Users’ purchase decisions are highly influenced not only by their general preferences, but also most recent interactions with a given platform or marketplace. In the case of gathering insights on user behaviour patterns, sequential recommendations are essential for any customer-oriented business, since they enable the prediction and suggestion of the next best basket or action for a user based on their past interactions. The advent of deep neural networks has enhanced the precision of analyzing patterns in sequential recommendations. However, these models are limited by their restricted memory scope and a tendency to prioritize currently popular items, which can decrease recommendation diversity. To address this issue, it is proposed to perform “injecting” of external item-item interaction knowledge, by synthesizing knowledge graph from previously known interactions, into the multi-head attention mechanism during model training, followed by knowledge distillation afterwards for faster inference. Obtained results indicate that the proposed model achieves high inference speed and exceptional precision, effectively addressing the limitations of existing sequential recommendation models.