This paper focuses on the inter-domain recommendation task in cross-domain recommendation, which aims to address the cold-start problem prevalent in traditional recommender systems. The primary concern is to map a source-domain user’s preference representation to the same latent space as target-domain items. Existing methods fail to strike a balance between generalization and customization during preference mapping, significantly limiting recommendation performance. In this paper, we propose a memory-augmented cluster-level preference mapping framework that captures users’ preference relationships between domains at the cluster level. We learn mapping functions for clusters to transform users’ representations from the source domain to the target domain. To store representations of clusters and mapping function parameters, we employ a key-value memory network. Through manual updates via Write operations after local adaptation on training overlapping users, both key vectors and value memory units of the memory network retain useful information throughout training. By attentively combining projected representations from all clusters using Read operations, we generate source-domain users’ representations in the target domain. Extensive experiments conducted on real-world datasets demonstrate our proposed model’s effectiveness compared with state-of-the-art methods.