PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Qiuyu Tian, Xiaowen Gu, Hang Su, Jianghan Chao, Haojie Yin, Fan Guo, Xin Zhang, Jinjing Shen, Ewing Luo, Youyong Kong, Yingce Xia, Zequn Liu

arXiv:2609.32401v1cs.CLcs.AI

TL;DR

Long-form writing needs explicit memory to maintain consistency as story histories grow and relevant information becomes harder to access. NarraWorld builds connected, evidence-grounded views of the storyworld and reconstructs request-specific memory; it achieves the strongest aggregate writing results across three benchmarks and also transfers to situated role-playing.

  • Problem

    Long-form writing needs explicit memory because story histories can outgrow model context, while heterogeneous narrative information can conflict and writing requests may omit dependencies.

    Source: §1 · pp. 1–2

  • Method

    NarraWorld derives world facts, per-character beliefs, developments, and hypothetical branches from a shared evidence-grounded graph, then uses hierarchical aggregation and planned reconstruction.

    Source: §1 · p. 2

  • Results

    Across three writing benchmarks, NarraWorld achieves the strongest aggregate writing results, transfers to situated role-playing, and largely preserves recall on a general-purpose long-term memory task.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    The same structured memory supports long-form writing, situated character action, and general-purpose memory use, with the admissible memory view conditioned on the task.

    Source: §5 · p. 10

    The system prioritizes explicit narrative structure at higher modeling cost, averaging about 200 modeling calls per story, and reducing extraction and planning overhead remains future work.

    Source: G LIMITATIONS · p. 31

Abstract

from arXiv · show

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.

1 INTRODUCTION

Long-form fiction requires explicit memory to keep new events consistent as story histories outgrow a model’s context. NarraWorld addresses the challenges of preserving distinct narrative statuses, integrating developments across passages, and retrieving dependencies omitted from writing requests.

1 INTRODUCTION

  • Explicit memory is needed because long-form story histories can outgrow a model’s context, and relevant information becomes harder to access when distant or in the input’s middle.Narrative consistency requires new events to agree with facts already established in the story.Source: §1 · p. 1
  • Memory must keep objective facts, character beliefs, plot developments, and foreshadowed possibilities distinct because they can conflict within the same storyworld.Collapsing these statuses can conflate information that should remain distinct and lead to story contradictions.Source: §1 · pp. 1–2
  • Narrative developments can accumulate across many passages, so memory must summarize them across multiple granularities without losing specific facts.Existing memories are often flat summaries, isolated facts, or event-level records.Source: §1 · p. 2
  • Similarity-based retrieval is bounded by the request text and can miss implicit dependencies, such as relationships affected by a requested scene revision.A request to rewrite a betrayal scene may also require memories of the character relationships it affects.Source: §1 · p. 2
  • NarraWorld derives four connected memory views from a shared evidence-grounded graph and uses planned reconstruction to infer implicit continuation dependencies before retrieval.Hierarchical aggregation with atomic closure organizes events into scenes, plotlines, and plots while preserving links to constituent events and source spans.Source: §1 · p. 2
  • Across three writing benchmarks, NarraWorld achieves the highest aggregate writing quality and transfers to situated role-playing.On a general-purpose long-term memory task, its memory largely preserves recall performance.Source: §1 · p. 2

2 NARRAWORLD: STRUCTURED MEMORY AS WORLD CREATION

NarraWorld constructs a shared narrative structure and reconstructs memory for each writing action. Its connected views, hierarchical organization, and planned retrieval preserve story evidence while adapting memory to narrative needs.

2 NARRAWORLD: STRUCTURED MEMORY AS WORLD CREATION

  • NarraWorld combines a shared evidence-grounded substrate, a hierarchy of events through plots, and planned reconstruction that infers dependencies before assembling admissible context under a budget.The planner uses the current narrative situation and a memory preview.Source: §2 · p. 3, §2.4 · p. 5

2.1 TASK FORMULATION

  • The writing request combines a scaffold with narrative-situation details that specify teller, recipient, focalizer, story time, discourse position, and applicable branch.The scaffold contains the global outline, rolling summary, preceding passage or chapter, current request, and output constraints.Source: §2.1 · p. 3

