Knowledge Incorporation In LLM Agents Should Emphasize Cross-phase Performance

Knowledge incorporation (KI) is essential for enabling large language model (LLM) agents to perform tasks that require new or task-specific knowledge. Unlike standalone LLMs, however, LLM agents acquire and incorporate knowledge across a multi-phase lifecycle, including initialization, development, and inference. As a result, KI cannot be fully treated as a within-phase procedure: when there is a phase transition, incorporated knowledge may fail to persist, while later KI may interfere with previously incorporated knowledge. Therefore, we argue that knowledge incorporation in LLM agents should emphasize cross-phase performance. Moreover, we propose two lifecycle meta-criteria, Gaining-ratio and Continuity, for measuring knowledge improvement and unintended interference across phases. These criteria serve both as evaluation tools and as explicit optimization constraints. Building on this formulation, we further propose practical design guidelines for specifying and validating cross-phase KI mechanisms. Overall, our suggested framework provides a unified basis for analyzing and designing KI mechanisms for LLM agents beyond isolated phase-specific updates.