Large Language Model Agents Are Not Always Faithful Self-Evolvers

Self-evolving large language model (LLM) agents continually improve by accumulating and reusing past experience, yet it remains unclear whether they faithfully rely on that experience to guide their behavior. We present the first systematic investigation of \emph{experience faithfulness}—the causal dependence of an agent's decisions on the experience it is given—in self-evolving LLM agents. Using controlled causal interventions on both raw and condensed forms of experience, we comprehensively evaluate four representative frameworks across 13 LLM backbones and 9 environments. Our analysis uncovers a striking asymmetry: while agents consistently depend on raw experience, they often disregard or misinterpret condensed experience, even when it is the only experience provided. This gap persists across single- and multi-agent configurations and across backbone scales. We trace its underlying causes to three factors: the semantic limitations of condensed content, internal processing biases that suppress experience, and task regimes where pretrained priors already suffice. These findings challenge prevailing assumptions about self-evolving methods and underscore the need for more faithful and reliable approaches to experience integration.

Self-evolving AI agents are designed to improve by learning from their past experience. For example, after completing previous tasks, they may store either full records of what they did or shorter summaries of what they learned, and then use these experiences to solve future tasks. However, it is unclear whether these agents truly use their stored experience, or whether their performance improves for other reasons. In this work, we test this by deliberately changing or damaging the agents’ stored experience and observing whether their behavior changes. We find that agents often rely on detailed past trajectories, but they frequently ignore or misuse shorter summarized experience, even when that summary is the only guidance available. This issue appears across different agent systems, model sizes, and task types. We also find several reasons behind this problem: summaries can be too vague, models may focus more on nearby task context than stored experience, and some tasks can already be solved from the model’s built-in knowledge. Our findings suggest that future self-evolving agents need better ways to write, retrieve, and use experience so that their improvements are more reliable and easier to trust.