acceptodds
Under review as a conference paper at ICLR 2027

BOOKMARKS: Checkpoint-Accelerated Active Grounding for Role-playing

Abstract

Memory systems are critical for role-playing agents (RPAs) to maintain long-horizon consistency. However, existing RPA memory methods (e.g., profiling) mainly rely on incremental summarization, whose compression inevitably discards important details. To address this issue, we propose a search-based memory framework called BOOKMARKS, which actively initializes, maintains, and updates task-relevant bookmarks for the current task (e.g., character acting). A bookmark is structured as the answer to a question at a specific point in the storyline, serving as knowledge checkpoints. For each current task, BOOKMARKS proposes search questions and goes through storyline for answers. If a proposed question matches an existing bookmark, the search will start from that checkpoint instead of the storyline beginning, which accelerates searching. The synchronized bookmark will also represent a closer checkpoint for faster future search. Compared with incremental summarization, BOOKMARKS offers (1) active grounding for capturing task-specific details and (2) passive updating to avoid unnecessary computation. Our implementation supports concept, behavior, and state searches, each powered by an efficient synchronization method. BOOKMARKS significantly outperforms RPA memory baselines on 85 characters from 16 artifacts, demonstrating the effectiveness of search-based memory for RPAs.The code and data will be released for reproduction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Role-playing, Persona Modeling, Grounding System

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_8QQMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BOOKMARKS, a search-based memory framework for role-playing agents that combines active grounding with checkpoint reuse. Given the current scene, the framework proposes relevant information sections, reuses semantically equivalent bookmarks or initializes new ones, and synchronizes their contents by processing only the storyline suffix after each stored checkpoint. Concept, state, and behavioral bookmarks provide entity descriptions, state-transition traces, and scene–action examples, respectively. Experiments across six narrative artifacts show improvements over retrieval-based and profiling baselines in next-action prediction, while accuracy on 284 human-authored reasoning questions increases from 39.44% to 45.42% over the strongest baseline. Additional analyses report benefits across model families and evaluation dimensions, with checkpoint reuse maintaining comparable prediction quality while reducing repeated updates and output-token costs.

Critical strengths
  1. BOOKMARKS combines task-dependent information selection with checkpoint reuse in a coherent framework, addressing both information loss in fixed profiles and redundant processing during active grounding.
  2. Consistent gains across narrative artifacts and model families, supported by human evaluation and human-authored reasoning questions, provide meaningful evidence of practical effectiveness.
Critical weaknesses
  1. This paper employs GPT-4.1 as the Large Language Model (LLM) judge; however, relying on a single LLM judge may introduce bias. It is recommended to use a combination of models from different model families for evaluation.

  2. Regarding human evaluation, the paper should clearly specify details concerning NAP validation and open-domain evaluation, including participant recruitment, informed consent, compensation arrangements, and relevant ethical review or exemption status. Furthermore, the paper does not disclose the number of human evaluators, the number of annotations per sample, or inter-annotator agreement.

Questions for authors

When grounding a target character, can BOOKMARKS retrieve events in the preceding storyline that the character has not witnessed or learned about? If so, how are such events handled in the reported role-playing evaluations?

3 comments
AuthorsOpenReview ↗

Author's Response: Evaluation Reliability and Human Annotation & Character Knowledge and Grounding Scope

Thank you for your thoughtful review and constructive suggestions. We appreciate your recognition of the coherent design of BOOKMARKS and its consistent improvements across narrative artifacts and model families. We have added validation with a judge from another model family and clarified the human annotation protocol and character-specific grounding mechanism. We have uploaded a revised version and these additions and clarifications are marked in blue in the manuscript.


Evaluation Reliability and Human Annotation (W1, W2)

We additionally evaluated the same 200 cases per dimension (1000 in total) using Claude Sonnet 4, obtaining 92.1% agreement with GPT-4.1. This cross-family consistency complements the human validation reported in the paper and provides further evidence for the reliability of our evaluation.

Our human evaluation involved three judges familiar with the evaluated artifacts. For NAP validation, each case was assessed by one judge against the ground-truth reference, making the task relatively direct and limiting subjective interpretation. For open-domain evaluation, each sample was assessed by all three judges, with the final judgment determined by majority vote. The judges were blind to the source model in this evaluation. Participant recruitment, informed consent, compensation, and ethics review or exemption followed the applicable institutional procedures.