2.2 MULTI-VIEW MEMORY OVER A SHARED NARRATIVE GRAPH

  • One evolving narrative graph projects into world-state, character-belief, development, and hypothetical-branch views, keeping shared facts distinct from perspective-dependent beliefs.Open developments can have bounded possible-world continuations; characters’ knowledge may differ from the world state.Source: §2.2 · pp. 3–4

2.3 ORGANIZING MEMORY ACROSS NARRATIVE SCALES

  • The hierarchy organizes atomic events into scenes, persistent plotlines, and plots, while atomic closure preserves links to member records and source spans.Higher-level units can be unfolded into relevant actions, beliefs, and source evidence; hypothetical records remain outside the hierarchy until realized in source text.Source: §2.3 · pp. 4–5

2.4 RECONSTRUCTING MEMORY FOR SITUATED WRITING

  • Reconstruction conditions access on narrative situation, plans for missing dependencies from a memory preview, retrieves at suitable narrative scales, and assembles evidence within a budget.The admissible view accounts for temporal, disclosure, and branch scope; higher-level matches unfold into supporting records.Source: §2.4 · pp. 5–6

3 EXPERIMENTS

NarraWorld leads the reported aggregate writing results across three benchmarks, with human evaluation supporting the writing conclusion. Ablations, case studies, and transfer evaluations examine its components and broader use.

3.1 EXPERIMENTAL SETUP

  • The experiments cover ConStory-Bench, TMAS, and LSGC, alongside RMTBench role-playing and LoCoMo recall, with shared task inputs and controlled writing models and protocols.Memory budgets are 12k tokens for ConStory-Bench and LSGC, 4k for RMTBench, and 8k for LoCoMo; LoCoMo sensitivity tests range from 2k to 16k.Source: §3.1 · p. 6

3.2 OVERALL RESULTS

  • NarraWorld leads all six reported ConStory-Bench measures, including CED of 0.185 versus 0.200 for Full Context, and leads LSGC Overall at 90.24 and TMAS Creativity, Development, and Quality.Human readers report LSGC Human Overall of 84.50 versus 80.50 for StructMem and 80.06 for Full Context; ConStory-Bench human CED is nearly identical for NarraWorld and Full Context, 0.095 and 0.096.Source: §3.2 · pp. 6–7, Table 1 · p. 7, Table 2 · p. 7

3.3 ABLATION STUDIES

  • Removing cross-event organization causes the broadest degradation, while removing character beliefs weakens perspective-sensitive continuity and removing possible worlds lowers creativity and development.Preview-guided retrieval planning improves every ConStory-Bench task with offline memory and budget fixed, with the largest gains on continuation and completion.Source: §3.3 · pp. 7–8, Figure 3 · p. 8

3.4 CASE STUDIES

  • A continuation case combines an unfinished fox drawing, the narrator’s knowledge about their mother’s box, and the father’s promise to give them space.The continuation completes the drawing while the father stays outside and does not ask to see it.Source: §3.4 · p. 8

3.5 EXTENSION TO SITUATED ROLE-PLAYING

  • On RMTBench, NarraWorld leads all four 1–5 dimensions, including plot advancement and character understanding, while Full Context is slightly higher on percentage-based CM and UPA.The table reports scores averaged over applicable turns; EE, EC, PA, and CU use 1–5 scores, while CM and UPA are percentage success rates.Source: §3.5 · p. 8, Table 5 · p. 8

3.6 GENERAL-PURPOSE MEMORY

  • At the shared 8k LoCoMo budget, NarraWorld leads or ties four of five measures, with Graphiti ahead only on single-hop questions.Exposing hypothetical-world records lowers Overall Main from 0.614 to 0.570; the paper contrasts writing’s use of hypothetical alternatives with QA’s benefit from established evidence.Source: §3.6 · pp. 8–9, Table 6 · p. 9, Figure 4 · p. 9

4 RELATED WORK

Prior work addresses persistent story state through writing workflows, narrative world models, and agent-memory designs. NarraWorld combines these concerns by modeling shared facts, private knowledge, cross-event development, and possibilities for sequential writing and situated action.

