🤖 AI

Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction

The paper introduces the Narrative World Model (NWM), a writer-memory system that leverages a narratology-grounded temporal-state graph and query-conditioned hybrid retrieval to significantly outperform existing agent-memory frameworks in answering complex, multi-hop questions about evolving story states in long-form fiction.

Original authors: Mohammad Saifullah, Thomas Kornmaier, Taaha Kazi, Vasu Sharma, Aditya Sanjiv Kanade, Aanand Kumar Yadav

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Saifullah, Thomas Kornmaier, Taaha Kazi, Vasu Sharma, Aditya Sanjiv Kanade, Aanand Kumar Yadav

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are writing a massive, 50-chapter novel. You are the author, and you have a brilliant memory for the first few chapters, but as the story grows, you start to forget the small details that matter most.

The Problem: You remember that the hero found a secret map in Chapter 3, but you forget who else knew about it. Did the villain find out in Chapter 10? Did the hero tell his friend in Chapter 15? If you forget these details, you might write a scene in Chapter 20 where the hero acts surprised by something he already knew, breaking the story's logic.

Current AI tools are like a librarian who can only find books based on keywords. If you ask, "Who knew the secret and when?", the librarian might hand you a page that mentions the secret, but it doesn't tell you who knew it or when they learned it. They might hand you a page from Chapter 25 when the answer was actually in Chapter 12.

The Solution: The "Narrative World Model" (NWM)

The authors of this paper built a new kind of "writer's assistant" called the Narrative World Model (NWM). Instead of just storing text, NWM stores the story logic itself.

Here is how it works, using a simple analogy:

1. The "Story Ledger" (Instead of a Text Dump)

Imagine a standard AI memory is like a giant pile of unorganized sticky notes. You have to read through thousands of them to find one specific detail.

NWM is different. It acts like a specialized ledger or a dramatic spreadsheet. When a chapter is finished, NWM doesn't just save the text; it breaks the story down into specific, labeled facts:

Who knows what? (e.g., "The villain knows the secret now, but the hero doesn't.")
When did things happen vs. when were they revealed? (e.g., "The murder happened in Chapter 2, but the detective only found out in Chapter 10.")
Did a promise get paid off? (e.g., "The hero promised to save the village in Chapter 5; that promise is now 'closed' in Chapter 20.")

It treats these details as distinct, typed entries rather than just words in a paragraph.

2. The "Time-Traveling Detective" (Retrieval)

When you ask the system a question (like "Who knew the secret in Chapter 15?"), a normal system might just search for the word "secret."

NWM acts like a detective with a time machine. It knows that in Chapter 15, the villain didn't know the secret yet, even though the reader found out in Chapter 10. It looks at its "ledger," finds the specific entry for "Villain's Knowledge" at "Chapter 15," and gives you the exact, correct answer. It ignores the rest of the story that isn't relevant to that specific moment in time.

3. The "Fair Test"

To prove this works, the authors set up a strict test:

They took a bunch of tricky questions about long stories that require connecting dots across different chapters (Multi-hop questions).
They gave the questions to different AI systems:
- The "Search Engine" (RAG): Just looks for keywords.
- The "Graph" (Graphiti): A smart system that maps relationships, but treats them like generic facts (e.g., "Person A is related to Person B").
- NWM: Their new system with the story-ledger.
The Catch: They used the same AI "reader" (a very smart model called Opus 4.8) to read the answers provided by all these systems. This ensures the reader wasn't biased; the difference was purely in how the information was retrieved.

The Results: Why NWM Won

The results were clear and significant:

The "Search Engine" got almost everything wrong because it couldn't connect the dots across chapters.
The "Generic Graph" (Graphiti) did better, but it still missed about 40% of the tricky questions. Why? Because it didn't understand story concepts. It knew "Person A" and "Person B" were connected, but it didn't understand the difference between "Event Order" (when something happened) and "Reveal Order" (when the character found out).
NWM got 89.8% of the tricky questions right.

The paper emphasizes that this wasn't because NWM had a "smarter" extractor or a bigger database. Even when they forced the other systems to use the same extraction tools, NWM still won. The advantage came from how the information was organized. NWM organized the story by narrative rules (who knows what, when, and why), while the others organized it by generic facts.

The Bottom Line

Think of it this way:

Old systems are like a library where you have to read every book to find a specific sentence.
NWM is like a story GPS. It doesn't just show you the road; it knows the traffic, the detours, and exactly where you are in the plot. It understands that in a story, who knows what is just as important as what happened.

The paper concludes that for writing long, complex fiction, you don't just need a memory that remembers facts; you need a memory that understands the drama and the timeline of the story. NWM provides that, making it much better at answering complex questions about evolving story states than current technology.

Technical Summary: Narrative World Model (NWM)

Problem Statement

Long-form fiction generation suffers from a disconnect between local fluency and durable narrative state. While models can generate coherent scenes, they often fail to maintain consistency regarding evolving story states, such as character knowledge boundaries (who knows a secret and when), the temporal distinction between event order and reveal order, the status of unresolved promises, and relationship dynamics.

General-purpose retrieval systems (RAG) and agent-memory frameworks typically represent entities and facts but lack the narratological structure required to answer multi-hop questions about story state. For instance, a standard graph might track that an object exists, but not its location relative to a specific chapter's "current" state, or whether a character's knowledge of an event precedes the narration of that event. Existing approaches often surface irrelevant evidence or fail to answer questions requiring evidence from two or more distinct chapters, specifically failing on queries involving focalization, dramatic function, and setup/payoff structures.

