Multimodal fake news detectors are typically trained to work on fixed distributions, making them hardly applicable to ever-changing events. Although it is possible to apply transfer learning to retrain a model on the most recent facts, it will tend to lose its ability to recognize old contents. We mitigate this problem by considering news as a stream of data that becomes available over time and by introducing a continual-learning solution that learns from new events as they become available. Our solution maintains good performance on previously known tasks without limiting the applicability of this solution to older news, leading to a substantial gain of \(+9.22\%\) accuracy on average compared to transfer learning and a \(+3.65\%\) increase in F1 score over the ideal scenario where you train the model on all data in one session. Besides this, we introduce the Tri-Encoder, a state-of-the-art multimodal model that allows the cross-attention mechanism between images and texts to be applied.