language:
- en
pretty_name: AutoMemoryBench
tags:
- benchmark
- agent-memory
- memory-systems
- llm-agents
- evaluation
AutoMemoryBench
State-Contract Evaluation for Auditable Agent Memory
AutoMemoryBench evaluates whether an agent uses the right memory, and only the admissible memory, under a query-time state contract. Each contract partitions memory into required, admissible, and prohibited sets, with prohibited memories typed as superseded, deleted, restricted, cross-namespace, or stale-tool.
Paper
Authors
Jiajun Wu, Jian Yang, Chen Li, Linzheng Chai, Ge Gao, Ensheng Shi, and Yuchi Ma.
Jiajun Wu and Jian Yang are affiliated with Beihang University. Ensheng Shi and Yuchi Ma are affiliated with Huawei Cloud Computing Technologies Co., Ltd.
Abstract
A deployed agent fails not only when it forgets, but when it uses memory that is stale, deleted, unauthorized, cross-namespace, or unsupported by current tool state. AutoMemoryBench introduces query-time memory admissibility as an executable evaluation contract. Its primary metric, StrictCore, counts a query only when the task is solved, required evidence is used, and no prohibited memory influences the output. The release spans eight domains with 7,200 case variants and 151,200 probes, paired with minimal single-axis interventions.
Citation
@misc{automemorybench,
title = {AutoMemoryBench: State-Contract Evaluation for Auditable Agent Memory},
author = {Wu, Jiajun and Yang, Jian and Li, Chen and Chai, Linzheng and
Gao, Ge and Shi, Ensheng and Ma, Yuchi},
year = {2027},
url = {https://github.com/wuyuVerse/AutoMemoryBench}
}
The formal Hugging Face Paper page and arXiv link will be added after the paper receives an arXiv identifier.
License
The code and benchmark data are released under Apache-2.0. The paper license will follow the license selected for the arXiv submission.