Mangrove ecosystems are intrinsically intertwined with human societies, providing vital ecosystem services while simultaneously facing escalating anthropogenic and climatic pressures. Although this complexity has stimulated a growing body of research, existing knowledge remains largely fragmented across isolated case studies. Such fragmentation hinders a holistic understanding of the social–ecological interactions, feedback mechanisms, and adaptive responses that characterize these tightly coupled systems. To address this gap, this study develops a computational framework that integrates Large Language Models (LLMs) with network science to synthesize disparate empirical evidence into a unified causal architecture of Mangrove Social–Ecological Systems (M-SES). Using LLM-assisted screening, we first reviewed 38,816 records, yielding a curated corpus of around 3,500 relevant empirical studies. Building on this foundation, the framework is designed to examine: (i) how thematic modeling can reveal latent semantic structures and delineate functional boundaries between socio-economic development and biophysical processes; (ii) how automated knowledge graph construction can systematically capture complex interactions among anthropogenic interventions, governance responses, livelihood dynamics, and ecosystem outcomes; and (iii) how network topological analysis can translate these relationships into Causal Loop Diagrams (CLDs), thereby uncovering generalized feedback loops and dynamic mechanisms underpinning system resilience. By mapping feedback mechanisms across empirical studies, this research provides a knowledge synthesis approach for understanding how mangrove social–ecological systems respond to disturbance, intervention, and change. The resulting causal architecture can inform adaptive management by identifying key feedbacks, potential trade-offs, and leverage points for sustaining ecosystem functions while supporting long-term social–ecological outcomes in coastal systems.