acceptodds
Under review as a conference paper at ICLR 2027

Shared Worlds, Private Minds: Structured Memory for Long-Form Writing as World Creation

Abstract

LLM agents that write long-form fiction need an explicit memory of the evolving storyworld to keep new events consistent with established facts. Such memory must keep heterogeneous narrative information distinct, integrate story developments across granularities, and recover dependencies that a writing request leaves implicit. We present NarraWorld, a structured memory system for long-form writing that treats memory construction as world creation. From a shared evidence-grounded graph, NarraWorld derives four connected views: world facts, per-character beliefs, open developments, and hypothetical branches (possible-world continuations). Hierarchical aggregation with atomic closure consolidates events into scenes, plotlines, and plots, keeping each higher-level node traceable to its constituent source spans. For retrieval, planned reconstruction infers a query's dependencies from the current narrative situation and a preview of memory, then assembles the relevant records within a token budget. Across three writing benchmarks, NarraWorld achieves the strongest aggregate results. Its memory also transfers to situated role-playing and largely preserves recall on a general-purpose long-term memory benchmark, paving the way for agents that sustain coherent storyworlds across diverse narrative tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Narrative Memory, Long-form Narrative Generation, Storyworld Modeling, Character Belief Tracking, Perspective-aware Retrieval, Possible-world Modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TUzpOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes NarraWorld, a structured memory system for long-form writing. It maintains four connected views over a shared narrative graph: world facts, character-specific beliefs, ongoing narrative developments, and hypothetical continuations. It organizes events into scenes, plotlines, and plots while retaining links to their underlying evidence. At retrieval time, a planner uses the writing context and a memory preview to identify implicit information needs and assemble relevant records within a token budget.

Critical strengths
  1. The distinction between world facts and character-specific beliefs is very important to assess narrative coherence, which is not widely covered by previous works. It is also related to the 'theory-of-mind' task, where language models still shows weakness.

  2. The combination of hierarchical memory and dependency planning is well motivated, and I think it can be adopted by other story generation methods.

  3. The evaluation includes plenty number of baselines, ablations studies, and human assessments.

Critical weaknesses
  1. Is there any problem with averaging the scores for the final summary and the 14 chapters? The quality of the overall summary may differ from that of the individual chapters, and including it in the average could inflate the overall score.

  2. This is the most serious concern. What is the practical improvement over the Full Context baseline? NarraWorld outperforms Full Context on LSGC quality and overall TMAS quality, but the gains do not seem substantial enough, considering additional inference cost. Moreover, the paper does not evaluate the same method across stories of different lengths.

  3. Is atomic closure accurate? The paper proves that links exist between higher-level and lower-level nodes, but does not establish whether those links are accurate or meaningful. Further analysis is needed to assess their quality.

Questions for authors
  1. Why didn’t the authors measure agreement among the LLM judges? Although the paper reports inter-human and human–LLM agreement, it does not report agreement among the LLM judges. I think this should also be measured. I understand that the inter-human and human–LLM agreement scores are relatively low, but this is common in story evaluation, so I will not focus on those values in my assessment.

  2. Providing an anonymous link to the generated stories would be very helpful, as it would allow readers to assess their quality directly.

I can give you higher overall score if you appropriately prove that NarraWorld has significant improvements than Full Context baseline even considering the huge amount of additional cost, and that these improvements hold across various lengths of stories. I recommend analyzing your method from other perspectives, such as its effect on diversity, or training models using your method.

8 comments
AuthorsOpenReview ↗

Response to Reviewer TUzp (1/3)

We thank you for highlighting the distinction between world facts and character beliefs, the memory hierarchy, and the evaluation coverage. We address the three weaknesses and two questions below, incorporating the additional diversity suggestion into W2.

W1: Chapter and final-summary score aggregation

Thank you for raising this concern. For LSGC, where summary evaluation contributes to the overall score, we report chapter-only, final-summary, and combined scores separately for all eight methods below. Scores are on a 0-100 scale, with higher values indicating better quality.

