Continual learning learns a sequence of tasks incrementally with the goal of achieving two main objectives: overcoming catastrophic forgetting (CF) and encouraging knowledge transfer (KT) across tasks. However, most existing techniques focus only on overcoming CF and have no mechanism to encourage KT, and thus do not do well in KT. Although several papers claimed that they could deal with both CF and KT, our experiments show that they suffer from serious CF when the tasks do not have much shared knowledge. In this paper, we study the continual learning of a sequence of natural language processing (NLP) tasks. In NLP, fine-tuning a BERT-like language model using in-domain data is regarded as one of the most effective approaches. However, this approach suffers from serious CF for continual learning. In this paper, we present a novel model called AFK to solve these problems. Experimental results demonstrate the effectiveness of AFK.