This paper proposes a framework called Self-Synthesized Rehearsal (SSR) to mitigate catastrophic forgetting in large language models (LLMs) during continual learning. Traditional rehearsal-based methods rely on previous training data to retain model knowledge, but this data may not be available in real-world applications. SSR addresses this by using the LLM to generate synthetic instances for rehearsal. The process involves in-context learning to generate synthetic instances, refining their outputs with the latest LLM, and selecting diverse, high-quality synthetic instances for future stages. Experimental results show that SSR achieves superior or comparable performance to conventional methods while being more data-efficient. It also preserves the generalization capabilities of LLMs in general domains. SSR is data-efficient and flexible, not requiring real data or additional generative models. The framework is evaluated on the SuperNI dataset, demonstrating its effectiveness in mitigating catastrophic forgetting. SSR outperforms other baselines in terms of AR and BWT metrics, and maintains strong performance in generalization tasks like AlpacaEval and MMLU. The framework is also shown to be robust to variations in rehearsal ratio and synthetic instance selection. SSR provides a promising solution for continual learning of LLMs in real-world settings, with potential for maintaining the acquired abilities of LLMs.This paper proposes a framework called Self-Synthesized Rehearsal (SSR) to mitigate catastrophic forgetting in large language models (LLMs) during continual learning. Traditional rehearsal-based methods rely on previous training data to retain model knowledge, but this data may not be available in real-world applications. SSR addresses this by using the LLM to generate synthetic instances for rehearsal. The process involves in-context learning to generate synthetic instances, refining their outputs with the latest LLM, and selecting diverse, high-quality synthetic instances for future stages. Experimental results show that SSR achieves superior or comparable performance to conventional methods while being more data-efficient. It also preserves the generalization capabilities of LLMs in general domains. SSR is data-efficient and flexible, not requiring real data or additional generative models. The framework is evaluated on the SuperNI dataset, demonstrating its effectiveness in mitigating catastrophic forgetting. SSR outperforms other baselines in terms of AR and BWT metrics, and maintains strong performance in generalization tasks like AlpacaEval and MMLU. The framework is also shown to be robust to variations in rehearsal ratio and synthetic instance selection. SSR provides a promising solution for continual learning of LLMs in real-world settings, with potential for maintaining the acquired abilities of LLMs.