Recursive self-improvement—the ability of an AI agent to turn accumulated experience into better future behavior—is a cornerstone of advanced agentic systems. Personal AI agents, which retain preferences, task histories, tool routines, and learned skills across sessions, offer a concrete setting to study this capability. Yet, until now, whether retained experience actually improves agents over time has not been systematically tested.
To address this gap, researchers introduce PAST-Bench, a benchmark designed to isolate the effect of retained experience. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that toggle retained experience on and off. The benchmark spans 26 scenarios and 204 episodes across four capability areas: memory, procedural reuse, information gathering, and update.
The study evaluates seven base models and four agent frameworks. Results show that improvement from retained experience is real but uneven across capabilities. Notably, agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended save, retrieve, and update pathway.
Guided by these findings, the team developed Hermes+, an extension of the Hermes agent 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gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced—though the effect remains capability- and model-dependent.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 merely retaining experience to systematically improving through it. The code is available on GitHub.