Source-linked AI summary
BOOKMARKS: Efficient Active Storyline Memory for Role-playing
Letian Peng, Ziche Liu, Yiming Huang, Longfei Yun, Kun Zhou, Yupeng Hou, Jingbo Shang
TL;DR
Existing RPA memory methods can lose details because filtering or profile compression exposes only part of the preceding storyline to grounding. BOOKMARKS instead maintains task-relevant question-answer bookmarks, and it consistently outperforms baseline methods across Fandom and Bandori benchmarks while improving memory efficiency.
Problem
Existing RPA memory methods may expose only a partial preceding storyline to grounding because retrieval filtering or profile compression discards details.
Method
BOOKMARKS actively grounds useful memory targets and passively updates task-relevant question-answer bookmarks synchronized to the current storyline point.
Results
BOOKMARKS consistently outperforms Vanilla, RICL, and ETA across both Fandom and Bandori benchmarks.
Takeaways & Limitations
BOOKMARKS provides a practical search-based alternative that improves long-horizon consistency and memory efficiency for role-playing agents.
Takeaways & Limitations
The current framework mainly maintains storyline-level states, behaviors, and concepts rather than finer-grained recognition information about which character knows what.
Abstract
from arXiv · showhide
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 recurrent 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 pieces of 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. For each current task, BOOKMARKS selects reusable existing bookmarks or initializes new ones (at storyline beginning) with useful questions. These bookmarks are then synchronized to the current story point, with their answers updated accordingly, so they can be efficiently reused in future grounding rounds. Compared with recurrent summarization, BOOKMARKS offers (1) active grounding for capturing task-specific details and (2) passive updating to avoid unnecessary computation. In implementation, BOOKMARKS 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.
1 Introduction
BOOKMARKS addresses the loss of important storyline details in retrieval and profile-compression methods by actively selecting task-relevant search targets and passively synchronizing reusable bookmarks. It improves role-playing memory performance and search efficiency across diverse long-horizon evaluations.
- Existing retrieval and profiling methods expose only partial storylines to grounding because filtering or compression discards details.
- BOOKMARKS maintains bookmarks as question-answer pairs synchronized to the storyline position where each answer is valid.
- BOOKMARKS proposes queries, matches or initializes bookmarks, synchronizes them to the current story point, and grounds acting with nearby bookmark information.
- Active grounding searches for information useful to the current scene, while passive updating avoids updating information that may not be reused.
- BOOKMARKS outperforms incremental profiling and retrieval-based grounding, especially on long-horizon-dependent storylines.
- 90%, saving over 70% search calculation cost, while ablation shows match-and-derive performs comparably to calculating from the storyline beginning.
2 Background and Related Work
Related work frames RPA memory as a choice of stored representation and update strategy, while BOOKMARKS combines dynamic retrieval with memory-pool updates as storylines unfold. This extends the principle of retaining enough state for the next action-dependent question and retrieving more only when needed.
- RPA research addresses personalized, consistent in-character actions through evaluation, training and inference, and dynamic memory systems.
- RPA evaluation includes holistic judgments and per-action scoring, with Fandom and Bandori supporting strict single-step scene-action comparisons.
- Training-time methods adapt model parameters through character experiences, synthetic dialogue, LoRA, personality-aware data, and reinforcement learning.
- RPA memory methods differ in what they store and how they update it, including static compressed profiles and dynamic retrieval from a fixed memory pool.
- BOOKMARKS combines dynamic-profile retrieval with memory-pool updates as the storyline unfolds, described as rolling self-augmentation not implemented by prior RPA memory methods.
3 Our BOOKMARKS Framework
BOOKMARKS builds task-relevant grounding from reusable bookmarks rather than compressing the entire storyline into one profile. It proposes useful questions, matches or initializes bookmarks, synchronizes selected items over newly unseen actions, and uses their answers to ground prediction.
- Bookmark Data Structure: BOOKMARKS maintains reusable bookmarks that track task-relevant questions and their answers at specific storyline synchronization points.Each bookmark stores a question, current answer, search type, synchronization point, and optional auxiliary memory.
- Active Grounding: At each prediction step, BOOKMARKS proposes task-conditioned questions instead of applying a fixed memory template across scenes.The proposal stage selects information worth tracking for the current action while supporting later maintenance and reuse.
- Passive Updating: Selected bookmarks are passively synchronized only when relevant, processing the unseen storyline suffix from each bookmark’s stored position to the current point.This avoids continuously updating irrelevant memory items and makes synchronization incremental.
