The Memory Tensor has released and open-sourced the memory-native foundational model MemT 1.0 series, which includes two models: MemT-1.0-Flash and MemT-1.0-Nano. The models feature built-in trainable memory modules that can autonomously organize and store information, eliminating the need to retrieve the entire historical record each time. Trained in two stages with 34 billion tokens, this series of models achieves up to nearly a 4-fold increase in end-to-end reasoning speed in long-history scenarios, while reducing persistent storage space by approximately 130 times compared to traditional solutions. Multiple memory benchmark tests show that its performance ranks among the top two parameterized memory methods, with average performance comparable to retrieval-augmented solutions, all without the need for real-time historical retrieval. Mounting memory does not affect the model's general capabilities. This model can reduce costs and increase efficiency for developers, providing a stable experience and supporting capability growth. The team will continue to optimize it, laying the foundation for achieving intelligent agents capable of continuous learning and evolution, as well as recursive self-improvement.