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

cs.CLSep 26, 2026

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

Authors: Qiuyu Tian, Xiaowen Gu, Hang Su, Jianghan Chao, Haojie Yin, Fan Guo, Xin Zhang, Jinjing Shen,

Organizations: Southeast University, Nanjing, China · Beijing Zhongguancun Academy, Beijing, China · School of Animation and Digital Arts, Communication University of China · Jiangsu Second Normal University ·

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.

Figures & tables

Figure 1: Illustrative comparison of conventional memory and NarraWorld. Leo leaves the key in the desk drawer, but Mara moves it into a hollow book while he is away. Conventional memory supplies the updated location, causing Leo to walk directly to the book. NarraWorld preserves both the shared fact that the key has moved and Leo’s private belief that it remains in the drawer, so the continuation is: he checks the drawer, discovers it empty, and asks Mara about the key.

Figure 2: Overview of the NarraWorld continual world-creation cycle. Narrative history is formed into shared-world and character-belief memory, then reconstructed for the next writing act or situated agent action. The resulting continuation becomes the next source unit in the evolving narrative.

Method	CED $\downarrow$	WPE $\uparrow$	Generation	Continuation	Expansion	Completion	Mem. tokens
NarraWorld	0.185	10,520.4	0.120	0.235	0.098	0.287	3,231
Graphiti	0.385	9,480.5	0.290	0.415	0.310	0.525	5,971
StructMem	0.450	9,215.3	0.355	0.495	0.380	0.570	3,891
Mem0	0.955	7,845.2	0.812	0.985	0.920	1.103	2,179
AriGraph	1.125	7,250.6	0.945	1.180	1.055	1.320	3,050
Hindsight	1.215	6,985.4	1.050	1.265	1.140	1.405	3,280

Table 1: ConStory-Bench consistency results. CED and task-specific errors are lower-is-better; WPE is an auxiliary higher-is-better word-per-error measure.

Table 4

Method	EE $\uparrow$	EC $\uparrow$	PA $\uparrow$	CU $\uparrow$	CM $\uparrow$	UPA $\uparrow$	Mem. tokens
NarraWorld	4.22	4.06	4.30	4.57	87.15%	68.85%	704
Mem0	3.98	3.56	3.74	4.25	85.00%	67.31%	344
Graphiti	4.02	3.61	3.82	4.41	86.15%	53.85%	1,215
Hindsight	4.01	3.52	3.78	4.21	84.60%	65.40%	654
AriGraph	3.95	3.60	3.71	4.28	85.35%	68.10%	429
Emotional RAG	3.99	3.55	3.73	4.22	82.10%	67.30%	600

Table 5: RMTBench multi-turn role-playing results (

\uparrow

). EE/EC/PA/CU use 1–5 scores; CM/UPA are percentage success rates. Scores are averaged over applicable turns.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Method	TMAS Quality	Prefix Continuity $\uparrow$	Words Generated	Unique
NarraWorld	83.01	87.29	905.6	0.3372
Mem0	81.79	86.98	931.6	0.3582
Graphiti	82.04	85.83	914.7	0.3599
Hindsight	77.87	80.48	920.9	0.3573
AriGraph	76.78	77.88	901.9	0.3595
StructMem	81.80	86.83	936.8	0.3414

Appendix

Table 7: TMAS aggregate and auxiliary diagnostics. TMAS Quality is the arithmetic mean of the four reported dimensions; Prefix Continuity, word counts, and Unique are descriptive measures.

Criterion	Mem0	AriGraph	Hindsight	Graphiti	StructMem	Narra World	Scaffold Only	Full Context
Believable Character Actions	89.07	78.13	80.84	88.69	90.02	91.04	70.96	86.80
Nuanced Characters	87.62	72.27	75.78	88.51	88.22	91.22	69.11	91.53
Pacing	70.93	71.02	72.98	81.29	85.20	87.71	69.09	81.76
World Building	86.00	74.80	79.06	88.29	87.29	87.49	66.56	92.62
Compelling Plot	80.22	75.62	79.91	88.76	85.60	87.25	67.18	88.42
Emotionally Engaging	87.58	80.53	82.38	89.00	89.84	94.60	71.02	91.09

Appendix

Table 8: LSGC final-rubric scores by criterion. Values are direction-corrected macro means over five starters and three runs on a 0–100 scale (

\uparrow

); negative criteria are inverted.

Starter	Mem0	AriGraph	Hindsight	Graphiti	StructMem	NarraWorld	Scaffold Only	Full Context
01	79.33	73.32	74.17	85.78	84.83	91.95	74.55	87.26
02	65.67	73.07	75.60	85.36	89.35	86.78	61.74	86.35
03	85.47	71.56	73.36	54.53	86.61	91.05	65.81	88.03
04	82.93	74.26	77.20	84.69	87.52	90.38	84.13	87.48
05	87.51	72.97	74.24	85.11	86.27	91.01	61.85	76.04
Macro mean	80.18	73.03	74.91	79.10	86.92	90.24	69.61	85.03

Appendix

Table 9: LSGC Overall scores by starter, averaged over three runs. Each value equally weights the 16 direction-corrected final criteria.