MethodChapter-only ↑Final summary ↑Combined ↑
NarraWorld90.2889.6890.24
StructMem86.9886.0886.92
Full Context85.1183.9185.03
Mem080.3078.5080.18
Graphiti79.2377.2879.10
Hindsight75.0772.6774.91
AriGraph73.2070.6573.03
Scaffold Only69.8166.8169.61

Final-summary scores are lower than chapter-only scores for every method, so including the summary does not inflate the absolute Overall scores in this comparison. The ranking is unchanged across all three columns. We retain the original combined score in the primary 40k table and report the decomposition alongside it; the new cross-length study uses chapter-only scores throughout.

Revision: Appendix I.2 (pp. 40–41, lines 2149–2175, Table 26) reports chapter-only, final-summary, and combined scores for all eight methods on LSGC.

W2: Advantages over Full Context: Long-Form Quality, Originality, and Diversity

Thank you for asking whether the gains over Full Context justify the inference cost. We directly compare both methods on stories generated on the LSGC dataset at 10k, 20k, 40k, and 80k words, using the same starters and three runs per setting.

Quality Across Lengths. DeepSeek-V4-Pro and Gemini (gemini-3.1-pro-preview) evaluate the same anonymized chapter texts using the same 16-criterion rubric. Scores are chapter-only averages on a 0-100 scale, excluding final summaries; higher is better. Δ is NarraWorld minus Full Context.

LengthJudgeFull ContextNarraWorldΔ
10kDeepSeek88.5090.89+2.39
10kGemini89.9892.34+2.36
20kDeepSeek87.1690.66+3.50
20kGemini88.6492.12+3.48
40kDeepSeek85.1190.28+5.17
40kGemini86.6291.75+5.13
80kDeepSeek80.5289.15+8.63
80kGemini82.0190.61+8.60

NarraWorld's quality advantage grows with story length. From 10k to 80k, its lead over Full Context increases from 2.39 to 8.63 points under DeepSeek and from 2.36 to 8.60 under Gemini. Both judges therefore show a widening advantage as the narrative horizon expands. The separate human evaluation gives NarraWorld the highest mean Human Overall at all four lengths; Appendix F.6 reports its sample sizes and results.

Chapter-Position Diagnostic. We additionally examined the chapter-only means by position for the 14-chapter 40k and 28-chapter 80k conditions. The late-minus-early change is −6.32 and −12.72 for Full Context, −3.73 and −7.50 for StructMem, and −1.48 and −3.53 for NarraWorld, respectively. This position-wise diagnostic shows a smaller decline for NarraWorld and a widening late-story margin over Full Context; the appendix reports the per-chapter points and descriptive LOWESS trends.

Computational Cost. Resource values are means per completed story, covering writing, memory construction, and planning, and excluding evaluation-judge usage. Tokens include input and output; runtime is end-to-end wall-clock time.

LengthMethodTotal tokens (k)Time (min)
10kFull Context105.311.5
10kNarraWorld340.817.2
20kFull Context277.322.8
20kNarraWorld516.031.4
40kFull Context940.651.0
40kNarraWorld1,115.861.5
80kFull Context3,448.0134.0
80kNarraWorld2,352.8126.8

NarraWorld's relative token and runtime costs decline as stories grow. From 10k to 80k, the token-cost ratio relative to Full Context falls from 3.24 to 0.68, and the runtime ratio from 1.50 to 0.95. At shorter lengths, the quality gains come with additional token and runtime costs; at 80k, NarraWorld delivers its largest quality gain while using 31.8% fewer tokens and taking 5.4% less time. Full Context repeatedly supplies the entire accumulated history for each writing step, whereas NarraWorld reconstructs the relevant memory context for the current request. The measured total costs include both context processing and memory construction and planning.

AuthorsOpenReview ↗

Response to Reviewer TUzp (2/3)

W2: Advantages over Full Context: Long-Form Quality, Originality, and Diversity (Continued)

Originality and Diversity. Following your suggestion, we use the publicly released MoPS collection (Ma et al., 2024) and adopt its Originality and Fascination metrics to assess originality and narrative appeal, complementing the long-story quality and cost comparisons above. MoPS provides varied premises for open-ended story generation, synthesized from components such as themes, settings, characters, and plots. We sample 200 premises from this collection and generate one story with each method per input, yielding 200 matched pairs. DeepSeek-V4-Pro and Gemini evaluate the generated stories on Originality, the perceived novelty relative to familiar stories, and Fascination, the perceived appeal of the narrative. Both use a 0-100 scale, with higher scores indicating better performance.

