Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TAME: A Trustworthy Test-Time Evolution of Agent Memory with Systematic Benchmarking

Yu Cheng, Jiuan Zhou, Yongkang Hu, Yihang Chen, Huichi Zhou, Mingang Chen, Zhizhong Zhang, Kun Shao, Yuan Xie, Zhaoxia Yin

TL;DR

The paper tackles memory misevolution in test-time learning by introducing the Trust-Memevo benchmark to quantify multi-dimensional trustworthiness during benign task evolution. It proposes TAME, a dual-layer memory framework that externalizes safety into an evaluator while the executor drives capability, operating in a closed loop of memory filtering, draft generation, trustworthy refinement, and dual-track memory updating. Across Math, Science, and Tool-use domains, Trust-Memevo reveals pervasive trust degradation with naive memory evolution, while TAME achieves joint improvements in trustworthiness and task performance, with a parallel refinement variant (TAME-S) further boosting outcomes. The work highlights the importance of endogenous safety regulation in self-evolving agents and provides a formal foundation and empirical validation for trustworthy test-time memory evolution with potential for future multimodal extensions and more granular trust analysis.

Abstract

Test-time evolution of agent memory serves as a pivotal paradigm for achieving AGI by bolstering complex reasoning through experience accumulation. However, even during benign task evolution, agent safety alignment remains vulnerable-a phenomenon known as Agent Memory Misevolution. To evaluate this phenomenon, we construct the Trust-Memevo benchmark to assess multi-dimensional trustworthiness during benign task evolution, revealing an overall decline in trustworthiness across various task domains and evaluation settings. To address this issue, we propose TAME, a dual-memory evolutionary framework that separately evolves executor memory to improve task performance by distilling generalizable methodologies, and evaluator memory to refine assessments of both safety and task utility based on historical feedback. Through a closed loop of memory filtering, draft generation, trustworthy refinement, execution, and dual-track memory updating, TAME preserves trustworthiness without sacrificing utility. Experiments demonstrate that TAME mitigates misevolution, achieving a joint improvement in both trustworthiness and task performance.

TAME: A Trustworthy Test-Time Evolution of Agent Memory with Systematic Benchmarking

TL;DR

The paper tackles memory misevolution in test-time learning by introducing the Trust-Memevo benchmark to quantify multi-dimensional trustworthiness during benign task evolution. It proposes TAME, a dual-layer memory framework that externalizes safety into an evaluator while the executor drives capability, operating in a closed loop of memory filtering, draft generation, trustworthy refinement, and dual-track memory updating. Across Math, Science, and Tool-use domains, Trust-Memevo reveals pervasive trust degradation with naive memory evolution, while TAME achieves joint improvements in trustworthiness and task performance, with a parallel refinement variant (TAME-S) further boosting outcomes. The work highlights the importance of endogenous safety regulation in self-evolving agents and provides a formal foundation and empirical validation for trustworthy test-time memory evolution with potential for future multimodal extensions and more granular trust analysis.

Abstract

Test-time evolution of agent memory serves as a pivotal paradigm for achieving AGI by bolstering complex reasoning through experience accumulation. However, even during benign task evolution, agent safety alignment remains vulnerable-a phenomenon known as Agent Memory Misevolution. To evaluate this phenomenon, we construct the Trust-Memevo benchmark to assess multi-dimensional trustworthiness during benign task evolution, revealing an overall decline in trustworthiness across various task domains and evaluation settings. To address this issue, we propose TAME, a dual-memory evolutionary framework that separately evolves executor memory to improve task performance by distilling generalizable methodologies, and evaluator memory to refine assessments of both safety and task utility based on historical feedback. Through a closed loop of memory filtering, draft generation, trustworthy refinement, execution, and dual-track memory updating, TAME preserves trustworthiness without sacrificing utility. Experiments demonstrate that TAME mitigates misevolution, achieving a joint improvement in both trustworthiness and task performance.
Paper Structure (32 sections, 13 equations, 3 figures, 9 tables)

This paper contains 32 sections, 13 equations, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Test-Time Learning
  4. Agent Memory Evolution
  5. Problem Formulation and The Trust-Memevo Benchmark
  6. Problem Formulation and Attribution of Misevolution
  7. Trust-Memevo: A Multi-dimensional Benchmark for Misevolution
  8. Empirical Revelation
  9. Methodology
  10. System Notation and Definitions
  11. Retrieval and Memory Filtering
  12. Similarity-based Retrieval.
  13. Utility-Prioritized Draft Generation
  14. Trustworthy Refinement and Execution
  15. Holistic Plan Refinement.
  16. ...and 17 more sections

Figures (3)

  • Figure 1: Comparison between existing memory evolution methods and our trustworthy TAME framework.
  • Figure 2: Cross-domain evolution sets and multi-dimensional trustworthiness assessment.
  • Figure 3: The overall framework of TAME: a closed-loop dual-layer memory evolution system.