Search papers, labs, and topics across Lattice.
This paper introduces REVERIEMEM, a novel three-layer memory architecture designed to enhance book-based role-playing agents by addressing two critical failures: Factual Overreach and Stylistic Monotony. By implementing episodic, semantic, and personality layers, the system allows characters to maintain a consistent perspective while generating contextually appropriate responses. Evaluation against the KBF-QA benchmark reveals a 34.6 percentage point improvement in Knowledge Boundary Fidelity, and a win rate of approximately 79% on narrative generation tasks, indicating significant advancements in character-grounded storytelling.
REVERIEMEM boosts character fidelity in role-playing agents, achieving a staggering 34.6% improvement in knowledge accuracy while preserving unique character voices.
Recent LLM role-playing systems build character agents from novels by extracting characters, scenes, and relations. Yet long-narrative role-playing suffers from two failures: Factual Overreach, where shared retrieval or parametric memory lets a character use facts outside its perspective, and Stylistic Monotony, where profile descriptions flatten a character into a fixed voice. To address these failures, we propose REVERIEMEM, a three-layer memory architecture for book-based character agents. The episodic layer stores first-person scene memories; the semantic layer stores visibility-tagged facts; and the personality layer stores situation-dependent speech and behaviour patterns. For evaluation, we construct KBF-QA, a 4,386-question benchmark over eight novels for testing knowledge boundaries. REVERIEMEM improves Knowledge Boundary Fidelity by 34.6 percentage points over the strongest prior method. On BOOKWORLD's five-dimension pairwise narrative protocol, REVERIEMEM achieves a ~ 79% win rate, suggesting that perspective-bounded memory improves both boundary fidelity and character-grounded narrative generation.
DAMO