Both judges favor NarraWorld on these two dimensions. The table reports paired mean differences (NarraWorld minus Full Context); brackets give 95% bootstrap intervals for Originality.

DimensionDeepSeek differenceGemini difference
Originality+15.67 [13.02, 18.29]+11.21 [9.31, 13.16]
Fascination+1.08+1.15

We additionally measure semantic diversity across the stories generated from these inputs using BGE-M3 embeddings (Chen et al., 2024). Pairwise distance and distance to the collection center measure overall semantic spread; nearest-neighbor distance measures local distinctiveness. Larger distances indicate greater diversity under this representation. NarraWorld improves all three collection-level measures:

Collection diversity measureChange relative to Full Context
Mean pairwise cosine distance+6.8%
Distance to the group center+7.3%
Mean 3-nearest-neighbor distance+11.4%

Taken together, these results show an increasingly favorable quality-efficiency trade-off as stories grow, alongside gains in originality, narrative appeal, and collection diversity. NarraWorld is therefore particularly well suited to long-form writing.

Revision: Section 3.2 (p. 8, lines 382–394) summarizes the length-dependent comparison and additional evaluation. Appendix I.2 (p. 40, lines 2120–2147, Table 25) reports quality, Appendix F.7 (pp. 35–36, lines 1884–1908, Figure 6) reports the chapter-position diagnostic, and Appendix F.6 (p. 35, lines 1844–1881, Table 24) reports full-story human evaluation. Appendix I.1 (p. 40, lines 2108–2118) defines resource accounting, Appendix I.3 (pp. 41–42, lines 2185–2220, Table 27) reports resource use, and Appendix L (pp. 48–49, lines 2540–2594, Tables 34–35) reports originality and collection diversity.

W3: Atomic closure accuracy and meaning

Thank you for asking us to verify the quality of atomic closure. We audited hierarchical memory from 15 complete NarraWorld LSGC stories, sampling 30 scenes, 30 plotlines, and 30 plots across early, middle, and late snapshots, together with 846 node-member associations.

Correctness and Meaningfulness. Correctness requires the linked record to be supported by the source text and correctly associated with that scene, plotline, or plot. Meaningfulness requires the association to provide substantive evidence for its content or development; merely sharing a character or background is insufficient. Both requirements must be met to pass the association check. We additionally report reference resolution and summary fidelity.

Evaluation Procedure. Two annotators checked memory against the source prefix available at each snapshot. Reference resolution and association accuracy use sampled links as their denominator; summary fidelity uses sampled nodes. Before discussion, annotator agreement on closure associations was 94.2%, with Cohen's kappa of 0.63.

LevelReferences resolvingCorrect and meaningful associationsNode summaries passing
Scene254/254 (100.0%)243/254 (95.7%)28/30 (93.3%)
Plotline290/292 (99.3%)271/292 (92.8%)26/30 (86.7%)
Plot297/300 (99.0%)272/300 (90.7%)25/30 (83.3%)
Overall841/846 (99.4%)786/846 (92.9%)79/90 (87.8%)

Overall, 92.9% of sampled links are both correct and meaningful, with pass rates above 90% at every level. These links provide substantive evidence for the corresponding abstractions, enabling higher-level memory to be traced back to the events and states that support it. Together with the summary-fidelity results, this supports the reliability and practical role of atomic closure in evidence reconstruction.

Revision: Section 3.2 (p. 8, lines 396–400) summarizes the audit. Appendix J.1 (p. 43, lines 2270–2303) describes sampling and annotation; Appendix J.3 (pp. 44–45, lines 2339–2380, Table 30) reports closure accuracy and summary fidelity.

AuthorsOpenReview ↗

Response to Reviewer TUzp (1/3)

Q1: Agreement among LLM judges

Thank you for suggesting this consistency check. We added Gemini (gemini-3.1-pro-preview) alongside DeepSeek-V4-Pro to evaluate stories generated on the LSGC dataset at 10k, 20k, 40k, and 80k words by Full Context, StructMem, and NarraWorld. Both judges evaluate the same anonymized chapter texts with identical prompts and the same 16-criterion rubric. Story scores average chapter scores on a 0-100 scale and exclude final summaries.

