In the evolving realm of online advertising, meme advertisements have become a powerful tool for brand messaging, blending humor with marketing content. Nonetheless, the subjective nature of humor and the diverse interpretations by audiences present difficulties in making sure that the content is both captivating, engaging and non-offensive. To address this, we introduce MemeAdSense, a novel framework that leverages Vision-Language Models (VLMs) and Large Language Models (LLMs) for careful offensive content detection and correction. Our approach employs a systematic question-answering pipeline to derive explicit insights from memes, enabling both binary and fine-grained multi-class classification with distinct forms of offensiveness, such as sexual/vulgar content, gender objectification, body shaming, and abuse. To enhance robustness, we incorporate CycleRAG, a retrieval-augmented mechanism that refines reasoning by retrieving real-world evidence from online sources, such as news and blogs, referencing past tone-deaf advertisements that harmed brand reputation. To support our work, we introduce MemeADClassification (MAC), a novel dataset curated for reasoning-guided meme classification and correction. Lastly, for derogatory memes, we employ an LLM-based correction mechanism, generating revised, non-offensive meme concepts while protecting humor and brand relevance. Our framework establishes a new benchmark for creating engaging yet ethically responsible meme advertisements.