Long-form Narrative Generation and Storyworld Modeling

  • Long-form writing systems use outlines, recursive drafting, multi-agent planning, or explicit story-state models, while NarraWorld treats writing as accepted narrative acts that update memory across steps.This sequential framing makes memory responsible for consequences, perspective, and unresolved developments across requests.Source: §4 · p. 9

Computational Narratology and Situated Character Minds

  • Narratology’s distinctions between story events, discourse ordering, narratorial voice, and character consciousness motivate NarraWorld’s shared-world, private-belief, development, and possible-world views.The possible-world view is grounded in possible-world accounts of narrative.Source: §4 · p. 9

Persistent Memory for Situated Language-Model Agents

  • Persistent-agent memories include memory streams, working or external memory, personalized retention, and graph or structured systems, but the cited systems do not jointly model canonical facts, private knowledge, cross-event development, and possibilities.NarraWorld addresses this gap with a shared substrate and task-conditioned reconstruction for writing and situated action.Source: §4 · p. 9

5 CONCLUSION

NarraWorld frames narrative memory as reconstructing a dynamic, perspective-dependent world, reporting strong writing results and transfer while identifying construction and planning costs and uncertainty about exposing speculative branches. Its construction process preserves evidence and organizes records into connected views and traceable hierarchies.

5 CONCLUSION

  • Across ConStory-Bench, Tell Me A Story, and LSGC, NarraWorld achieves the strongest aggregate writing results, while the richer representation brings construction and planning costs and leaves open when speculative branches should be exposed.The conclusion also reports support for situated character action and general-purpose memory use, and says admissible memory views should depend on task.Source: §5 · p. 10

A NARRAWORLD MEMORY CONSTRUCTION

  • Each accepted source unit is parsed, integrated into the narrative structure, and made available through four memory views.The update uses trusted narrative text with source order and optional metadata, preserving exact source spans and segmenting at verified boundaries.Source: A NARRAWORLD MEMORY CONSTRUCTION · p. 15, §A.1 · p. 15

A.1 SOURCE UNITS AND CANONICAL RECORDS

  • The construction pipeline grounds entities, extracts events and propositions, models character access, and emits state, relation, and development changes; committed records retain provenance and prior validity intervals.A state change closes the prior interval and creates a new record rather than overwriting historical observations.Source: §A.1 · p. 15

A.2 INCREMENTAL UPDATE AND VIEW PROJECTION

  • The four views project the same committed records: shared facts, character-specific beliefs, cross-event developments, and conditional continuations for unresolved developments, retrieved under temporal, perspective, and branch constraints.A record may appear in multiple projections while retaining one provenance chain.Source: §A.2 · p. 15

A.3 HIERARCHICAL ORGANIZATION AND EVIDENCE CLOSURE

  • Scenes, plotlines, and plots are consolidated only when supported by content continuity or development signals, and atomic closure traces selected high-level nodes back to events, propositions, and exact source spans.When consolidation checks fail, lower-level records remain and no unsupported summary is committed.Source: §A.3 · pp. 15–16

A.4 WORKED MULTI-VIEW UPDATE

  • In the worked key example, Mara’s move updates the shared location and her belief, but not Leo’s; the event advances his search, while possible continuations remain hypothetical.The candidates are never written back as established world facts.Source: §A.4 · p. 16

A.5 SELECTED PROMPT INSTRUCTIONS

  • The selected instructions keep events distinct from propositions, ground character knowledge in evidence, preserve unresolved commitments, and connect scenes only through supported relations.They distinguish plans from execution and chronology from causation, and require scene connections to draw on at least two distinct scenes.Source: §A.5 · pp. 17–18

B WRITING BENCHMARK ADAPTATIONS

The benchmark adaptations cover fixed-prefix continuation, continual task-adaptive writing, and long-form generation, with shared inputs and withheld future text. TMAS is evaluated through rolling checkpoints that provide a source-faithful test of one continuation from a fixed history.

