Narrative World Model (NWM) beats agent memory for long-form fiction QA

Long-form fiction writers and AI assistants struggle to answer complex story-state questions: Did a character know a secret before it was revealed? Was a setup paid off? General-purpose retrieval and agent-memory systems treat facts and entities but ignore the narratological structure that underpins these multi-hop queries. A new preprint from Mohammad Saifullah and five co-authors introduces the Narrative World Model (NWM), a dedicated writer-memory system that explicitly models temporal and causal relationships in fiction.

NWM pairs a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. To isolate memory performance, the team evaluated each system using a single, fixed reader model (Opus 4.8) over identical chapter-safe evidence. On a reproducible public corpus and a custom multi-hop benchmark, NWM substantially and significantly outperformed Graphiti/Zep (Rasmussen et al., 2025), the strongest existing temporal-knowledge-graph agent-memory framework, as well as GraphRAG and flat retrieval. The advantage held even when the baseline was rebuilt with NWM's own extractor, confirming the core benefit comes from its narratology-grounded representation and query-conditioned retrieval, not graph size or extraction quality.