Staying In Character: Perspective-Bounded Memory For Book-Based Role-Playing Agents

Researchers introduce REVERIEMEM, a three-layer memory architecture that enables large language model-based character agents to maintain perspective-bounded knowledge and distinct personalities when roleplaying in book-based narratives. The system addresses key limitations in current LLM roleplay systems by preventing characters from accessing facts outside their perspective and eliminating flattened, monotonous characterization.

This research addresses a fundamental challenge in using large language models for narrative generation: maintaining consistent character knowledge boundaries and voice diversity. Current LLM roleplay systems suffer from two critical failures—characters inappropriately access information they shouldn't know, and profile descriptions reduce complex personalities to static patterns. REVERIEMEM's three-layer architecture elegantly solves these problems by separating episodic memories (first-person scene experiences), semantic knowledge (visibility-tagged facts), and behavioral patterns (situation-dependent speech and actions).

The work builds on growing interest in applying LLMs to complex narrative tasks, where character consistency and knowledge constraints become increasingly important. As language models improve, researchers recognize that raw capability alone produces unconvincing roleplay—characters need genuine boundaries and nuanced personality expression to feel authentic. This research represents a meaningful step toward more sophisticated narrative AI.

The practical implications extend beyond entertainment applications. Publishing, game development, and interactive storytelling platforms all depend on believable character agents that respect narrative constraints. The KBF-QA benchmark with 4,386 questions across eight novels provides a rigorous evaluation framework, while the 34.6 percentage point improvement in knowledge boundary fidelity demonstrates substantial technical progress. The ~79% win rate on narrative quality tests suggests the approach succeeds at both technical correctness and user-facing experience quality.

Future development should focus on scaling REVERIEMEM across diverse literary genres and examining how the architecture handles dynamic information revelation throughout longer narratives. Researchers should also explore whether these memory structures could apply to other constrained roleplay scenarios beyond literature.

• →REVERIEMEM introduces a three-layer memory system that prevents LLM characters from accessing information outside their narrative perspective.
• →The system achieves 34.6 percentage point improvements in knowledge boundary fidelity compared to prior methods.
• →A new 4,386-question benchmark (KBF-QA) provides rigorous evaluation of character knowledge constraints across eight novels.
• →The architecture separately manages episodic memories, visibility-tagged facts, and situation-dependent behavioral patterns for authentic characterization.
• →~79% win rate on narrative quality tests indicates the approach improves both technical accuracy and user experience.