Agreement Metrics. Spearman correlation measures agreement in story rankings; MAE measures the average absolute score difference; mean offset is Gemini minus DeepSeek. The last column reports how often the two judges agree on the sign of the NarraWorld-minus-Full Context score difference across 15 matched starter–run pairs; this reports agreement in comparison direction.

LengthSpearman ρMAEMean offsetNW vs. FC direction agreement
10k0.731.82+1.4713/15 (86.7%)
20k0.701.95+1.4814/15 (93.3%)
40k0.682.12+1.4914/15 (93.3%)
80k0.632.39+1.4815/15 (100.0%)

At each length, these metrics cover stories from all three methods. The two judges show substantial positive correlation in story rankings and small absolute score differences. Ranking methods by their mean story scores also yields the same ordering at every length: NarraWorld, StructMem, then Full Context. These results provide consistent support from both evaluators for our main comparative findings.

Revision: Appendix F.5 (pp. 34–35, lines 1782–1842, Tables 22–23) reports cross-judge agreement. The blue-marked addition in Section 3.2 (p. 8, lines 379–380) directs readers to this analysis.

Q2: Anonymous access to generated stories

Thank you for requesting access to the generated stories. They are available through the following anonymous repository:

Anonymous generated-story examples

Finally, thank you again for your thoughtful feedback! Your questions helped us sharpen the evaluation and clarify how the results should be interpreted. We learned a great deal from your suggestions, particularly about examining narrative quality from complementary perspectives, and explored more useful and interesting datasets&benchmarks in the field of AI Writing. The additional analyses and revisions have strengthened the paper, and we appreciate the time and care you have devoted to reviewing our work.

Reviewer_fVx3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes NarraWorld, a structured memory system for long-form story generation. Instead of storing narrative history as one flat memory, it separates information into shared world facts, character-specific beliefs, ongoing developments, and possible future branches. It also organizes events at different levels and uses a planning step to retrieve the memory needed for the next writing action.

Critical strengths
  1. The memory design matches the needs of long-form fiction. Separating shared facts from what individual characters know or believe is useful for avoiding inconsistencies in character behavior. The distinction between world state, private beliefs, developments, and possible worlds is also easy to understand.

  2. The paper studies the method across several tasks. The evaluation includes three long-form writing benchmarks, role-playing, and a general long-term memory benchmark. NarraWorld performs well across most of these settings and often outperforms existing memory systems and full-context baselines.

  3. The ablation studies are useful. The paper separately removes character beliefs, possible worlds, and cross-event organization. The results show that these components contribute differently, with cross-event organization having a large effect on long-form writing quality.

Critical weaknesses
  1. Limited architectural novelty: NarraWorld does not introduce a new neural architecture. Its main contribution is the organization of narrative memory into several structured views over a shared graph, including world facts, character beliefs, developments, and possible worlds. While this design is useful for long-form storytelling, many of the underlying ideas, such as graph-based memory, typed memory stores, hierarchical organization, and retrieval planning, have been explored in prior work. The paper should more clearly explain what is technically new in the proposed memory architecture beyond combining these existing ideas for narrative generation.

  2. The contribution appears partly system-level rather than algorithmically new. Several ideas used here, such as structured memory, graph-based memory, hierarchical summaries, retrieval planning, and character state tracking, have been explored separately in prior work. The main contribution is their organization into a narrative-specific memory framework. The paper should explain more clearly which parts are technically new compared with methods such as StructMem, Graphiti, Hindsight, and other story-state tracking approaches.

  3. The evaluation depends substantially on LLM-based judges. Many of the reported writing-quality scores are based on automated evaluation. The paper includes human evaluation, which helps, but the human study is more limited than the automatic evaluation. Since qualities such as plot, creativity, and character consistency can be subjective, broader human evaluation would strengthen the conclusions.

  4. The computational cost is not studied in enough detail. The proposed system performs several memory-construction and planning operations for each update. The authors mention construction and planning cost as a limitation, but there is little analysis of latency, number of LLM calls, token usage, or monetary cost compared with simpler memory methods.

  5. It is not clear how well the system scales to much longer stories. The experiments show benefits on the selected benchmarks, but the paper would be stronger with a direct analysis of performance as narrative length increases. In particular, it would be useful to study whether the memory graph, hierarchical structure, and retrieval process remain effective for book-length stories with many characters and overlapping plotlines.

