We embed personalized LLM Watermarks responsively into generated text with Sparse Autoencoders, with only inference-time sampling on black-box access LLMs and no extra training, and achieved high accuracy while preserving text quality.