In the era of social media, internet memes have become powerful tools for constructing, reinforcing, and disseminating narratives quickly and accessibly. Understanding memes at this narrative level requires a comprehensive theoretical framework and theory-grounded computational methods. However, current social science approaches focus on manual, qualitative analyses, while scalable computational methods often lack solid theoretical grounding, and no integrated perspective currently combines these two perspectives. To address this, we propose a comprehensive perspective for analyzing meme narratives that integrates theoretical constructs from the social sciences and aligns them with state-of-the-art AI methods and benchmarks. Our framework views memes as fragmented, nontraditional narrative material that can be studied through their specific elements: social identities relating to entities and their roles, legitimation strategies shaping meme messages, and the emergence of shared narratives that reinforce perceptions of social and political in-groups and out-groups. We observe a significant disconnect between theory and computational methods in evaluation practices and model development, a lack of accounting for real-world knowledge and social media rhetoric, and fragmentation across computational approaches that prioritize either scale or depth. These shortcomings highlight the opportunity to operationalize our framework by developing theory-grounded AI benchmarks and systems that enable large-scale narrative analysis.