Method	Overall mean $\uparrow$	Run range	Run SD
NarraWorld	90.24	2.59	1.35
Mem0	80.18	5.38	2.69
Graphiti	79.10	3.03	1.51
Hindsight	74.91	4.06	2.10
AriGraph	73.03	3.57	1.90
StructMem	86.92	3.34	1.69

Appendix

Table 10: Run-to-run variability of LSGC Overall scores. Each run is first aggregated over five starters and the 16 criteria; range and SD are computed over the three independent runs. SD is the sample standard deviation.

Dimension	Subcriteria
Plot	Connected structure; plot progression; logical coherence; intentional turns.
Creativity	Engaging specificity; non-generic ideas; purposeful use of tropes; original contribution.
Development	Character context; setting context; believable detail; motivated development.
Language Use	Sentence variety; lexical richness; literary effect; repetition control; appropriateness of voice.
Prefix Continuity	Continuity of events, characters, perspective and voice, and facts.

Appendix

Table 11: TMAS evaluation subcriteria, each scored from 0 to 4.

Method	Budget	Overall Main	Multi-hop	Single-hop	Temporal	Open-domain
Mem0	2k	0.4736	0.3718	0.5241	0.4919	0.2691
	4k	0.4982	0.4047	0.5392	0.5372	0.2834
	8k	0.5010	0.4100	0.5410	0.5470	0.2700
	12k	0.4683	0.3984	0.5125	0.4758	0.2620
	16k	0.4742	0.4047	0.5125	0.4963	0.2682
StructMem	2k	0.4962	0.3565	0.5275	0.5827	0.3435

Appendix

Table 12: LoCoMo budget sensitivity for NarraWorld, StructMem, and Mem0. All scores are higher-is-better.

Figure 5: Conceptual comparison of retrieval paradigms. Query-centric methods interpret the request and match the resulting query against stored memory; both deterministic parsing and question-only LLM planning instantiate this left-hand path. NarraWorld instead conditions narrative planning on an admissible memory preview before expanding and searching for supporting evidence. The diagram summarizes the two information flows.

Retrieval setting	Overall Main	Multi-hop	Single-hop	Temporal	Open-domain
Deterministic Retrieval	0.5963	0.4479	0.6467	0.6513	0.4069
Retrieval w/o Dependency Planning	0.5982	0.4487	0.6488	0.6449	0.4383
Preview-Guided Reconstruction (NarraWorld)	0.6144	0.4580	0.6790	0.6490	0.3920

Appendix

Table 13: LoCoMo retrieval ablation with shared offline

W/B/D

memory, an 8k final memory budget, and the same answer model. All metrics are higher-is-better.

Retrieval setting	CED $\downarrow$	Generation	Continuation	Expansion	Completion
Retrieval w/o Dependency Planning	0.315	0.210	0.380	0.185	0.485
Preview-Guided Reconstruction (NarraWorld)	0.185	0.120	0.235	0.098	0.287

Appendix

Table 14: ConStory-Bench retrieval ablation under matched writing scaffolds, offline memory, memory budgets, and generation settings. All columns report CED (

\downarrow

Setting	Plot	Creativity	Development	Language	TMAS Quality	Prefix Continuity
Full NarraWorld	78.92	84.90	85.60	82.62	83.01	87.29
Without character beliefs ( $B$ )	76.54	82.15	81.33	81.58	80.40	83.21
Without possible worlds ( $H$ )	77.20	78.85	80.15	82.10	79.57	85.65
Without cross-event organization	74.35	81.12	78.90	80.45	78.70	79.42

Appendix

Table 15: TMAS ablations. All quality dimensions are higher-is-better.

Criterion	Full NarraWorld	Without $B$	Without $H$	Without cross-event
Believable Character Actions	91.04	83.33	92.02	82.51
Nuanced Characters	91.22	81.93	90.09	81.16
Pacing	87.71	84.22	85.36	72.44
World Building	87.49	85.96	88.44	74.84
Compelling Plot	87.25	82.49	84.33	76.82
Emotionally Engaging	94.60	87.36	88.27	82.49

Appendix

Table 16: LSGC ablations by criterion. Scores are direction-corrected macro means over 15 samples (5 starters

\times

3 runs) on a 0–100 scale (

\uparrow

Setting	Overall mean $\uparrow$	Run range	Run SD
Full NarraWorld	90.24	2.59	1.35
Without $B$	86.04	3.26	1.68
Without $H$	88.74	1.53	0.80
Without cross-event	81.05	4.02	2.09

Appendix

Table 17: Run-to-run variability for LSGC ablations. Overall means are direction-corrected scores; range and sample SD are computed over the three independent runs after aggregating each run over five starters and 16 criteria.

