Few-shot event detection intends to identify events of emerging types given merely a few annotated examples for model training. It is quite a challenging task due to insufficient event information and poor generalization owing to limited data. Although promising results have been achieved, previous methods still ignore the semantic relevance between newly emerging event types and old ones. In this work, we propose MetaRAED, a meta learning prototype-based retrieval augmented model for few-shot event detection. Inspired by retrieval augmented generation (RAG) in large language models, we propose to maintain an event knowledge base storing observed events and then retrieve relevant features from it to augment the detection of novel types of events in a discriminative way called retrieval augmented discrimination (RAD). Additionally, we introduce an estimate-then-sample strategy, which models the distributions of events, as well as the model-agnostic meta learning paradigm for training to further strengthen the model generalization. Extensive experiments on a widely-used large few-shot event detection dataset prove the effectiveness and robustness of our proposed model.