Existing methods for anomaly detection based on memory-augmented autoencoder (AE) have the following drawbacks: (1) Establishing a memory bank requires additional memory space. (2) The fixed number of prototypes from subjective assumptions ignores the data feature differences and diversity. To overcome these drawbacks, we introduce DLAN-AC, a Dynamic Local Aggregation Network with Adaptive Clusterer, for anomaly detection. First, The proposed DLAN can automatically learn and aggregate high-level features from the AE to obtain more representative prototypes, while freeing up extra memory space. Second, The proposed AC can adaptively cluster video data to derive initial prototypes with prior information. In addition, we also propose a dynamic redundant clustering strategy (DRCS) to enable DLAN for automatically eliminating feature clusters that do not contribute to the construction of prototypes. Extensive experiments on benchmarks demonstrate that DLAN-AC outperforms most existing methods, validating the effectiveness of our method. Our code is publicly available at https://github.com/Beyond-Zw/DLAN-AC.