PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
2026-08-04 • Computation and Language
Computation and Language
AI summaryⓘ
The authors studied whether AI agents that remember past experiences actually get better over time. They created PAST-Bench, a test that compares agents with memory turned on versus off across many tasks. They found that most agents show some improvement, but how they improve varies a lot. Using these insights, the authors improved an agent called Hermes+ to boost gains from memory, especially when updating outdated information. Their work offers a way to measure and understand how AI agents can learn and improve by using what they remember.
recursive self-improvementpersonal AI agentsexperience retentionPAST-BenchHermes+task memoryprocedural reuseinformation gatheringagent evaluationmachine learning benchmarks
Authors
Shuhan Xue, Zixin Ding, Yichen Shen, Yinjie Wang, Zhenfei Yin, Yingcheng Wu, Yuxin Chen, Mengdi Wang, Ling Yang
Abstract
Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained experience actually improves them over time has not been systematically tested. We introduce PAST-Bench, a benchmark designed to isolate this question. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that turn retained experience on and off. It spans 26 scenarios and 204 episodes across memory, procedural reuse, information gathering, and update. We report both later-task gains and whether those gains follow the intended save, retrieve, and update pathway. Across seven base models and four agent frameworks, improvement is real but uneven across capabilities. Agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended pathway. Guided by these findings, we develop Hermes+, which extends Hermes with five targeted interventions across stages of the agent loop. Hermes+ raises the average gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced, although the effect remains capability- and model-dependent. Together, PAST-Bench and Hermes+ provide an evaluation and diagnostic foundation for studying how persistent agents can progress from retaining experience to systematically improving through it. Code: https://github.com/Gen-Verse/PAST-Bench