Abstract: Semantic ID (SID)-based generative recommender systems represent items as sequences of discrete identifiers and generate the identifier of the next item based on a user’s interaction history, constituting a representative application of generative retrieval in intelligent information systems. Upon receiving an interaction-level deletion request, a trained model must reduce the influence of the deleted interaction, preserve recommendation utility, and ensure that generated identifiers correspond to real items. Existing recommendation unlearning methods are primarily designed for matrix factorization, embedding-based models, or sharded training, making it difficult to simultaneously address sliding-window propagation and valid-identifier constraints in SID-based generative recommendation. To address these challenges, we propose an interaction-level machine unlearning method for SID-based generative recommender systems. Based on the distinct roles of a deleted interaction as either a prediction target or part of the historical context in sliding-window training, the method establishes a traceable mapping from deletion requests to the corresponding training windows. Using the original model as a reference, it suppresses the generation preference for deleted SIDs in the target forgetting set, reconstructs affected contexts, and combines the reconstructed and unaffected windows to construct a retained training set, thereby jointly optimizing the forgetting and retention objectives. At inference time, beam search is constrained by a prefix trie of valid SIDs. Experiments with 1% interaction-level deletion on the Amazon Beauty, Sports, and Toys datasets demonstrate that the proposed method closes 97.8%–99.7% of the DelTarget Hit@10 gap between the original model and the retrained reference, while preserving recommendation utility on the post-deletion data distribution and ensuring valid SID generation, with substantially lower training cost than retraining.