The Titans architecture and MIRAS framework aim to enhance AI models' long-term memory and processing speed. Traditional Transformer models struggle with lengthy contexts due to increasing computational costs. While some alternatives like linear RNNs and state space models have emerged, they often compress context into fixed sizes, losing critical information. Titans combines the speed of RNNs with the accuracy of transformers, allowing models to adapt in real-time by incorporating unexpected information without needing offline retraining. Titans introduces a novel long-term memory module that functions like a multi-layer perceptron, offering greater expressive power. It utilizes a "surprise metric" to determine the importance of new inputs, allowing the model to prioritize which information to retain. For example, if the model encounters a surprising input while summarizing a financial report, it recognizes the anomaly and stores it in long-term memory. Key features include momentum, which captures both current and past surprises, and a forgetting mechanism that regulates memory capacity by discarding less relevant information. MIRAS offers a broader theoretical framework for analyzing memory in sequence models, identifying four main design elements: memory architecture, attentional bias, retention gates, and memory algorithms. By moving beyond mean squared error (MSE) for optimization, MIRAS allows for the development of more robust models that can handle outliers better. The framework has led to the creation of models like YAAD, which uses a gentler penalty for errors, MONETA, which explores stricter penalties for stability, and MEMORA, which focuses on maintaining a balanced memory state. These innovations position Titans and MIRAS as significant advancements in the field of AI memory and sequence modeling.