- Matching: Proposed queries are matched by search type and may reuse an existing bookmark, derive a new one, or create a fresh bookmark.Reuse preserves an existing memory slot, derivation initializes a related bookmark from a synchronized answer, and fresh bookmarks begin without an existing suitable candidate.
- Synchronization: BOOKMARKS supports concept, state, and behavioral searches with type-specific synchronization mechanisms.State synchronization tracks evolving facts, behavioral synchronization accumulates evidence for behavior patterns, and concept synchronization gathers contextual occurrences.
4 Benchmark
The benchmark evaluates role-playing agents on sequentialized storylines by comparing predicted character actions with ground truth. It uses examples from two benchmark collections and a strict exact-match criterion over 15.2K test instances.
- Instance Examples: Table 1 presents benchmark instance examples, while real benchmark scenes contain 10 preceding actions.The examples illustrate the preceding-action context used for role-playing prediction.
- Datasets: Two benchmark collections cover 85 characters across 16 artifacts, including 45 characters from 8 artifacts and 40 characters from 8 Bandori stories.The collections contain 20,778 and 7,866 actions from benchmarked characters, respectively.
- Criterion: 15.2K test instances evaluate predictions made from preceding storyline actions.Each character’s storyline is split into two halves, using the first for information collection and the second for evaluation.
- Criterion: Exact match judges whether the key move of each predicted action matches the reference action.GPT-4.1 serves as the judge, with manual validation finding agreement in 483 of 500 cases, or 96.6%.
5 Experiment
BOOKMARKS is evaluated against retrieval, profiling, and no-grounding baselines across benchmarked role-playing tasks, with additional efficiency, ablation, live-storyline, and multi-action analyses. The experiments consistently support search-based storyline memory, especially through synchronized bookmarks and active, task-relevant grounding.
- 5.2 Main Results: BOOKMARKS consistently outperforms Vanilla, RICL, and ETA across Fandom and Bandori benchmarks.The comparison uses Key Move Exact Match Rate on the two benchmarks.
- 5.3 Reusing Hit Rate: BOOKMARKS improves efficiency by reusing or deriving bookmarks instead of recomputing updates from the storyline beginning.Reuse handles identical queries, while derivation initializes a new bookmark from a useful synchronized basis; the hit rate fluctuates with narrative changes.
- 5.4 Ablation Study: Removing near bookmarks reduces performance, while removing derivation or reuse mainly weakens BOOKMARKS’s efficiency advantage.The ablation is conducted on the PoPiPa dataset with five characters.
- 5.4 Ablation Study: Replacing behavioral synchronization with incremental behavior updating hurts performance, supporting verified action evidence for maintaining behavior bookmarks.The ablation also attributes grounding quality to active proposal, passive updating, and type-specific synchronization.
- 5.5 Live Evaluation: BOOKMARKS achieves the best overall performance on a post-cutoff storyline and outperforms Vanilla, RICL, and ETA across most characters.The live evaluation uses BanG Dream! Girls Band Party! Event 321, released on February 8, 2026.
- 5.6 Case Study: Multi-action Generation: In a multi-action case study, BOOKMARKS preserves local narrative focus and produces a continuation closer to the reference than simple ETA profile grounding.Its reusable anchors cover location, preceding events, group activity, speaking style, and character reactions.
6 Conclusion and Future Work
The paper concludes by presenting BOOKMARKS as a search-based memory framework for long-horizon role-playing agents and outlining future extensions.
- BOOKMARKS maintains task-relevant bookmarks as synchronized question-answer pairs along the storyline.
- BOOKMARKS is positioned as an alternative to retrieval-only grounding and incremental profile compression.
- Future work includes recognition management, self-refinement integration, and customized update policies for query types and narrative structures.
Limitations
The framework establishes general search-based memory for long-horizon role-playing while leaving several extensions for future exploration.
- BOOKMARKS currently maintains storyline-level states, behaviors, and concepts rather than finer-grained recognition information.
- Future work may track which character knows which information at each story point.
- Future directions include integrating bookmark evidence with self-refinement and developing more customized update policies.
A Statistics
The experiments use benchmark statistics summarized in Table 5.
- Table 5 presents the benchmark statistics used in the experiments.
B Character & Artifact Background Information
The appendix provides concise background information for the experimental artifacts, bands, stories, and characters through Tables 6–9.
- Tables 6–9 provide concise descriptions of artifacts and character-related background information used in the experiments.
- Table 6 summarizes stories from the Fine-grained Fandom Benchmark.
- Tables 7–9 summarize bands and character background information for the Bandori and Fandom benchmarks.