Recent advancements in large language models (LLMs) have significantly empowered the development of Role-Playing Language Agents (RPLAs), providing promising advancements such as interactive companions, copilots, personalized decision-making, and scenario reconstruction. However, existing works indicate RPLAs suffer from behaving hallucination and counterfactuality, resulting less human-like and abnormal behaviors. Previous works attempt to alleviate the phenomenon via various methods, including precise descriptive datasets, nonparametric prompting, and parametric training. Although their works are promising, there remains deficiencies in terms of effectiveness and efficiency for real-world deployment. Concretely, constructing datasets incurs expensive labor, creating, and learning costs. Nonparametric prompting raises knowledge conflict and long-context issues. Parametric training requires at least full-scale LLMs, which incur substantial costs associated with inference and storage. To address the shortages, this research novelly exploits role-playing pruning, which aims to prune LLMs’ character-unrelated knowledge to alleviate hallucination and counterfactuality while maintain character consistency and memory. The knowledge pruning not only leads to better role-playing ability, but also derives compact RPLAs for deployment. The impact of this research encompasses novelty, effectiveness of character fidelity, and efficiency for deployment.