Setting	Overall Main	Multi-hop	Single-hop	Temporal	Open-domain
NarraWorld (QA view: $H$ excluded)	0.614	0.458	0.679	0.649	0.392
NarraWorld (QA view: $H$ included)	0.570	0.398	0.655	0.559	0.366
Without character beliefs ( $B$ )	0.586	0.415	0.650	0.635	0.370
Without cross-event organization	0.537	0.388	0.618	0.512	0.341

Appendix

Table 18: LoCoMo memory-view and component ablations at the shared 8k budget. The default QA view excludes hypothetical-world records (

H

); the exposed-

H

condition includes them. Overall Main aggregates the four reported question subsets (

\uparrow

Method	Human CED $\downarrow$	Auto CED $\downarrow$	$\Delta$	TP	FP	FN	Precision	Recall	F1
NarraWorld	0.095	0.185	+0.090	1	1	0	50.0%	100.0%	66.7%
Graphiti	0.289	0.385	+0.096	2	2	1	50.0%	66.7%	57.1%
Full Context	0.096	0.200	+0.104	1	1	0	50.0%	100.0%	66.7%

Appendix

Table 19: ConStory-Bench human evaluation and judge alignment on 12 prompts (12 stories per method).

\Delta

is Auto CED minus Human CED. Detection metrics use adjudicated error-subtype presence as the reference.

Method	Character	Plot	World/Coh.	Style/Creat.	Prompt	Human Overall	Auto Overall
NarraWorld	86.00	84.00	86.00	82.00	94.00	84.50	90.24
StructMem	84.00	81.00	82.00	76.00	92.00	80.50	86.92
Full Context	83.00	82.00	88.00	72.00	96.00	80.06	85.03

Appendix

Table 20: LSGC human scores across all five starters on a 0–100 scale (

\uparrow

). Auto Overall is shown alongside the human ratings for comparison.

Dimension	Spearman $\rho$	MAE	Bias	Human–Human MAE
Character	0.52	5.40	+4.89	6.20
Plot	0.48	6.80	+6.42	7.50
World/Coherence	0.61	5.10	+4.67	5.50
Style/Creativity	0.38	7.60	+7.17	9.00
Prompt	0.68	3.20	+2.33	4.20
Overall	0.55	6.20	+5.71	6.48

Appendix

Table 21: LSGC human–judge alignment across 15 stories. Bias is the mean automatic minus human score. Both MAE columns use the 0–100 scale.

Figure 6: Memory and continuation in ConStory-Bench prompt 128, unit 5. Solid colored arrows connect recurring details; dashed arrows connect related developments.

Figure 7: Memory and revelation in LSGC starter-01, Chapter 10. Solid arrows connect recurring content; dashed arrows connect related developments and reinterpretations.

Memory-Augmented Language Model Agents

Memory-augmented language model agents use persistent facts, interaction histories, and skills to support personalization and long-horizon tasks.

See all 428 papers

Narrative World Model: Narratology-Grounded Writer Memory for Long-Form Fiction
Jul 6, 2026Mohammad Saifullah, Thomas Kornmaier, Taaha Kazi +3Multi-Hop QA Question Answering
PocketFM
Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory
Jun 9, 2026Suozhao Ji, Baodong Wu, Zehao Wang +8LLM Agent Memory Persistent Memory for Language Models
1Infinigence AI · 2Tsinghua University · 3Shanghai Jiaotong University
From Personas to Plot: Character-Grounded Multi-Agent Story Generation for Long-Form Narratives
Jul 1, 2026Aayush Aluru, Chloe Ho, Muhammad Hammouri +5Grounded Text Generation Multi-Agent LLMs
Pocket FM · Princeton University · University of Michigan +2

Jul 6, 2026J/K move · Enter open · S save

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

Mohammad Saifullah, Thomas Kornmaier, Taaha Kazi, Vasu Sharma+2

PocketFM

Multi-Hop QA Question Answering Knowledge Representation Temporal KGs Temporal QA

Open paper AskPDF arXiv

Long-form fiction writers need memory that answers multi-hop questions about evolving story state: who knows a secret and when they learned it, whether an event preceded 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 turn on, so they surface the wrong evidence or none at all. We introduce the Narrative World Model (NWM), a writer-memory system that pairs a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. To measure memory rather than the answerer, we read every system through a single held-constant Opus 4.8 reader over only that system's chapter-safe evidence, on a reproducible public corpus and a validated multi-hop benchmark, and we compare against the strongest existing temporal-knowledge-graph agent-memory framework, Graphiti/Zep (Rasmussen et al., 2025). NWM substantially and significantly outperforms this baseline on multi-hop narratological QA across both corpora, and far exceeds GraphRAG and flat retrieval. The advantage is representational rather than an artifact of extraction: it survives rebuilding the baseline with NWM's own extractor, and traces to its narratology-grounded structure and query-conditioned retrieval, not to graph size or extractor quality.

Abstract

Memory-Augmented Language Model Agents

Explore similar work

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

Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory

From Personas to Plot: Character-Grounded Multi-Agent Story Generation for Long-Form Narratives