Narrative World Model: A Memory System for AI Writers Based on Narratology

Creating long works of fiction requires AI systems to have a memory capable of answering complex, multi-step questions about the state of the plot: who knows the secret and when they learned it, whether an event preceded the narrative that revealed it, whether a plot device worked, and how the characters’ relationships have changed.

Existing agent memory systems and general-purpose retrieval solutions store entities and facts, but do not account for the narratological structure upon which such questions are based. As a result, they return irrelevant evidence or fail to find it at all.

The Narrative World Model Approach

The proposed NWM system combines two components:

• A typified temporal state graph based on the principles of narratology;
• Hybrid retrieval, determined by the type of query.

This structure allows the system to track not just facts, but their role in the narrative—temporal connections, the revelation of secrets, the development of relationships, and dramatic arcs.

Evaluation Methodology

To measure memory quality specifically—rather than the model’s ability to answer questions—the authors ran all compared systems through a single reader—Opus 4.8—with access only to evidence relevant to a specific chapter. Testing was conducted on a reproducible public corpus and a validated benchmark for multi-step questions.

The baseline comparison was Graphiti/Zep—the strongest of the existing agent memory frameworks based on temporal knowledge graphs—served as the baseline comparison. Comparisons were also made with GraphRAG and flat retrieval.

Results

NWM significantly and statistically outperforms Graphiti/Zep on multi-step narratological QA on both corpora, and also noticeably outperforms GraphRAG and flat retrieval.

Key conclusion: the advantage is representational in nature, rather than extraction-based. This advantage persists even when the baseline system is re-engineered using NWM’s own extractor and can be traced specifically to the narratological structure and query-driven retrieval, rather than to the graph size or the quality of fact extraction.

Source: cs.AI updates on arXiv.org