This chapter studies machine unlearning in complex settings where models must forget not only specific data samples, but also multimodal relations and skills. We introduce MU-Bench, a benchmark for evaluating unlearning methods across diverse tasks and data modalities, including text, image, audio, and video. We introduce MultiDelete, a method for multimodal unlearning that removes targeted cross-modal associations while preserving essential knowledge. For large language models (LLMs) that have learned to use external tools, we introduce ToolDelete, which removes the ability to use specific tools without affecting general language modeling performance. These resources and methods demonstrate that unlearning can be made more effective, scalable, and aligned with real-world needs.