We propose a benchmark and method for continual multimodal knowledge editing with reliable compositional reasoning.