Mathematics is important in the school curriculum, but its abstract nature makes it difficult for children to understand. Traditional education follows a teacher-centred, textbook-based and one-size-fits-all approach, which struggles to accommodate individual learning needs. The context-based learning, such as educational stories, is student-centred and has been shown to help children understand and solve mathematical problems. Based on this, we propose a system that automatically generates educational stories using generative artificial intelligence to provide support for understanding specific mathematical problems and solving them. Presently, we have developed a multi-agent framework based on large language models for story text generation and established a set of criteria for story evaluation. The results indicate that the multi-agent framework enhances the accuracy of the mathematical description in the educational stories, but the generated stories still lack creativity. Future work will focus on producing more personalised multimodal content that aligns better with children’s interests and learning needs.