues, developing adaptive long-term memory management, and validating in real-world long-term interactions. ENPMR-Bench offers a vital tool for advancing affective AI, guiding the development of systems that truly understand and support human emotional needs.
Deep Analysis
Background
Affective dialogue systems已成为人机交互的重要研究方向。早期如ELIZA、ChatterBot主要关注任务导向的问答，随后发展出情感识别与表达（如DialoGPT、Meena），强调理解和表达情感。深度学习推动记忆增强模型（如ReMemBERT、Longformer）应用于长时上下文理解，但多集中于事实知识检索，缺乏对用户潜在情感需求的主动识别。心理学研究表明，情感支持的有效性依赖于对用户深层需求的理解和匹配，尤其在多轮长时交互中，记忆的合理利用至关重要。现有评估多偏静态、单维，难以全面衡量模型在需求识别和记忆匹配中的表现，亟需系统化基准。
Core Problem
当前情感支持系统在主动识别用户潜在情感需求方面存在明显不足，主要原因是缺乏结构化的需求映射和多模态支持。模型多依赖语义相似性，难以捕捉深层次情感状态，导致记忆匹配偏差，影响共情效果。此外，缺乏真实长时交互验证，限制了实际应用。需要提出更有效的需求导向检索策略，以实现更自然、更贴心的情感交互。
Innovation
本研究创新点包括：1）引入基于马斯洛需求层次的结构化记忆映射，为记忆检索提供理论指导；2）结合多模态信息（文本、视觉）提升需求识别能力；3）设计多指标评估体系（如Recall、nDCG、情感得分），系统衡量模型在需求匹配中的表现；4）采用链式推理（CoT）增强模型对潜在需求的理解。通过这些创新，突破了以往静态、单维度的评估范式，推动主动记忆检索在情感支持中的应用。
Methodology
- �� 构建多模态用户画像，涵盖基本信息和多类别记忆条目。
- �� 设计基于马斯洛需求的四个核心需求维度（生理、归属、尊重、自我实现），定义不同需求对应的记忆类型。
- �� 采集1800+多模态对话，结合专家标注，建立多层次情感支持场景。
- �� 利用多模态特征（文本、视觉）训练embedding模型（如Qwen3-Embedding-8B），评估其在Recall和nDCG指标上的表现。
- �� 采用大模型（GPT-4o、DeepSeek-v3）进行需求识别和记忆匹配，结合链式推理（CoT）提升识别准确率。
- �� 设计多任务评估体系，测量记忆检索、情感共鸣和响应生成的效果，分析模型偏差和不足。
Experiments
采用专家设计的1800+对话样本，涵盖不同需求层次和场景。模型包括多种embedding（Qwen3-Embedding-8B、doubao-embedding）和大模型（GPT-4o、DeepSeek-v3、Qwen-Max）。评估指标包括Recall@10、Top-1准确率、nDCG，以及情感得分。通过对比不同模型和策略（如链式推理、Golden记忆），分析其在需求识别、记忆匹配和响应质量上的表现。还进行ablation研究，验证结构化需求映射和多模态信息的贡献。
Results
模型在Recall@10最高仅达46.41%，Top-1准确率低于10%，显示语义相似性不足以捕获情感相关记忆。链式推理带来一定提升，但仍远低于Golden记忆条件，情感得分在4.91左右，说明记忆匹配仍是瓶颈。不同模型偏好不同记忆类型，偏向突出highlight记忆，导致情感共鸣不足。这些数据表明，当前模型在理解复杂情感需求和匹配合适记忆方面存在系统性偏差，亟需改进。
Applications
该研究为情感支持机器人、心理咨询助手等提供评估工具，有助于提升系统的个性化和共情能力。未来可结合多模态信息，优化需求识别和记忆匹配策略，推动智能陪伴和心理健康服务的广泛应用。
Limitations & Outlook
模型在多需求同时存在时表现不佳，需求理解仍有限。评估主要基于静态对话，缺乏真实长时交互验证。记忆匹配偏差导致情感共鸣不足，未来需结合多模态和情感分析技术，提升系统鲁棒性和实际应用能力。
Plain Language Accessible to non-experts
想象你在一家厨房做饭，每次准备食材都要根据菜谱和你的口味选择不同的调料。系统就像这个厨房的厨师，它需要记住你喜欢吃辣还是不辣，喜欢甜还是咸。过去的经验（记忆）帮助厨师做出更符合你口味的菜肴，但如果厨师只记得一些表面信息，比如你喜欢的菜名，而忘了你更喜欢辣味，那做出来的菜就可能不合心意。这个研究就像在教厨师如何更聪明地记住你的深层需求，主动找出最适合你的调料，从而做出让你感动的菜肴。它强调理解你的真正需求，而不是仅仅依赖表面信息。未来的系统会变得更像一个懂你心思的厨师，能提前准备你喜欢的菜肴，让你每次用餐都觉得温暖贴心。
ELI14 Explained like you're 14
想象你和朋友聊天，有时候你会偷偷告诉他一些心里的事，比如觉得孤单或者烦恼。朋友如果能记住这些细节，下次你遇到类似问题时，他就能主动帮你出主意或者陪你玩。这就像是一个超级聪明的朋友，能记住你所有的小秘密和喜欢的事情，然后在你需要时帮你解决问题。这个研究的目标就是让人工智能变得像那个懂你心思的朋友一样，不仅能记住你说过的话，还能理解你的情感需求，主动提供帮助。它通过分析你以前说过的事情，判断你现在可能的心情，然后找出最合适的“记忆”来回应你。虽然现在还做得不够好，但未来它会变得更懂你，让你每次和它聊天都像和老朋友一样温暖、贴心。
Abstract
Memory-augmented language agents are increasingly deployed in affective applications such as emotional support, where understanding and responding to users' latent emotional needs is critical. However, existing research often treats memory as a tool for factual retrieval, overlooking its role in shaping users' emotional experiences. In this work, we introduce ENPMR-Bench, a benchmark for evaluating Emotional Need-aware Proactive Memory Retrieval (ENPMR), a core capability that enables agents to infer users' latent emotional needs and proactively retrieve appropriate memories to support empathetic interaction. Grounded in Maslow's hierarchy of needs, ENPMR-Bench includes over 1,800 memory-augmented dialogues and defines structured mappings between emotional needs and supportive memory types. Experimental results demonstrate that current retrieval paradigms, including both embedding-based and LLM-driven approaches, exhibit substantial deficiencies, with empathy scores significantly lagging behind golden memory conditions. While chain-of-thought prompting improves the alignment between inferred emotional needs and retrieved memories to some extent, a notable performance gap remains. Together, these findings reveal critical limitations in current agents and outline directions for advancing personalized emotional support through need-sensitive memory retrieval.