BOOKMARKS: Efficient Active Storyline Memory for Role-playing
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 6 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Letian Peng et al. |
| Year | 2026 |
| HF Upvotes | 6 |
| arXiv | 2605.14169 |
| Download | |
| Code | https://github.com/KomeijiForce/BOOKMARKS_Koishiday_2026 |
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 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.
Engineering Breakdown
Plain English
BOOKMARKS proposes a search-based memory system for role-playing agents that maintains long-horizon consistency by storing structured question-answer pairs (bookmarks) tied to storyline points, rather than relying on lossy summarization. The system actively selects which existing bookmarks to reuse for a current task or creates new ones, then synchronizes answers as the story progresses—avoiding the information loss that occurs when compressing memory through summarization.
Key Engineering Insight
Instead of compressing memory through summarization (which loses details), BOOKMARKS uses a retrieval-based approach with question-answer pairs anchored to story points. This shifts memory management from lossy sequential compression to explicit, addressable facts that can be selectively retrieved and updated—making memory bottlenecks a retrieval problem, not a compression problem.
Why It Matters for Engineers
Production dialogue and game agents today struggle with consistency over long conversations because they either lose context through compression or hit token limits through verbatim memory. BOOKMARKS provides a concrete pattern for structured memory that scales: question-answer pairs are compact, queryable, and can be selectively synchronized, making it tractable to maintain character consistency across hour-long conversations without exponential context growth.
Research Context
Prior RPA memory methods relied on periodic summarization to keep context bounded, which inevitably discards details needed for consistent characterization. BOOKMARKS advances this by introducing a structured, search-indexed memory format that trades sequential summarization for active retrieval and targeted updates. This enables a new class of agents that can handle multi-hour storylines while maintaining fine-grained consistency—shifting from passive compression to active memory management.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::