Appendix D now details both the cross-family judge comparison and the human annotation protocol, with a reference from the evaluation description in Section 4.


Character Knowledge and Grounding Scope (Q1)

BOOKMARKS can consult the complete preceding storyline when grounding a character. During state updates, the update rules instruct the model to skip events that the target character has neither witnessed nor learned about. Such events are therefore not automatically incorporated as the character’s observations. Incremental updating supports this distinction by following when information becomes available to the character and how it changes their state, preserving the character’s perspective as the story unfolds. We have made this character-specific update rule explicit in Section 3.4.


Thank you again for helping us make the evaluation and grounding protocol clearer. We hope you find the corresponding explanations in the revised manuscript helpful as you revisit these points. We fully respect the time and effort you have devoted to this review and will remain actively engaged throughout the entire discussion period. We would be glad to hear your further thoughts and work through any remaining questions with you.

Dear reviewer, we fully respect your recommendation and suggestions on our work. We hope our response has addressed your concerns and will stay active until the final day of the rebuttal period. Please do not hesitate to raise any further questions or share any additional thoughts.

Thank you for the detailed response and additional clarifications. My concerns have been largely addressed. Since my original rating already reflected my positive assessment of the paper’s contributions and empirical results, I will maintain my score of Weak Accept.

Reviewer_gwA4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes BOOKMARKS, a memory system for AI role-playing characters. It selects relevant topics from the story, maintains notes about them, and updates those notes only when needed. This helps characters use earlier events without repeatedly processing the whole story. The BOOKMARKS is evaluated across six narrative artifacts against non-active grounding baselines, showing improvement in next-action fidelity (emotion, intent, causality, position, and content) and mystery-game reasoning accuracy.

Critical strengths
  1. The idea is clear and the authors aim to address a practical problem.
  2. Results consistently improve over the tested baselines across several stories.
  3. Ablation experiments help explain which parts contribute.
Critical weaknesses
  1. The experiments compare BOOKMARKS against retrieval and profile-summary methods, but do not include established adaptive memory systems such as MemGPT, A-MEM, or Generative Agents’ reflection memory.

  2. The headline '8.1–32.6X' savings compares measured output-token usage with an estimated no-reuse counterfactual. Although Figure 4(b) includes baseline comparisons, output tokens do not capture total computational cost.

  3. Evidence for long-horizon deployment is limited. The main evaluation supplies the true preceding story at every prediction step, so generated mistakes do not enter subsequent memory. This evaluates memory-assisted prediction under a correct history, but does not establish robustness to compounding errors or behavioral drift. Best@1gains over the strongest baseline range from 1.22 to 3.13 points points, averaging 2.07 points across seven test sets.

Questions for authors
  1. How do the authors distinguish the contribution of the memory system from the model’s prior knowledge of familiar storylines, and assess whether that prior knowledge increases or reduces the observed gains?
  2. Which models perform bookmark proposal, matching, and updating? In the cross-model experiments, do these components also use the evaluated role-playing model?
  3. The no-reuse ablation suggests that checkpointing mainly improves efficiency while preserving accuracy. Can the authors isolate how much of the accuracy improvement comes from on-demand section selection vs. type-specific memory representations and update rules?
6 comments
AuthorsOpenReview ↗

Author's Response (1/4): Performance: Agentic Memory Comparisons and Component Contributions

Thank you for your constructive suggestions and the recognition of the practical problem we address, the consistent improvements across stories, and the value of the ablation analyses. Based on your suggestions, we have further strengthened the evaluation with agentic-memory comparisons and autoregressive rollouts, and supplemented the efficiency analysis with detailed resource measurements. We have uploaded a revised version and the corresponding additions and clarifications are marked in blue in the manuscript. We hope these updates will address your concerns on the current submission and look forward for further discussion from you.


Performance: Agentic Memory Comparisons and Component Contributions (W1; Q3)

Comparison with agentic memory. Our original baseline selection follows the methodological progression in Section 5.1 from ETA’s single profiles, through 3-way ETA’s selection among fixed sections, to BOOKMARKS’s dynamic section creation and type-specific updates. These directly connected comparisons examine the benefits of increasingly flexible memory organization. The mentioned agentic memory systems, including A-Mem, represent a parallel methodological line, making them valuable for broader architectural comparisons beyond this progression. We thus extend the evaluation to complementary agentic memory architectures. We select A-Mem [1] as a representative method, motivated by its stronger performance than MemGPT [2] in A-Mem’s published evaluation. We evaluate both A-Mem alone and BOOKMARKS + A-Mem. In the integrated method, retrieved A-Mem notes inform bookmark proposal and final grounding.

