Abstract: With the rapid proliferation of social media, dynamic character memes have become a pivotal vehicle for the dissemination of internet culture, posing significant challenges to the multimodal understanding and creative text generation capabilities of modern models. Existing methods, predominantly designed for static images, struggle to process the temporal dynamics of videos and often rely on costly large-scale datasets for training. To address these issues, this paper introduces DCMC-Agent, an autonomous agent framework that simulates the complete human creative workflow of “observe-think-express” for meme generation. The framework operates in a training-free, zero-shot manner and is capable of empowering any Vision-Language Model. Specifically, its Grid Image Embedding Module(GIE) efficiently encodes dynamic video signals into a static visual representation comprehensible to a VLM. Concurrently, the Semantic Richness Prompting Module(SR-Prompt) guides the model through deep analysis and creative ideation via multi-dimensional prompts, while the Retrieval-Augmented Adaptive Generation Module(RAAG) integrates external knowledge and user feedback for continuous content optimization. Experiments conducted on the DFM-QA dataset demonstrate that DCMC-Agent, empowering the DeepSeek-VL2 model, surpassed all baselines, including GPT-4o and fine-tuned models, across all metrics. This validates the framework's advanced design and effectiveness in handling complex, domain-specific multimodal tasks.