  6. The role of possible worlds needs more justification. Possible future branches help on creative writing tasks, but they can also introduce speculative information into memory. The LoCoMo experiment itself shows that exposing these records can reduce factual recall. The paper should discuss more clearly when hypothetical memory should be created, retained, removed, or hidden from the model.

  7. The method is complex and depends heavily on LLM-generated memory structures. NarraWorld requires entity resolution, fact extraction, character-access tracking, development tracking, possible-world construction, hierarchical aggregation, dependency planning, and retrieval. Errors at any of these stages may propagate to the final generation. The paper does not provide enough analysis of how often these intermediate memory structures are incorrect.

Questions for authors

Please refer to weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer fVx3 (1/4)

We thank you for recognizing the narrative memory design, broad evaluation, and ablation studies. We address the seven concerns below.

W1: Technical contribution

Thank you for encouraging us to clarify the technical contribution. We agree that graph-based memory, hierarchical organization, and retrieval planning are established primitives. Our contribution is to incorporate narrative-specific information status and access constraints throughout the memory computation process.

Perspective-Aware Memory. NarraWorld maintains shared world facts alongside character-specific beliefs and their acquisition paths. An event can therefore update the shared world without changing what every character knows. These distinctions are preserved through narrative development, so a continuation can respect both the established world state and a character's incomplete or outdated knowledge.

Cross-Event Evidence Organization. NarraWorld organizes accepted events into scenes, plotlines, and plots while retaining atomic records, source spans, and character-specific perspectives through atomic closure. This allows a multi-scene development to be retrieved as a higher-level unit and then unfolded into the evidence admissible for the current narrative situation.

Situated Memory Reconstruction. NarraWorld conditions reconstruction on story time, disclosure state, focalization, and branch. A planner reads a preview of the admissible memory to infer implicit dependencies, retrieves across narrative scales, and expands selected abstractions into supporting evidence within a token budget. The same stored history can therefore support different narrative situations while preserving belief holders and evidential status.

Together, these designs address a central challenge of long-form fiction: factual consistency alone does not determine what a character may know, how a relationship has developed, or when a revelation may be disclosed. We clarify these distinctions and their relationship to prior work in the revised paper.

Revision: The blue-marked edits to the contribution statements in Section 1 (p. 2, lines 90–101) clarify the contribution. The added sentences in Section 2.2 (p. 4, lines 194–196) and Section 2.4 (p. 5, lines 235–237) explain how information status and acquisition relations established during construction guide access and planning and constrain evidence expansion and assembly.

W2: Comparison with prior memory and story-state methods

Thank you for requesting a direct comparison. We clarify the key distinctions and their implications for narrative generation below.

StructMem: Narrative Views and Reconstruction. StructMem binds factual and relational entries and consolidates cross-event information. NarraWorld connects hierarchical abstractions to character-held beliefs, acquisition paths, development states, and conditional branches, then reconstructs the evidence admissible for the current writing situation. This preserves local facts and character perspectives when retrieving a longer development.

Graphiti: Temporal and Character-Specific Access. Graphiti maintains temporal facts, source episodes, and community summaries. NarraWorld also records who acquired information and how. A world-state change can therefore coexist with a character's earlier belief, supporting continuations that respect differences between what happened and what each character knows.

Hindsight: Narrative Dependencies. Hindsight separates facts from opinions and supports evidence-based updates and agentic retrieval. NarraWorld organizes character beliefs, unresolved developments, and conditional continuations for writing. Its planner uses a preview of admissible memory to identify evidence dependencies that the current request leaves implicit.

Story-State Methods: Reconstructing Writing Context. FactTrack tracks temporal facts; DOME interleaves outlining and writing with temporal-graph retrieval; MAGNET models character actions, state updates, and goals; and character-belief planners represent differing knowledge. NarraWorld connects these narrative distinctions to memory extracted from accepted prose and to context reconstruction across events, scenes, plotlines, and plots.

Revision: The blue-marked edits in Section 4 expand the comparisons with story-state methods (pp. 9–10, lines 483–489), character-belief planners (p. 10, lines 498–501), and StructMem, Graphiti, and Hindsight (p. 10, lines 510–517).