Table 1. Comparison with A-Mem and its integration with BOOKMARKS using Best@1 as the metric.

MethodHaruhiJOJOFMAAGOTPoPiPa BS0PoPiPa E321Manosaba NAPManosaba MCQ
Vanilla32.2134.1034.2141.3036.0544.5738.1035.21
A-Mem33.8636.2836.0342.8638.2747.8140.7237.68
BOOKMARKS36.3738.8737.8844.4139.8749.3543.9745.42
BOOKMARKS + A-Mem37.5238.2538.6645.0840.4949.7144.3746.83

A-Mem alone is less effective than BOOKMARKS, and their integration supports treating them as parallel, complementary approaches. A-Mem develops associations among memories, and BOOKMARKS determines the information needed for the current task, allowing the associative network to support grounding on demand. We have added a dedicated Section 5.3 for this comparison and expanded the agentic-memory discussion in Section 2.

Contributions to prediction quality. The submitted revision has already provided relevant evidence on on-demand section selection and type-specific memory representations and update rules through Table 2 and Figure 5. Section 5.1 now makes the connection between these components and the existing evidence more explicit. As observed, the no-reuse ablation yields comparable prediction quality, supporting checkpoint reuse primarily as an efficiency mechanism.

In Table 2 of the submission, 3-way ETA generally improves over ETA while retaining ETA-style updates, supporting the benefit of selecting and synchronizing task-relevant sections. BOOKMARKS extends this intermediate formulation with dynamic section creation and type-specific representations and update rules, achieving further improvements across all evaluated storylines and both tasks.

Figure 5a-b further examines the contributions of specialized memories. Removing behavioral bookmarks particularly affects Emotion and Intent, whereas removing state bookmarks has a larger effect on Causality and Content. On Manosaba MCQ, these removals reduce accuracy by 2.97 and 5.06 percentage points, respectively. These distinct effects support the complementary contributions of behavioral examples to character responses and state traces to the evolving circumstances behind actions.


References

[1] A-Mem: Agentic Memory for LLM Agents. NeurIPS 2025.

[2] MemGPT: Towards LLMs as Operating Systems. arXiv 2023.

AuthorsOpenReview ↗

Author's Response (2/4): Performance: Long-Horizon Evaluation

Performance: Long-Horizon Evaluation (W3)

We evaluate autoregressive rollouts of K consecutive actions, using K=8 and K=32 in non-overlapping windows. Each window starts from the true preceding storyline, with all methods following the same character-turn schedule. Generated actions then enter subsequent history and memory updates, allowing errors to accumulate within the rollout.

We evaluate each character’s trajectory as a whole against ground truth using the same five consistency criteria as NAP. The judge considers whether the character’s overall behavior matches the reference, allowing minor differences in individual actions. Our annotated storylines provide the complete character-level references required for this evaluation.

Table 2. Multi-step rollout consistency evaluated by Best@1

KMethodHaruhiJOJOFMAAGOTPoPiPa BS0PoPiPa E321Manosaba NAP
8Vanilla36.1936.3031.7537.4420.5819.7121.73
3-way ETA38.2238.5536.8440.2922.4323.9428.05
BOOKMARKS40.0240.6738.0843.0025.9527.5832.11
32Vanilla35.6030.9220.4535.2114.8220.306.77
3-way ETA38.8333.2925.2339.1119.8422.0217.58
BOOKMARKS41.4236.8730.8543.4826.1524.5522.71

Averaged across test sets, BOOKMARKS achieves larger relative improvements over 3-way ETA in both rollout settings than in single-step NAP, with the larger gain at K=32. These results support the value of grounding under accumulating generated history. We have added this evaluation as a dedicated Section 5.4, with the rollout and ground-truth comparison protocol detailed in Appendix G.

