level: research
long-form fiction writers need memory that can answer multi-hop questions about changing story state. these questions include who knows a secret and when they learned it, whether an event happened before the narration that revealed it, whether a setup paid off, and how a relationship shifted. general-purpose retrieval and agent-memory systems represent entities and facts but not the narratological structure these questions depend on. as a result, they often surface wrong evidence or none at all.
the narrative world model (nwm) is a writer-memory system that combines a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. it is designed to capture the structural elements of stories, such as plot, character knowledge, and temporal order. the system was tested on a reproducible public corpus and a validated multi-hop benchmark. to measure memory rather than the answering ability, all systems were read through a single held-constant opus 4.8 reader using only chapter-safe evidence.
the nwm was compared against the strongest existing temporal-knowledge-graph agent-memory framework. the approach focuses on providing writers with accurate and relevant story state information, which is critical for maintaining consistency in long narratives. by grounding memory in narratology, the system aims to support complex storytelling tasks that require tracking evolving relationships and events over time.
why it matters: better story memory can improve ai-generated fiction and other long-form text by keeping track of complex plot and character details.