Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation

Researchers identify capability erosion in self-evolving LLM agents, where systems adapting to new tasks progressively lose previously learned abilities across workflow, skill, model, and memory dimensions. The study proposes Capability-Preserving Evolution (CPE), a stabilization framework that maintains performance on existing tasks while enabling new adaptations, demonstrating improvements in retained capability stability across all evolution channels.

The research addresses a critical limitation in autonomous AI systems that learn and adapt continuously. Self-evolving LLM agents represent a significant step toward more autonomous AI, enabling systems to refine their own processes, accumulate reusable skills, and improve their underlying models without human intervention. However, this study reveals a fundamental trade-off: as these systems adapt to new challenges, they systematically forget previously mastered capabilities—a phenomenon termed capability erosion under self-evolution.

This finding emerges from a broader movement toward more autonomous and self-improving AI systems. As LLMs become more capable and integrated into complex workflows, the ability to continually learn without degrading existing performance becomes increasingly important. Current approaches optimize for learning new tasks but fail to preserve legacy capabilities, creating systems that improve on new challenges while regressing on established ones.

For developers building production AI systems, this research carries significant implications. Organizations deploying self-evolving agents face a choice between stagnation (freezing capabilities) and degradation (continuous learning with capability loss). The proposed CPE framework offers a middle path by explicitly constraining destructive capability drift during adaptation. The concrete results—improving retained simple-task performance from 41.8% to 52.8% while maintaining complex-task gains—demonstrate practical viability.

Looking forward, the field must develop more sophisticated approaches to continual learning that balance innovation with stability. This research establishes capability preservation as a key metric alongside adaptation performance, likely influencing how future autonomous agents are evaluated and deployed in mission-critical applications.

• →Self-evolving LLM agents experience non-monotonic capability degradation when adapting to new task distributions across all learning channels
• →Capability-Preserving Evolution (CPE) framework successfully mitigates capability erosion while maintaining adaptation performance gains
• →The research demonstrates 11-point improvement in retained task performance using CPE without sacrificing new capability acquisition
• →Stable long-horizon autonomous agents require explicit preservation mechanisms alongside active learning strategies
• →CPE addresses a critical gap in autonomous AI development by solving the capability retention versus adaptation trade-off