AuthorsOpenReview ↗

Response to Reviewer fVx3 (2/4)

W3: Expanded full-story human evaluation

Thank you for highlighting the need for broader human evaluation. We expanded the full-story study to 54 LSGC stories across 10k, 20k, 40k, and 80k words, covering Full Context, StructMem, and NarraWorld. This adds 39 stories to the original 15-story 40k sample, which is reused and counted only once. That original sample comprises the fixed run_1 across all five starters for each method (15 complete 14-chapter stories, each with its final summary); its human–judge alignment pairs scores for those exact story instances.

Evaluation Scope and Protocol. We sample complete stories for human evaluation: each method contributes 5, 5, 5, and 3 stories at the four lengths, respectively. The 15-story human totals at 10k, 20k, and 40k each combine five stories from three methods; automatic evaluation covers all 15 generated instances per method and length. The smaller 80k sample reflects the greater reading effort required at that length. Two blinded assessors independently rate each complete story, yielding 108 story-level rating sets, including 78 new ratings. Human Overall uses the five dimensions Character, Plot, World/Coherence, Style/Creativity, and Prompt Adherence, with weights of 3:4:2:6:1. We average the two readers within each story before averaging stories within each method and length.

Results and Annotator Agreement. Scores use a 0–100 scale. Counts are listed as Full Context / StructMem / NarraWorld (FC / SM / NW). Rater MAE measures the mean absolute difference between the two assessors' story-level Human Overall scores, pooling all three methods at each length: 15, 15, 15, and 9 stories, respectively.

Length / sampleStories (FC / SM / NW)Full ContextStructMemNarraWorldRater MAE
10k / new5 / 5 / 582.6082.6785.505.55
20k / new5 / 5 / 582.1181.6884.886.00
40k / original5 / 5 / 580.0680.5084.506.48
80k / new3 / 3 / 374.5177.8782.966.58

NarraWorld has higher mean Human Overall than both baselines at all four lengths. Its margins over Full Context are +2.90, +2.77, +4.44, and +8.45 points; those over StructMem are +2.83, +3.20, +4.00, and +5.09 points.

Revision: The blue-marked summary in Section 3.2 (pp. 7–8, lines 377–379) reports the expanded full-story human evaluation and its main finding. Supporting details on sampling, scores, and rater MAE appear in Appendix F.6 (p. 35, lines 1844–1881, Table 24).

W4: Computational cost

Thank you for asking for a clearer cost analysis. We directly compare NarraWorld with StructMem, a simpler structured-memory baseline, at 10k, 20k, 40k, and 80k words. The evaluation uses stories generated on the LSGC dataset, with five starters and three runs for each method-length condition.

Calls, Tokens, Runtime, and Monetary Cost. The table reports mean LLM request counts, total input/output tokens, end-to-end wall-clock time, and estimated LLM token cost per completed story. Accounting includes writing and auxiliary operations such as memory construction and planning, and excludes evaluation-judge usage. Tokens are in thousands; time is in minutes and includes API latency.

LengthMethodLLM callsTotal tokens (k)Time (min)Est. LLM cost (USD)
10kStructMem42.00215.014.80.323
10kNarraWorld68.00340.817.20.554
20kStructMem76.00399.027.50.583
20kNarraWorld122.00516.031.40.661
40kStructMem154.13866.955.21.222
40kNarraWorld246.001,115.861.51.391
80kStructMem318.001,871.2116.52.516
80kNarraWorld495.002,352.8126.82.844

USD estimates apply the official standard, uncached input/output rates to the writer and auxiliary token breakdowns in Table 27: GPT-5 for writing ($1.25/$10.00 per million tokens) and GPT-5-mini for auxiliary operations ($0.25/$2.00), checked on November 14, 2026. These are token-based estimates without caching or batch discounts; they exclude embedding and reranking costs.

Quality-Efficiency Trade-off. NarraWorld requires more LLM calls and tokens than StructMem. This translates into a moderate increase in end-to-end runtime, while NarraWorld achieves stronger writing quality. As stories grow, its relative overhead decreases, making the quality-efficiency trade-off increasingly favorable for long-form generation.

Revision: Section 3.2 (p. 8, lines 382–387) summarizes the resource comparison. Appendix I.1 (p. 40, lines 2108–2118) defines accounting, and Appendix I.3 (pp. 41–42, lines 2185–2220, Table 27) reports requests, tokens, time, and estimated LLM costs, with the pricing formula and sources.

