Narrative World Model outperforms Graphiti/Zep on narratology QA

The Narrative World Model, authored by Mohammad Saifullah and five coauthors and submitted to arXiv on 6 Jul 2026, introduces a writer-memory system for long-form fiction that pairs a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. The paper reports that this system substantially and significantly outperforms the strong temporal-knowledge-graph agent-memory baseline Graphiti/Zep (Rasmussen et al., 2025) on multi-hop narratological question answering across two corpora.

What is the Narrative World Model?

The Narrative World Model, or NWM, is a writer-memory architecture that represents story state as a "narratology-grounded typed temporal-state graph" and couples that representation with query-conditioned hybrid retrieval. The graph is typed and temporal to encode narratological structure such as who knows a secret and when, event ordering relative to narration, payoffs of setups, and relationship shifts; retrieval is conditioned on the query to surface chapter-safe evidence for the reader.

The paper frames the problem as memory for long-form fiction where writers need answers to multi-hop narratological questions, and argues that general-purpose retrieval and agent-memory systems fail because they lack narratological structure. The authors provide a reproducible public corpus and a validated multi-hop benchmark to evaluate memory rather than downstream answerers.

How was NWM evaluated and how did it perform?

The authors read every system through a single held-constant Opus 4.8 reader over only that system's chapter-safe evidence, and evaluated on both a reproducible public corpus and a validated multi-hop benchmark; under that setup, NWM substantially and significantly outperformed Graphiti/Zep, and far exceeded GraphRAG and flat retrieval. The evaluation isolates memory by holding the reader constant and restricting evidence to chapter-safe material.

The paper emphasizes that the advantage is representational rather than an extraction artifact: rebuilding the baseline with NWM's extractor did not remove NWM's lead, and the gains trace to the narratology-grounded graph structure and query-conditioned retrieval rather than graph size or extractor quality. The manuscript runs 23 pages with 4 figures and comprises a 9-page main text plus appendix, and cites Graphiti/Zep as the strongest existing temporal-knowledge-graph agent-memory framework (Rasmussen et al., 2025).

Why it matters

NWM tackles a practical gap for long-form fiction: writers and story systems need memory that answers multi-hop questions about evolving story state, not only entity facts. By encoding narratological relations explicitly in a typed temporal-state graph and by conditioning retrieval on the question, NWM changes what evidence is surfaced. If the representational claims hold across genres, memory systems built around narratology could reduce wrong or missing evidence in story-aware applications and QA.

The paper's controlled evaluation design — a single Opus 4.8 reader and chapter-safe evidence — strengthens the case that representation, not auxiliary engineering, drives the improvement.

What to watch

Look for release of the code, extractor, and dataset links listed under the paper's Code, Data and Media section; those materials are already surfaced on the arXiv entry as toggles. Also watch for replication of NWM beyond the paper's public corpus and validated benchmark, and for how other agent-memory frameworks adapt narratology-grounded structures.

References and specifics: the paper is arXiv:2607.05577, submitted on 6 Jul 2026, and compares NWM to Graphiti/Zep (Rasmussen et al., 2025), GraphRAG, and flat retrieval. The authors describe their core structure as a "narratology-grounded typed temporal-state graph" paired with query-conditioned hybrid retrieval.