AuthorsOpenReview ↗

Author's Response (3/4): Efficiency and Computational Cost

Efficiency: Efficiency and Computational Cost (W2)

We report input and output tokens, actual API cost, and processing time alongside performance gains, comparing methods across different profile-length limits and maximum numbers of bookmarks in the proposal stage. We show the comparison between BOOKMARKS and ETA, as ETA is a baseline that does not apply either active grounding or passive update in its framework.

Resource statistics cover memory initialization, maintenance, and final prediction. Token usage is summed across requests, with tokens and time averaged per prediction. Actual API cost is reported per 1,000 predictions. Time denotes cumulative server-side processing time, excluding communication overhead.

Table 3. Performance gains and resource usage. (Following Figure 4b, performance gain denotes the Best@1 improvement over Vanilla in percentage points.)

Method / configurationPerformance gainInput / output tokensActual API cost, USD / 1,000 predictionsTime, s
ETA, max 128 words1.912,194.3 / 245.25.15.25
ETA, max 256 words2.363,680.0 / 470.89.08.19
ETA, max 512 words2.126,076.5 / 879.115.814.25
BOOKMARKS, max 3 bookmarks3.622,701.2 / 364.16.76.35
BOOKMARKS, max 5 bookmarks5.875,098.6 / 661.912.29.69
BOOKMARKS, max 10 bookmarks6.919,152.4 / 1,074.721.916.04

As shown in the Table, even the smallest tested BOOKMARKS configuration outperforms all ETA configurations. Compared with the ETA configurations using larger profiles, it also consumes fewer input and output tokens, incurs lower API cost, and requires less processing time. As consistently shown in Figure 4b, increasing ETA’s profile capacity yields diminishing gains and eventually a slight performance decline, whereas allowing more bookmarks in the proposal stage continues to improve predictions. This supports allocating resources to task-aware grounding over the complete preceding storyline as grounding needs change. For example, a gem introduced in the prologue may be mentioned repeatedly before its current location becomes necessary for a later prediction. Meeting that new grounding need requires tracing its state changes through the preceding storyline, which might begin from a very early point.

These measurements now appear in Appendix E and are discussed alongside Figure 4b in Section 5.5, making the relationship between resource use and prediction quality easier to assess.

Reviewer_xLW2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BOOKMARKS, a memory framework for role-playing agents. The main idea is to let the agent actively search for useful information from previous story events, instead of only using a fixed character profile. The method uses three types of memory: concept, state, and behavioral memory. It also uses checkpoints to avoid reading the same history again when updating memory. The authors test the method on six stories and show improvements in next-action prediction and reasoning tasks.

Critical strengths
  1. The paper is well written and easy to follow. The motivation is clear, and the framework is explained with good examples.

  2. The idea of combining active memory search with checkpoint-based updates is interesting. It can help reduce repeated computation in long conversations.

  3. The experiments cover different stories and models. The results show consistent improvements over the baselines, especially on reasoning questions.

Critical weaknesses
  1. Active retrieval, structured memory, and incremental memory updates have already been studied in previous work. The main contribution seems to be combining these ideas, rather than proposing a completely new method.

  2. Most baselines use simple retrieval or profile summarization. It would be helpful to compare with more recent memory systems, such as A-MEM or MOOM, or a simple agentic RAG method that actively searches the story.

  3. The method includes dynamic section creation, three memory types, and checkpoint reuse. However, it is still unclear how much each part contributes to the final performance.

  4. The paper shows large token savings compared with updating from scratch, but does not clearly report the total running time and cost compared with other memory methods. Creating a new bookmark may still require reading a long story.

Questions for authors
  1. What is the main technical difference between BOOKMARKS and existing methods that use active retrieval and dynamic memory updates?

  2. Have the authors tried a simpler baseline that first generates search queries and then retrieves relevant story events using BM25 or dense retrieval? How would it compare with BOOKMARKS?

  3. How much improvement comes from dynamic bookmark creation, and how much comes from using different memory types? Could the authors provide a more detailed ablation study?

  4. If a bookmark loses important information during an earlier update, can the system recover that information later? Does checkpoint reuse make this problem worse?

  5. Can the authors report the total input and output tokens, number of LLM calls, and actual inference time? This would help show whether BOOKMARKS is more efficient in practice.