AuthorsOpenReview ↗

Response to Reviewer fVx3 (3/4)

W5: Performance across story lengths

Thank you for suggesting a direct evaluation across story lengths. We compare NarraWorld, Full Context, and StructMem on complete stories generated on the LSGC dataset at target lengths of 10k, 20k, 40k, and 80k words, corresponding to 4, 7, 14, and 28 chapters. The 80k-word condition corresponds to book-length stories.

Quality Across Lengths. DeepSeek-V4-Pro and Gemini (gemini-3.1-pro-preview) evaluate the same anonymized chapter texts using the same 16-criterion rubric. Scores are direction-corrected to a 0–100 scale, where higher is better. We average chapter scores within each story and then average stories; final-summary judgments are excluded at every length.

Target lengthJudgeFull ContextStructMemNarraWorldΔ vs. Full Context
10kDeepSeek88.5089.1590.89+2.39
10kGemini89.9890.6292.34+2.36
20kDeepSeek87.1688.1690.66+3.50
20kGemini88.6489.6592.12+3.48
40kDeepSeek85.1186.9890.28+5.17
40kGemini86.6288.4891.75+5.13
80kDeepSeek80.5284.3589.15+8.63
80kGemini82.0185.8390.61+8.60

Length-Dependent Advantage. Both judges rank NarraWorld first at every tested length. Its mean advantage over Full Context increases from 2.39 to 8.63 points under DeepSeek and from 2.36 to 8.60 under Gemini. Its advantage over StructMem increases from 1.74 to 4.80 and from 1.72 to 4.78 points, respectively. These automatic assessments show a widening advantage over both baselines as stories grow. The separate human evaluation in W3 also gives NarraWorld the highest mean Human Overall at all four lengths. W4 reports the associated resource trade-offs.

Chapter-Position Diagnostic. We also examined the chapter-only means by position for the 14-chapter 40k and 28-chapter 80k conditions. The late-minus-early change is −6.32 and −12.72 for Full Context, −3.73 and −7.50 for StructMem, and −1.48 and −3.53 for NarraWorld, respectively. Thus, the position-wise diagnostic shows a smaller decline for NarraWorld and a widening late-story margin over Full Context. The appendix reports the per-chapter points and descriptive LOWESS trends.

Together, its performance across story lengths, its advantage at 80k words, and its stronger relative performance in later chapters support NarraWorld's ability to sustain writing quality over extended narratives.

Revision: Section 3.2 (p. 8, lines 382–387) and Appendix I.2 (p. 40, lines 2120–2147, Table 25) report quality across lengths; Appendix F.7 (pp. 35–36, lines 1884–1908, Figure 6) adds the chapter-position diagnostic. Appendix F.5 (p. 34, lines 1782–1811, Table 22) reports the paired comparisons under both judges.

W6: Possible-world lifecycle and access

Thank you for asking us to clarify the role and lifecycle of hypothetical memory. H is an optional prospective-memory view: it is enabled for writing tasks and disabled for general-purpose factual recall. When enabled, it supplies conditional alternatives for prospective query planning and creative continuation. Removing H lowers LSGC Overall from 90.24 to 88.74 and reduces creativity and development on TMAS.

Creation, Retention, and Access. We use the following rules:

  • Create candidates for unresolved developments when their premises are supported by the current world and character views.
  • Retain candidates while their premises remain valid and the associated development remains unresolved, even when the latest passage does not mention them. If partially realized, retain only the unrealized continuation.
  • Remove candidates from the active view when their premises are negated, their development is resolved or abandoned, or they are fully realized. Historical records remain stored.
  • Enable H for writing, exposing branches as explicitly hypothetical alternatives. Disable H for general-purpose factual recall, excluding these branches from the model-accessible memory. On LoCoMo, exposing H reduces Overall Main from 0.614 to 0.570 under the same 8k budget, supporting the default exclusion of H in factual recall.

Separation from Established Facts. Retrieving or selecting a branch does not make it a world fact; realization must be supported by accepted source text. This allows NarraWorld to use future alternatives to support creative development while keeping factual recall grounded in established evidence.

