The ever-growing number of digital financial transactions requires increasingly enhanced mechanisms of detecting anomalies to combat novel fraudsters and money laundering practices. Traditional methods and preconceived graph solutions are often unable to determine the time-specificity of financial criminal activity. In this research, T-GAE, an unsupervised Temporal Graph Autoencoder based on innovative methods that train transactional data as continuous-time graphs and apply temporal graph neural networks (TGNNs) with memory and attention mechanisms to detect anomalies, is introduced. The suggested scheme describes both structural connections (accounts, merchants, and users) and how transactions change over time. The anomaly identification is done by analysis of reconstruction errors, where deviations from the normal pattern of transactions are identified. The validation of T-GAE with the DGraph dataset (1.2 million transactions) and further validation with the El-liptic dataset demonstrates T-GAE to outperform substantially static graph and classic machine learning baselines with an increase in F1-score by 22% and false positives decrease by 18%. This research paper identifies the essentiality of the temporal modeling as the tool to enhance the financial fraud detection.