5 comments
AuthorsOpenReview ↗

Author's Response (1/3): Technical Novelty and Baseline Comparisons

We are grateful for your positive and thoughtful feedback. We especially appreciate your recognition of the clear motivation, the value of active memory search with checkpoint reuse, and the consistent improvements across stories and models. Following your suggestions, we have broadened the empirical comparisons and clarified the resource requirements and contributions of individual components. We have uploaded a revised version and these additions and clarifications are marked in blue in the manuscript. We value your advice and look foward for further discussion during the rebuttal period.


Technical Novelty and Baseline Comparisons (W1, W2; Q1, Q2)

The technical contribution of BOOKMARKS is demand-driven memory construction and maintenance over the complete preceding storyline. At each prediction step, current grounding needs determine which information sections to create, reuse, and synchronize.

For related works, TimeChara [1] supports retrieval from earlier events, but evaluates independent character interviews. CFSM [2] studies sequential character-state updates, while BOOKMARKS further allows newly arising information needs to trigger memory construction from the complete preceding storyline. To our knowledge, BOOKMARKS is the first framework to introduce this form of demand-driven memory construction and maintenance over the complete preceding storyline in sequential character-action prediction. Each bookmark couples a section and its grounded content with an independent checkpoint. New needs can access the full history, while existing bookmarks process only the suffix after their checkpoints.

Comparison with agentic memory. Our original baselines follow Section 5.1’s progression from ETA’s single profiles, through 3-way ETA’s selection among fixed sections, to BOOKMARKS’s dynamic section creation and type-specific updates. These comparisons examine increasingly flexible memory organization. The mentioned agentic memory systems pursue a parallel methodological line. We extend the evaluation to complementary agentic memory architectures. We select A-Mem [3] for its stronger performance than MemGPT [4] in A-Mem’s published evaluation. We evaluate A-Mem alone and BOOKMARKS + A-Mem, where retrieved A-Mem notes inform bookmark proposal and final grounding.

Table 1. Comparison with A-Mem and its integration with BOOKMARKS using Best@1 as the metric.

MethodHaruhiJOJOFMAAGOTPoPiPa BS0PoPiPa E321Manosaba NAPManosaba MCQ
Vanilla32.2134.1034.2141.3036.0544.5738.1035.21
A-Mem33.8636.2836.0342.8638.2747.8140.7237.68
BOOKMARKS36.3738.8737.8844.4139.8749.3543.9745.42
BOOKMARKS + A-Mem37.5238.2538.6645.0840.4949.7144.3746.83

BOOKMARKS outperforms A-Mem alone, while their integration supports treating them as parallel, complementary approaches. A-Mem develops associations among memories, and BOOKMARKS activates this network according to current grounding needs. This comparison appears in Section 5.3, with the agentic-memory discussion expanded in Section 2.

Comparison with Simple Agentic Retrieval. We implement Q2E–BM25 [5], which generates keywords for BM25 retrieval, and RRR–Dense [6], which generates a query for dense retrieval. RICL directly uses the current scene for dense retrieval, providing a comparison without query generation.

Table 4. Comparison with simple agentic retrieval. Metrics and reporting follow Table 1.

MethodHaruhiJOJOFMAAGOTPoPiPa BS0PoPiPa E321Manosaba NAPManosaba MCQ
RICL32.8635.5235.2842.2837.5046.8239.8236.62
Q2E–BM2533.7536.0136.4342.8136.6547.4841.3837.68
RRR–Dense32.4434.6335.8541.6736.8146.5838.7536.97
BOOKMARKS36.3738.8737.8844.4139.8749.3543.9745.42

RICL outperforms RRR on mean NAP, favoring direct scene retrieval over query rewriting in this setting. BOOKMARKS improves further, supporting task-relevant memory construction and maintenance beyond retrieving individual events. Details appear in Appendix F, referenced from Section 5.3.

References

[1] TimeChara: Evaluating Point-in-Time Character Hallucination of Role-Playing Large Language Models. Findings of ACL 2024.

[2] Codified Finite-State Machines for Role-playing. ICLR 2026.

[3] A-Mem: Agentic Memory for LLM Agents. NeurIPS 2025.

[4] MemGPT: Towards LLMs as Operating Systems. arXiv 2023.

[5] Query Expansion by Prompting Large Language Models. Gen-IR Workshop at SIGIR 2023.

[6] Query Rewriting in Retrieval-Augmented Large Language Models. EMNLP 2023.