Revision: The blue-marked additions in Section 2.2 (p. 4, lines 191–194) clarify H's lifecycle rules and separation from established facts. Appendix K (pp. 46–47, lines 2430–2498, Tables 32–33) reports the lifecycle evaluation and its interpretation.

Reviewer_UrVKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents NarraWorld, a new memory system for LLM-driven long-form story generation. The approach relies on creating and maintaining structural representations tailored to long-form narrative writing specifically. For example, the memory contains information not only about the actual state of the fictional world, but also current plot lines, each character's beliefs, and so on. The resulting system shows strong overall performance across a range of benchmarks.

Critical strengths
  • The paper is structured in a clear and focused manner.
  • The authors spend sufficient time on ablations, investigating the contributions of different parts of the proposed system.
  • The authors complement automated benchmark evaluation with human-in-the-loop assessment, which strengthens the reliability of the obtained results.
  • The proposed method shows strong performance overall.
Critical weaknesses

I did not find any critical weaknesses.

One thing I would like to see addressed more fully is the discussion of the potential impact of the paper on human writers and writing industry in general. AI-assisted writing is a controversial topic in the writing community, and I believe it needs to be, in the very least, acknowledged.

Questions for authors
  1. "When an event changes an existing state, NarraWorld preserves the validity period of the previous state and establishes a new current state instead of overwriting the earlier record." I am not sure I fully understand this? Does it imply that every update creates a full copy of the world state, rather than modifying the existing one?

  2. "Human readers support the main writing conclusion: on ConStory-Bench, NarraWorld and Full Context have nearly identical human CED (0.095 and 0.096), both below Graphiti (0.289)." minor, but it'd be good to remind the reader here that "lower is better" for this metric.

2 comments

We thank you for the positive assessment of the presentation, ablations, and human evaluation, and for the suggestions below.

W1: Impact on human writers and the writing industry

Thank you for highlighting the relevance of this work to human writers and the writing industry. This research investigates structured memory for maintaining consistency in fictional narratives. NarraWorld can serve as a tool for inspiring creative ideas and managing narrative states: authors could inspect world states, unresolved developments, character knowledge, and supporting evidence when checking continuity or making targeted corrections. Authors retain control over creative direction and the final acceptance of generated material. The research does not advocate unauthorized use of writers' work or presenting AI-generated material as human-authored. Any application to creative writing should respect authors' creative agency and the rights associated with source materials.

Revision: Section 5 (Discussion; p. 10, lines 521–527) describes the potential author workflow and authors' control over creative decisions. The blue-marked paragraph in the Ethics Statement (p. 11, lines 566–570) clarifies the research scope, creative agency, and respect for source rights.

Q1: Record-level state updates

Updates operate on the affected state record. When an event changes a property, NarraWorld closes that property's previous validity interval and appends a record for the new value. Shared entity, event, and provenance links are reused; unrelated state records remain unchanged.

For example, if a key moves from a drawer to a hollow book, the drawer-location record remains as historical evidence with a closed validity interval, and a new book-location record becomes current. Time-conditioned access can recover the appropriate location at an earlier or current point in the story. This is record-level temporal versioning in an append-only store; an update does not copy the entire world state.

Revision: The blue-marked addition in Section 2.2 (p. 4, lines 183–186) clarifies record-level temporal versioning and provides the key-location example.

Q2: CED direction

Thank you for this suggestion. CED measures consistency-error density, so lower is better. The human CED scores of 0.095 for NarraWorld and 0.096 for Full Context indicate nearly identical performance, with both below Graphiti's 0.289. We explicitly repeat the direction beside these scores in the main text, alongside the existing metric definition and table caption.

Revision: Section 3.2 (p. 7, line 373) adds "CED is lower-is-better" beside the human scores.

Thank you again for your helpful suggestions. We have added a discussion of potential uses for human writers, authors’ creative control, and respect for source rights. We also clarify that state updates version only the affected records and explicitly state that lower CED is better beside the human-evaluation scores.

The corresponding revisions are marked in the updated manuscript. We would appreciate knowing whether these changes address your questions, and would be happy to clarify anything further.

Comments on the paper

1 comment

Dear reviewers,

The authors have submitted their rebuttals. Please take a look and interact with the authors if necessary.

Thanks a lot.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.