The proliferation of Internet memes has transformed social media into a highly visual landscape, concurrently creating a sophisticated vehicle for multimodal hate speech. Detecting such content is notoriously difficult due to the "semantic gap", where hateful intent arises solely from the intersection of benign text and neutral imagery. Current State-of-the-Art Large Language Models (LLMs), while possessing advanced reasoning capabilities, suffer from significant limitations: they function as "black boxes," are prone to hallucinations, and exhibit structural biases induced by safety training. To address these challenges, this thesis proposes a novel architecture based on Knowledge Injection. Instead of relying on a single end-to-end model, we introduce a modular framework where a Generative "Meta-Reasoner" arbitrates conflicting signals provided by specialized "Discriminative Experts".