The persistent evolution of cyber threats necessitates intrusion detection systems (IDS) that adapt dynamically to novel attacks while preserving knowledge of previous ones. Conventional IDS models face significant challenges, including catastrophic forgetting during incremental updates, a scarcity of labeled data, and pronounced class imbalance. While incremental learning (IL) and semi-supervised learning (SSL) offer promising avenues, existing approaches often overlook critical practical issues like high-dimensional, sparse network traffic, blurred task boundaries, and uneven class distribution shifts. To overcome these limitations, we propose SS-DIL, a semi-supervised incremental learning framework. Our method incorporates: a DP-Means-based memory selection mechanism to preserve informative historical samples; a multi-level knowledge distillation loss aligning model logits, attention maps, and intermediate features to mitigate catastrophic forgetting; and a meta-learning-derived weighting strategy to balance learning across imbalanced classes from labeled and unlabeled data. Extensive experiments on realistic benchmarks demonstrate that SS-DIL achieves state-of-the-art performance in retaining past knowledge, detecting rare attacks, and generalizing from limited labeled data.