In disentangled representation learning, the goal is to achieve a compactrepresentation that consists of all interpretable generative factors in theobservational data. Learning disentangled representations for graphs becomesincreasingly important as graph data rapidly grows. Existing approaches oftenrely on Variational Auto-Encoder (VAE) or its causal structure learning-basedrefinement, which suffer from sub-optimality in VAEs due to the independencefactor assumption and unavailability of concept labels, respectively. In thispaper, we propose an unsupervised solution, dubbed concept-free causaldisentanglement, built on a theoretically provable tight upper boundapproximating the optimal factor. This results in an SCM-like causal structuremodeling that directly learns concept structures from data. Based on this idea,we propose Concept-free Causal VGAE (CCVGAE) by incorporating a novel causaldisentanglement layer into Variational Graph Auto-Encoder. Furthermore, weprove concept consistency under our concept-free causal disentanglementframework, hence employing it to enhance the meta-learning framework, calledconcept-free causal Meta-Graph (CC-Meta-Graph). We conduct extensiveexperiments to demonstrate the superiority of the proposed models: CCVGAE andCC-Meta-Graph, reaching up to $29\%$ and $11\%$ absolute improvements overbaselines in terms of AUC, respectively.