Methodology: The Narrative World Model (NWM)

The authors propose the Narrative World Model (NWM), a writer-memory system designed to pair a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. The system operates on finalized chapter prose, treating memory as a record of accepted state rather than future intent.

1. Causal Publish Flow

NWM treats finalized chapters ( $C_n$ ) as the source of truth. Upon publication, an extractor (Claude Sonnet 4.5) reads the prose and emits typed memory records directly, rather than free-text summaries. These records are stamped with their source chapter and evidence spans. The system enforces a strict causal cutoff: queries for chapter $n+1$ can only access updates completed through chapter $n$ . Republishing an edited chapter invalidates downstream edges dependent on the superseded state.

2. Schema and Global Registries

Instead of generic facts, NWM extracts a structured, evidence-backed writer-memory structure containing specific narratological fields:

Character State: Goals, knowledge/unknowns (epistemic boundaries), and relationship deltas.
Plot/Promise: Structural roles, open/closed status, and promised payoffs.
Narrative Function: Focalized observers, dramatic beats, and turn/reversal markers.
Temporal Validity: Records include validity intervals (e.g., when a fact was true vs. when it was revealed).

These records merge into cumulative registries (characters, relationships, objects, threads) that expose the latest causally valid slice for a given checkpoint.

3. Temporal Knowledge Graph (KG)

The KG projects schema states into temporal relations. Edges are typed by the narrative relation they assert (e.g., knowledge, location, relationship-polarity change, thread resolution). Crucially, edges store validity intervals, allowing the graph to return the "state as of chapter $n$ " by selecting the latest valid edge for an entity while preserving history. This distinguishes NWM from flat snapshots or rolling summaries that lose old-but-current states.

4. Query-Conditioned Hybrid Retrieval

To answer a writer query for chapter $n+1$ , NWM employs a four-stage retrieval process:

Causal Restriction: Filters the store to records with source chapters $\le n$ .
Hybrid Node Ranking: Ranks entity/seed nodes using a fusion of BM25 and dense vectors (reciprocal-rank fusion).
One-Hop Expansion: Expands the top-ranked nodes to include their typed one-hop neighborhoods, surfacing the specific narrative structure (e.g., knowledge boundaries, reveal orders) required for the query.
Bounded Packet Assembly: Merges and truncates the evidence into a bounded packet for the reader.

This "conditioning" step is critical; it differs from "NWM State Memory," which simply serializes all current state without query-specific graph search.

5. Evaluation Protocol

The authors evaluate NWM using a held-constant reader (Opus 4.8) to isolate memory and retrieval performance from generator behavior. The reader answers questions based only on the evidence packet provided by each system.

Benchmarks: A private corpus of 5 production-style serialized books (50 chapters each) with a curated 176-item multi-hop benchmark, and a public corpus of 12 public-domain books with a 576-item set (including a 110-item multi-hop subset).
Baselines: Simple Search, RAG (BGE+BM25), GraphRAG, and Graphiti/Zep (the strongest existing temporal-KG framework).
Fairness Controls: Graphiti is extracted using NWM's own extractor (Sonnet 4.5) to ensure the comparison is not confounded by extraction quality. A robustness check re-ingests Graphiti with a cheaper extractor (GPT-4.1-mini), showing no significant accuracy change.

Key Results

Performance on Multi-Hop QA

NWM Graph Retrieval significantly outperforms all baselines on multi-hop narratological questions:

Private Corpus (176 multi-hop items): NWM Graph Retrieval achieved 0.898 accuracy, compared to 0.574 for Graphiti (extractor-matched). The paired McNemar test yielded a result of 64–7 ( $p < 10^{-5}$ ).
Public Corpus (576 items): NWM Graph Retrieval achieved 0.625 accuracy versus 0.516 for Graphiti ( $p = 0.0001$ ).
Multi-hop Subset (Public): NWM reached 0.709 versus Graphiti's 0.582.

Ablation and Analysis

Representation vs. Extraction: The performance gap is representational, not an artifact of extraction. Re-ingesting Graphiti with a cheaper extractor did not change its accuracy ( $p=0.89$ ), and Graphiti's generic schema failed to represent narratological structures (e.g., focalization, reveal order) even when using the same extractor.
Retrieval Conditioning: "NWM State Memory" (serializing all state without query-conditioned search) scored only 0.358 on multi-hop items, failing to beat Graphiti. This demonstrates that the advantage comes from query-conditioned retrieval surfacing specific typed edges, not merely the presence of the typed store.
Failure Modes: On the 64 items where NWM succeeded and Graphiti failed, Graphiti abstained in every case because the gold fact was absent from its retrieved evidence. The failure was due to the absence of narratological structure in the representation, not reasoning errors.
Graph Size: NWM's typed graph is smaller than the source GraphRAG graph yet answers far more questions, indicating that narrative-structured decomposition is the driver, not graph density.

Significance and Claims

The paper claims that narratology-grounded memory is a necessary precondition for high-quality long-form fiction generation. The primary contribution is demonstrating that representing story state through typed narratological units (focalization, reveal order, promise/payoff) combined with query-conditioned retrieval allows a system to answer multi-hop story-state questions with significantly higher accuracy than generic temporal knowledge graphs or standard RAG.

The authors emphasize that the gain is representational: the system succeeds because it explicitly models the structure of narrative (who knows what, when, and how it functions dramatically) rather than just generic entity relations. The paper concludes that while NWM provides the necessary "chapter-safe evidence" for writer questions, the evaluation of whether this leads to better generation (continuation) remains future work. The current results isolate the memory and retrieval components, proving that the representation and retrieval conditioning are the operative mechanisms for success in multi-hop narratological reasoning.