arXiv cs.AI· Xinyuan Song, Bowen Zhu, Hasibul Haque, Liang Zhao·· 2 天前AI 评分40
MEMONDEMAND：面向大规模企业数据的记忆管理系统
MEMONDEMAND: A Memory Management System for Large-Scale Enterprise Data
AI 导读
MEMONDEMAND 是一个面向大规模企业仓库的记忆管理系统，通过动态多级层级、每层双记忆（蒸馏路由与详细证据分离）和按需记忆提升三种机制协同工作。在 EnterpriseRAG-Bench 上，它从 10M token 到完整 618M token 集合的每个评测规模均超越已公开的 LB#1 最强结果，10M 时提升 12.23%，618M 时提升 4.66%。
正文 · 原文
Abstract:Enterprise repositories are large, heteroge- neous, and continuously updated, making re- trieval difficult when efficient access, source- faithful evidence, and cross-query adaptation must be supported together. Enterprise mem- ory extends retrieval beyond the model con- text, but existing systems do not jointly address collection-specific hierarchy construction, low- cost routing, detailed evidence loading, and workload-aware memory updates at this scale. We introduce MEMONDEMAND, short for On- Demand Memory, a memory management sys- tem with three coordinated mechanisms: a dy- namic multi-level hierarchy that determines the abstraction structure and depth for each col- lection, dual memory at every hierarchy level that separates distilled routing from detailed evidence, and on-demand memory promotion that updates node priority under a bounded active-state budget. On EnterpriseRAG-Bench, MEMONDEMAND outperforms the strongest published LB#1 result at every evaluated scale from 10M tokens through the complete 618M- token collection, with gains of 12.23% at 10M and 4.66% at 618M. Results on FinanceBench, HotpotQA, and FRAMES further show strong performance across financial, multi-hop, and fact-retrieval settings. Together, these results establish MEMONDEMAND as an accurate, ef- ficient, and scalable memory solution for very large enterprise repositories across data scales, domains, and evidence requirements. Our code is available at this https URL xfab-xinyuansong/MemOnDemand.git.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.22141 [cs.AI] |
| (or arXiv:2608.22141v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.22141 arXiv-issued DOI via DataCite |
Submission history
From: Bowen Zhu [view email]
[v1]
Sat, 22 Aug 2026 23:54:38 UTC (4,299 KB)
[v2]
Thu, 15 Oct 2026 05:14:48 UTC (2,408 KB)
来源：arXiv cs.AI · arxiv.org