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

Researchers introduce TAME, a trust-aware memory evolution framework that addresses the vulnerability of AI agents to safety misalignment during test-time learning. The system uses paired Executor and Evaluator components to selectively reinforce and reuse agent memories, demonstrating 14.6 percentage point accuracy improvements on mathematical benchmarks while maintaining trustworthiness.

The research tackles a critical challenge in developing advanced AI systems: maintaining safety alignment as agents learn and evolve through experience without parameter updates. Traditional approaches to agent memory assume all accumulated experiences are equally valuable, but TAME recognizes that uncurated memory accumulation can degrade safety properties—a phenomenon termed Agent Memory Misevolution. This distinction matters because it highlights a fundamental tension in AGI development between capability advancement and safety preservation.

The architectural innovation centers on a collaborative governance model where the Executor handles practical task execution while the Evaluator provides quality assurance through trust feedback mechanisms. This separation of concerns mirrors human organizational structures and reflects growing recognition that capability and safety are interdependent rather than competing objectives. The Trust-Memevo benchmark itself represents a contribution by establishing systematic evaluation criteria for agent trustworthiness during evolution, addressing a measurement gap in the field.

For the broader AI development community, these findings suggest that memory curation mechanisms are not peripheral optimizations but core components of safe AGI systems. The demonstrated performance gains—particularly the 14.6 percentage point improvement on AIME benchmarks—indicate that safety-aware designs need not sacrifice capability. This counterintuitive result challenges assumptions that safety constraints inherently limit performance, potentially influencing how subsequent research priorities are balanced.

The research points toward memory-augmented architectures as central to next-generation AI systems, with implications for how developers design retrieval mechanisms and feedback loops in large language models and reasoning systems.

• →TAME framework maintains AI agent safety during test-time learning by introducing trust-aware memory governance between Executor and Evaluator components.
• →Agent Memory Misevolution occurs when uncurated experience accumulation degrades safety alignment despite task performance improvements.
• →The Trust-Memevo benchmark establishes systematic evaluation criteria for trustworthiness during agent memory evolution.
• →TAME achieves 14.6 percentage point accuracy improvement on GPT-5.2 AIME while preserving competitive trustworthiness scores.
• →Safety-aware memory curation appears to enhance rather than constrain AI reasoning performance.