AuthorsOpenReview ↗

Author's Response (2/3): Efficiency and Computational Cost

Efficiency and Computational Cost (W4; Q5)

In complement to existing efficiency experiment in the submission, we further report input and output tokens, actual API cost, and processing time alongside performance gains, comparing methods across different profile-length limits and maximum numbers of bookmarks in the proposal stage.

Resource statistics cover memory initialization, maintenance, and final prediction. Token usage is summed across requests, with tokens and time averaged per prediction. Actual API cost is reported per 1,000 predictions. Time denotes cumulative server-side processing time, excluding communication overhead.

Table 3. Performance gains and resource usage. (Following Figure 4b, performance gain denotes the Best@1 improvement over Vanilla in percentage points.)

Method / configurationPerformance gainInput / output tokensActual API cost, USD / 1,000 predictionsTime, s
ETA, max 128 words1.912,194.3 / 245.25.15.25
ETA, max 256 words2.363,680.0 / 470.89.08.19
ETA, max 512 words2.126,076.5 / 879.115.814.25
BOOKMARKS, max 3 bookmarks3.622,701.2 / 364.16.76.35
BOOKMARKS, max 5 bookmarks5.875,098.6 / 661.912.29.69
BOOKMARKS, max 10 bookmarks6.919,152.4 / 1,074.721.916.04

Even the smallest tested BOOKMARKS configuration outperforms all ETA configurations in Table 3. Compared with the ETA configurations using larger profiles, it also consumes fewer input and output tokens, incurs lower API cost, and requires less processing time. As consistently shown in Figure 4b, increasing ETA’s profile capacity yields diminishing gains and eventually a slight performance decline, whereas allowing more bookmarks in the proposal stage continues to improve predictions. This supports allocating resources to task-aware grounding over the complete preceding storyline as grounding needs change. For example, a gem introduced in the prologue may be mentioned repeatedly before its current location becomes necessary for a later prediction. Meeting that new grounding need requires tracing its state changes through the preceding storyline.

These measurements now appear in Appendix E and are discussed alongside Figure 4b in Section 5.5, making the relationship between resource use and prediction quality easier to assess.

AuthorsOpenReview ↗

Author's Response (3/3): Contributions of Individual Components

Contributions of Individual Components (W3; Q3, Q4)

The submitted paper provides complementary evidence on the components’ respective roles through Table 2 and Figures 4-5. To make these connections easier to follow, Section 5.1 now explicitly links the methodological progression to the evidence on section selection, type-specific updates, and checkpoint reuse.

Memory organization and section selection are examined through the progression in Table 2. ETA maintains a single generic profile, while 3-way ETA introduces selection and synchronization among fixed sections using ETA-style updates. This intermediate formulation generally improves over ETA. BOOKMARKS further enables demand-driven section creation and type-specific updates, improving performance across all evaluated storylines and both tasks. This progression supports organizing memory around the information needed for the current grounding task.

The complementary roles of specialized memories are examined in Figure 5a–b. Removing behavioral bookmarks particularly affects Emotion and Intent, whereas removing state bookmarks has a larger effect on Causality and Content. On Manosaba MCQ, these removals reduce accuracy by 2.97 and 5.06 percentage points, respectively. These distinct effects support the contributions of behavioral examples to character responses and state traces to the evolving circumstances behind actions.

Checkpoint reuse primarily contributes to efficiency. Figure 5a-b shows comparable prediction quality when reuse is disabled. Together with the savings from avoiding repeated updates in Figure 4a, this supports checkpoint reuse as a mechanism for preserving grounding quality while reducing redundant computation.

Relevant details can be overlooked during memory updating, retrieval, or even direct interpretation of a storyline. BOOKMARKS retains access to the complete preceding storyline independently of bookmark checkpoints, providing an additional recovery path. When a new grounding need leads to a section without a matching bookmark, the system constructs a new bookmark from the beginning of the observed history. This allows later grounding needs to bring previously omitted evidence back into consideration, reducing dependence on what earlier updates happened to preserve. We have clarified this recovery path in the matching discussion in Section 3.4.


Thank you again for your careful reading and constructive suggestions. Your suggestions have helped make the evaluation more complete and the methodological distinctions clearer. We hope you find the fuller discussions in the revised manuscript helpful as you revisit these points. We fully respect the time and effort you have devoted to this review and will remain actively engaged throughout the entire discussion period. We would be glad to hear your further thoughts and work through any remaining questions with you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.