Molecular generation is crucial for drug discovery and materials science, especially under data-scarce conditions where traditional generative models often fail to generalize conditionally. To address this, we propose MetaMolGen, a first-order meta-learning-based molecular generator tailored for few-shot and property-conditioned generation. MetaMolGen standardizes graph motifs by mapping them into a normalized latent space and employs a lightweight autoregressive model to generate SMILES sequences that accurately reflect molecular structures. A learnable property projector is integrated into the generation process to support target property conditioning. Experiments show that MetaMolGen consistently generates valid and diverse molecules in low-data regimes, outperforming conventional baselines and demonstrating strong adaptability and efficient conditional generation.