Prior work has shown that the multi-relational embedding objective can be reformulated to learn dynamic knowledge graphs, enabling incremental class learning. The core contribution of our work is Incremental Semantic Initialization, which enables the multi-relational embedding parameters for a novel concept to be initialized in relation to previously learned embeddings of semantically similar concepts. We present three variants of our approach: Entity Similarity Initialization, Relational Similarity Initialization, and Hybrid Similarity Initialization, that reason about entities, relations between entities, or both, respectively. When evaluated on the mined AI2Thor dataset, our experiments show that incremental semantic initialization improves immediate query performance by 21.3 MRR* percentage points, on average. Additionally, the best performing proposed method reduced the number of epochs required to approach joint-learning performance by 57.4\% on average.