Knowledge base question answering (KBQA) aims to retrieve and reason over structured data to answer natural language questions. Traditional semantic parsing approaches often struggle in few-shot scenarios due to limited annotated data and reliance on rigid query structures, while retrieval-augmented generation (RAG) methods lack explicit symbolic reasoning and are prone to factual errors. We propose FSR-KBQA, a fuzzy symbolic reasoning framework inspired by case-based reasoning, which enables large language models to perform stepwise inference over retrieved knowledge using approximate symbolic programs. This design enhances robustness and interpretability without requiring precise logical forms. Experiments on KQA Pro and GrailQA show that FSR-KBQA achieves strong accuracy and generalization, reaching 68.25% on KQA Pro with Llama-3.1 (70B), and surpassing GPT-4 on GrailQA despite using smaller, open-weight models. These results demonstrate the effectiveness of fuzzy symbolic reasoning as a scalable approach for few-shot KBQA, and we release code and resources at https://github.com/Husense6/FSR-KBQA .