B WRITING BENCHMARK ADAPTATIONS

  • TMAS tests fixed-prefix continuation, ConStory-Bench tests continual writing across four task types, and LSGC extends continual generation to 14 chapters and at least 40,000 words per story.All conditions receive the same task inputs, and future text and evaluation annotations are withheld from memory construction and generation.Source: B WRITING BENCHMARK ADAPTATIONS · pp. 18–19

B.1 TMAS: SOURCE-FAITHFUL ROLLING CHECKPOINTS

  • TMAS evaluates 168 source-faithful checkpoints in a paired rolling protocol, initializing independent memory and ingesting each visible prefix before requesting one continuation.The evaluator-only human continuation is withheld; reported primary dimensions are Plot, Creativity, Development, and Language Use, with TMAS Quality as their arithmetic mean.Source: §B.1 · p. 19, Table 7 · p. 19

B.2 CONSTORY-BENCH: TASK-ADAPTIVE CONTINUAL WRITING

The section specifies continual-writing protocols across four ConStory task types and LSGC, then details how writing, role-playing, and long-term-memory outputs are scored. Protocols preserve task-specific constraints while allowing later writing units to retrieve from accepted history.

B.2 CONSTORY-BENCH: TASK-ADAPTIVE CONTINUAL WRITING

  • ConStory-Bench contains 200 stories across generation, continuation, expansion, and completion, assembled as 1,250 units under task-specific instructions and writing schedules.The first unit establishes each story’s initial state; 1,050 later units can retrieve from previously accepted writing, while future units and evaluator targets remain hidden.Source: §B.2 · pp. 19–20

B.3 LSGC: CHAPTER-CONTINUAL LONG-STORY GENERATION

  • LSGC extends five official starters into 14-chapter stories targeting 42,000 words, with memory retrieved once per chapter under a 12,000-token ceiling.The final-piece judgment uses a rolling summary rather than the complete 40,000-word story; completed-story instances, not chapters, are the primary statistical units.Source: §B.3 · pp. 20–21

C EVALUATION PROTOCOLS, METRICS, AND JUDGE PROMPTS

C.2 TELL ME A STORY: ADAPTED CONTINUATION QUALITY

  • TMAS compares continuations against a shared prefix using Plot, Creativity, Development, and Language Use, with the human continuation serving as a calibration anchor rather than the sole acceptable outcome.Prefix Continuity is auxiliary and does not enter the equal-weight TMAS Quality aggregate.Source: §C.2 · pp. 22–23

C.3 LSGC: FULL-STORY QUALITY UNDER THE LONGFORM WRITING BENCH RUBRIC

  • LSGC scores chapters and a final-story view against 16 criteria, direction-corrects negative criteria, and aggregates judgments within each completed story instance.Overall is the equal-weight mean of all 16 criteria, not the mean of the five criterion groups.Source: §C.3 · p. 23

C.4 RMTBENCH: TURN-LEVEL ROLE-PLAY SCORES

  • RMTBench retains its mixed scales: four dimensions use 1–5 scores, while character maintenance and user-preference adherence use binary success values when applicable.Inapplicable dimensions are excluded from their own denominator, and no cross-metric average is reported.Source: §C.4 · p. 24

C.5 LOCOMO: CATEGORY-AWARE DETERMINISTIC SCORING

  • LoCoMo’s main comparison uses category-aware scoring on 1,540 multi-hop, temporal, open-domain, and single-hop questions, excluding adversarial C5 questions.Overall Main is the question-weighted mean over the C1–C4 subset.Source: §C.5 · p. 24

C.1 CONSTORY-BENCH: CONSISTENCY ERROR DENSITY

The section defines consistency diagnostics for assembled ConStory stories and distinguishes them from literary-quality scores.

C.1 CONSTORY-BENCH: CONSISTENCY ERROR DENSITY

  • CED uses five categories covering 19 error subtypes and counts whether each subtype has at least one supported error in an assembled story.Multiple instances of the same subtype contribute one unit, and task-specific values use story-level means.Source: §C.1 · p. 22
  • Words per Penalized Error (WPE) is an auxiliary, length-normalized measure adapted from the benchmark’s intermediate score.Its formula uses story length and penalized-error count; both WPE and CED are consistency diagnostics, not literary-quality scores.Source: §C.1 · p. 22
  • CED and WPE are reported as consistency diagnostics rather than as measures of literary quality.Both measures depend on story length.Source: §C.1 · p. 22

D RETRIEVAL PLANNING ABLATIONS

The ablations compare query-centric retrieval with preview-conditioned dependency planning under matched memory and generation conditions. Preview-Guided Reconstruction leads on the reported LoCoMo aggregate and reduces ConStory consistency errors across task types.

