Memory Papers
Memory Architecture

ENPMR-Bench: Benchmarking Proactive Memory Retrieval for Emotional Support Agents

AuthorsXing Fu, Yulin Hu, Mengtong Ji et al.

arXiv 20262026
arXivPDF

TL;DR

ENPMR-Bench uses a Maslow-grounded need–memory mapping to show that Recall@10 on emotional memories tops out at 46.42% with empathy capped far below golden-memory responses.

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Emotional agents mis-retrieve memories with Recall@10 stuck at 46.42% and Top-1 below 10%

ENPMR-Bench shows that even the best embedding model reaches only 46.42% Recall@10 and under 10% Top-1 on emotional memories.

In emotional support dialogue, this means ENPMR-Bench exposes agents that miss users’ latent needs, degrading empathy scores compared to golden-memory conditions.

HOW IT WORKS

ENPMR-Bench — Emotional Need-aware Proactive Memory Retrieval

ENPMR-Bench combines Structured Memory Retrieval Guidelines, User Profile Generation, Emotional Support Dialogue Generation, and the Memory Retrieval Task to test need-aware memory selection.

You can think of ENPMR-Bench like a psychological card catalog: Maslow-based need labels route queries to specific memory card types instead of nearest-neighbor text matches.

This design lets ENPMR-Bench probe whether agents can infer latent needs and pick supportive memories, something a plain context window or vanilla RAG cannot capture.

DIAGRAM

Need-aware proactive memory retrieval flow in ENPMR-Bench

This diagram shows how ENPMR-Bench routes each user turn through need inference, memory selection, and response generation in the task formulation at Equation (1).

DIAGRAM

ENPMR-Bench data construction and evaluation pipeline

This diagram shows how ENPMR-Bench builds user profiles, generates memory-augmented dialogues, and then evaluates retrieval and response generation.

PROCESS

How ENPMR-Bench Handles an Emotional Support Dialogue Session

  1. 01

    User Profile Generation

    ENPMR-Bench runs User Profile Generation to structure personas from PersonaHub into demographics plus themed memories like relationships, powers, goals, highlights, and preferences.

  2. 02

    Generate Sessions From Profile

    ENPMR-Bench applies Generate Sessions From Profile to turn each memory entry into temporally ordered user–AI snippets forming long-term dialogue histories.

  3. 03

    Emotional Support Dialogue Generation

    ENPMR-Bench uses Emotional Support Dialogue Generation to sample target memories by need type and create short-turn support dialogues grounded only in those memories.

  4. 04

    Memory Retrieval Task

    ENPMR-Bench finally runs the Memory Retrieval Task and Response Generation Task, letting models select memories and produce empathetic replies evaluated by Recall, nDCG, and empathy.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Comprehensive ENPMR evaluation protocol

    ENPMR-Bench defines a comprehensive evaluation protocol for Emotional Need-aware Proactive Memory Retrieval using Structured Memory Retrieval Guidelines grounded in Maslow’s hierarchy and Equation (1).

  • 02

    Systematic comparison of memory systems

    ENPMR-Bench systematically compares embedding-based retrieval, Agentic Memory systems like Mem0, Memu, MemOS, and LLMs such as GPT-4o and DeepSeek-V3 on identical emotional support dialogues.

  • 03

    Identified challenges for emotional support agents

    ENPMR-Bench identifies key challenges, showing Qwen3-Embedding-8B reaches only 46.42% Recall@10 and DeepSeek-V3 empathy rises from 4.38 to 4.91 with golden memories.

RESULTS

By the Numbers

Recall@10

46.42%

+5.55 over gte-Qwen2-7B-instruct

Recall@1

9.46%

vs gte-Qwen2-7B-instruct at 6.30%

Empathy score w/ retrieved

4.38

DeepSeek-V3 base vs 4.91 with golden memories

Memory Type Match Rate

0.59

DeepSeek-V3 and GPT-4o w/CoT vs lower base TMR values

ENPMR-Bench evaluates emotional need-aware retrieval and support on 1,872 Chinese emotional support dialogues with 11,846 memories. The 46.42% Recall@10 and 4.91 empathy upper bound show ENPMR-Bench exposes retrieval as the main bottleneck for emotional intelligence.

BENCHMARK

By the Numbers

ENPMR-Bench evaluates emotional need-aware retrieval and support on 1,872 Chinese emotional support dialogues with 11,846 memories. The 46.42% Recall@10 and 4.91 empathy upper bound show ENPMR-Bench exposes retrieval as the main bottleneck for emotional intelligence.

BENCHMARK

Retrieval performance comparison of different embedding models on ENPMR-Bench

Recall@10 on ENPMR-Bench memory retrieval task.

KEY INSIGHT

The Counterintuitive Finding

ENPMR-Bench shows that even with topic-filtered memories, LLMs achieve a maximum Memory Type Match Rate of only 0.53 and empathy below golden-memory runs.

This is surprising because many assume strong LLMs can implicitly infer needs, yet ENPMR-Bench reveals systematic misalignment between user needs and retrieved memory types.

WHY IT MATTERS

What this unlocks for the field

ENPMR-Bench unlocks a way to stress-test emotional need inference and proactive memory retrieval rather than just factual recall in long-term agents.

With ENPMR-Bench, builders can design and benchmark memory systems that explicitly align Maslow-style needs with memory types, enabling more trustworthy emotional support agents.

~12 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

Paper: ENPMR-Bench: Benchmarking Proactive Memory Retrieval for Emotional Support Agents

Answers use this explainer on Memory Papers.

Checking…