Large language models (LLMs) have revolutionized how we interact with technology, but they face a significant hurdle: continual learning. Imagine teaching an LLM new things without it forgetting what it already knows—a tricky task, much like trying to add new files to a full hard drive without deleting anything. This 'catastrophic forgetting' is a major challenge in AI, especially in class-incremental learning (CIL), where an LLM needs to learn new classes without losing its grip on previously learned ones. Traditional methods try to combat this by fine-tuning the LLM, adjusting its internal parameters as it learns. But this often leads to the very forgetting we're trying to avoid. In-context learning offers a promising alternative, where the LLM learns from examples provided within the text prompt, without changing its core parameters. However, this method runs into a roadblock: the limited 'context window' of LLMs. As new classes are introduced, the prompt becomes excessively long, exceeding the LLM's capacity, much like overflowing a mailbox with too many letters. A new approach called InCA (In-context Continual Learning Assisted by an External Continual Learner) tackles this challenge by introducing a clever helper: an External Continual Learner (ECL). Think of the ECL as a smart filter that pre-selects the most relevant information for the LLM. It works by generating descriptive 'tags' from the input text—like keywords or topics—and using these tags to identify the most likely classes for the given input. Then, it provides the LLM with concise summaries of these classes, keeping the prompt short and focused. This method offers several advantages. First, it avoids catastrophic forgetting because the core LLM parameters remain untouched. Second, it tackles the 'inter-task class separation' problem, where the LLM struggles to distinguish between new and old classes. By representing each class with a distinct statistical distribution, InCA helps the LLM maintain clear boundaries. Finally, it makes in-context continual learning scalable, allowing LLMs to learn a large number of classes efficiently. Experiments show InCA significantly outperforms traditional CIL methods. Interestingly, even long-context LLMs, designed to handle larger prompts, perform worse than InCA when overloaded with too much information. This highlights the importance of focused learning, providing the LLM with the right information at the right time. While the research primarily focuses on text classification, future work aims to extend InCA to other NLP tasks like dialogue generation and summarization, paving the way for more versatile and continuously learning LLMs. This innovative approach represents a crucial step forward in building AI systems that can learn and adapt continuously, mimicking the way humans acquire knowledge throughout their lives.