D RETRIEVAL PLANNING ABLATIONS

  • Preview-Guided Reconstruction inspects an admissible memory preview to plan dependencies before formal retrieval, unlike query-centric controls that plan from the question alone.The LoCoMo comparison holds offline memory, final memory budget, and answer model fixed; the writing comparison likewise matches its scaffold and generation conditions.Source: D RETRIEVAL PLANNING ABLATIONS · pp. 24–25, §D.1 · p. 25, §D.2 · p. 26

D.1 LOCOMO: QUESTION-CENTRIC RETRIEVAL

  • 0.6144 Overall Main was highest for Preview-Guided Reconstruction, exceeding Deterministic Retrieval by 0.0181 and Retrieval w/o Dependency Planning by 0.0162.It also led multi-hop and single-hop accuracy; Table 13 reports the matched retrieval ablation.Source: Table 13 · p. 26, §D.1 · p. 26

D.2 CONSTORY-BENCH: WRITING-CONDITIONED RETRIEVAL

  • 0.130 overall CED reduction, with improvements in every task type, followed preview-guided reconstruction on ConStory-Bench.The largest absolute reduction was 0.198 on Completion, followed by 0.145 on Continuation; the comparison holds the writing scaffold and other conditions fixed.Source: §D.2 · p. 26, Table 14 · p. 26

E MEMORY ABLATIONS ACROSS TASKS

Memory ablations test the contribution of beliefs, possible worlds, and cross-event organization across writing and question answering. Their effects vary by task: cross-event organization has the largest reported LSGC quality impact, while excluding hypothetical-world records benefits LoCoMo QA.

E.1 TMAS: CONTINUATION QUALITY

  • The ablations compare removal of character beliefs, possible worlds, or cross-event organization while retaining the other memory components and matched evaluation settings.TMAS continuation quality complements consistency measurement because possible-world memory may affect creativity and development even without changing consistency error density.Source: E MEMORY ABLATIONS ACROSS TASKS · pp. 26–27, §E.1 · p. 27, Table 15 · p. 27

E.2 LSGC: LONG-FORM NARRATIVE QUALITY

  • Removing cross-event organization lowers LSGC Overall from 90.24 to 81.05, a 9.19-point decrease, the largest reported ablation drop.Removing beliefs lowers Overall to 86.04, while removing possible worlds yields 88.74; the full system is less stable than the H-removed condition across runs.Source: §E.2 · p. 27, Table 16 · p. 28, Table 17 · p. 28

E.3 LOCOMO: QA-TIME MEMORY-VIEW ACCESS

  • Excluding hypothetical-world records improves LoCoMo Overall Main by 4.4 points and improves every reported subset.This QA-time comparison isolates access to H; the writing ablations retain H to preserve unresolved alternatives during continuation.Source: §E.3 · p. 27

F.1 EVALUATION PROTOCOL

Human evaluation compared narrative consistency and quality across systems, alongside alignment between automatic and human scores. The rubric assessed five dimensions of long-form narrative quality.

F.1 EVALUATION PROTOCOL

  • Across 684 ConStory story–subtype pairs, adjudication yielded five positive labels; pooled agreement was 98.4% (κ = 0.55), while positive agreement was 56.0%.Four assessors evaluated anonymized texts; two independently read each complete story, and a blinded coordinator adjudicated disagreements.Source: §F.1 · p. 28

F.2 CONSTORY-BENCH: HUMAN CONSISTENCY ASSESSMENT

  • 0.095 was NarraWorld’s human CED, similar to Full Context’s 0.096 and below Graphiti’s 0.289 on 12 prompts.Human assessment identified one positive error-subtype label each for NarraWorld and Full Context, compared with three for Graphiti; the automatic judge flagged more errors than human readers.Source: §F.2 · p. 29, Table 19 · p. 29

F.3 LSGC: HUMAN QUALITY AND SCORE ALIGNMENT

  • 84.50 was NarraWorld’s highest LSGC Human Overall score, versus 80.50 for StructMem and 80.06 for Full Context; Full Context led World/Coherence and Prompt Adherence.NarraWorld led Character, Plot, and Style/Creativity. Automatic–human correlation was moderate overall (ρ = 0.55), with automatic scores averaging 5.71 points higher and lowest correlation for Style/Creativity (ρ = 0.38).Source: §F.3 · p. 29, Table 20 · p. 29, Table 21 · p. 30

