The emergence of new deepfake techniques poses significant challenges for forgery detection. As countermeasures, researchers have developed extensible detection methods using incremental learning but suffer from the overfitting issue to old data, especially for identity information. In this work, we analyze the identity dependency problem, where existing detection models tend to memorize irrelevant identity information. To tackle this problem, we propose an identity-agnostic incremental learning framework for extensible deepfake detection. Specifically, an identity disentanglement-based feature extraction module is first proposed to decouple identity and artifact components, supervised by reconstructing content-forgery mixed samples. Additionally, to reduce interference from specific identity content, a replay set construction module is developed, which selects and reorganizes salient forged patches. Experiments demonstrate the effectiveness of our identity-independent strategy for incremental deepfake detection.