From Facts to Insights: A Persona-Driven Dual Memory Framework and Dataset for Role-Playing Agents

2026-05-25

AI Summary

This work addresses the degradation of role consistency and response genericness in long-term dialogues with role-playing agents, which stems from reliance on role-agnostic memory summarization. To mitigate this, the authors propose DualMem, a novel framework featuring a dual-memory mechanism that decouples memory into two distinct channels: factual knowledge and role-driven insights, with an explicit emphasis on interpreting facts from the character’s perspective. Leveraging a newly curated RoleMemo dataset, they train a 4B-parameter model through supervised fine-tuning and reinforcement learning, further integrating an external memory architecture to better model role consistency. Experimental results demonstrate that DualMem substantially outperforms zero-shot, role-agnostic baselines built upon DeepSeek-V3.2, significantly enhancing role fidelity in extended conversations.

Abstract

While role-playing agents excel in short-term interactions, long-term conversations overwhelm context windows, motivating external memory frameworks. Current systems typically rely on persona-agnostic summarization, which records facts without persona-specific interpretation, yielding generic responses that compromise persona fidelity. To bridge this gap, we introduce RoleMemo, a dataset featuring four reasoning tasks where the factual fragments must be interpreted through the persona to reach the correct answer. Evaluation on RoleMemo exposes critical limitations of persona-agnostic frameworks. We thus propose DualMem, which decouples memory into two streams: factual cognition and persona-conditioned insight. Trained through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), our framework with a 4B-parameter model outperforms zero-shot persona-agnostic frameworks powered by DeepSeek-V3.2 for sustained persona fidelity. Our resources are available at https://github.com/role2026/rolememo.

Problem

• role-playing agents
• persona fidelity
• external memory
• long-term conversation
• persona-agnostic summarization

Innovation

• persona-driven memory
• dual memory framework
• role-playing agents
• persona fidelity
• long-term conversation