F.4 HUMAN EVALUATION RUBRIC FOR LONG-FORM NARRATIVE

  • The human rubric grouped 16 criteria into Character, Plot, World/Coherence, Style/Creativity, and Prompt Adherence, each scored from 0 to 100.Assessors read complete stories and rated character depth and consistency, plot arc and pacing, storyworld coherence, prose quality and originality, and fidelity to the premise and constraints.Source: §F.4 · p. 30

G LIMITATIONS

NarraWorld’s structured modeling has a construction-cost trade-off, and its possible-world view is most suited to long-form writing. Selected case studies illustrate memory–narrative correspondences but do not estimate their frequency or isolate retrieval’s causal contribution.

G LIMITATIONS

  • NarraWorld averages about 200 successful modeling calls per story, within baseline workloads of roughly 20 to 800; its richer modeling is more expensive than a compact passage store.The authors describe a quality–cost trade-off and identify reducing extraction and planning overhead as future work.Source: G LIMITATIONS · p. 31
  • The possible-world view is designed for long-form writing, where unresolved alternatives can support continuation quality, but hypothetical content is not equally useful for factual QA.NarraWorld uses task-conditioned access to this view; a more adaptive gate is future work.Source: G LIMITATIONS · p. 31

H TRACING MEMORY INTO NARRATIVE: QUALITATIVE CASE STUDIES

  • The selected case studies illustrate individual memory–narrative correspondences, but do not estimate how often patterns occur or isolate retrieval’s causal contribution.The diagrams were not writer inputs; LSGC cases also examine retrieved text presented to the writer.Source: H TRACING MEMORY INTO NARRATIVE:… · p. 31

H.1 CONSTORY-BENCH: OBJECTS, PROMISES, AND RECONCILIATION

  • NarraWorld’s remembered promise appears as the father’s restraint while the narrator completes the fox drawing, though the sketchbook cover changes from green to blue.The case illustrates relational and thematic continuity while leaving object-state tracking as a separate challenge.Source: §H.1 · p. 32

H.2 LSGC: EVIDENCE, PERSPECTIVE, AND REASSESSING LOYALTY

  • NarraWorld reframes Percy’s concealment as coerced, but only partly distinguishes evidence from Ron’s interpretation; Graphiti’s successive accounts of Fred’s fate are also incompletely reconciled.The case develops the revelation through reassessment of prior testimony and relationships.Source: §H.2 · pp. 32–33

H.3 FURTHER EXAMPLES OF CONSTRAINTS AND CONSEQUENCES

  • Further cases show memory supporting rule-constrained rescue, responsibility in daily actions, revised clue interpretations, unresolved crisis at a chapter boundary, and a relationship after magical access ends.In the rescue case, treatment proceeds through a sealed channel while the indicator remains red; in the clue case, revised interpretations do not establish that Froshi’s sister is alive.Source: §H.3 · p. 33

H.4 QUALITATIVE LIMITATIONS

The qualitative cases show that retaining uncertainty and relevant constraints in memory does not ensure they are scrutinized or executed during generation. The scenes can preserve emotional or thematic continuity while leaving evidence attribution, investigation, or object-state consistency incomplete.

H.4 QUALITATIVE LIMITATIONS

  • The LSGC scene advances emotional conflict but only partly separates what evidence supports from Ron’s provisional interpretation.The case illustrates a gap between retaining uncertainty in memory and examining new evidence during generation.Source: §H.4 · pp. 33–34
  • In the ConStory case, thematic continuity coexists with an inconsistent sketchbook color, showing that remembered constraints still require execution.Returning relevant people and unresolved questions does not specify how a scene should test competing explanations; rapid dialogue revelation can leave little space for investigation.Source: §H.4 · p. 34
  • The authors identify targeted checks of consequential constraints as a direction for future work.Source: §H.4 · p. 34
  • Shared writing summaries should preserve claim attribution and uncertainty because stronger wording can override distinctions retained in memory.Source: §H.4 · p. 34

Related papers

Loading 2609.32401v1…