MemT 1.0 has been officially released as the industry's first memory-native foundational model, offering two versions with different parameter specifications. It is currently the largest open-source parameterized memory model in the industry. The model breaks through traditional external memory modes by encoding historical information into dynamically updatable internal memory states, parallelizing the original attention mechanism with a parameterized memory module within the Transformer layers. Through the collaborative operation of dynamic local memory and hypermemory, combined with multi-type gating and hybrid storage mechanisms, it achieves native memory reading/writing and cross-session continuity. This design eliminates the need for repeatedly injecting historical texts or frequently fine-tuning the foundational model. The training process employs a three-track data pipeline and distillation alignment techniques, surpassing existing parameterized memory baseline models in agent tasks, multilingual processing, and memory benchmark tests. Particularly in long-history scenarios, its inference speed and storage efficiency significantly outperform full-context solutions, decoupling persistent session states from historical length growth and enabling continuous cost optimization as the model is used. The model fills the gap in memory architectures for large-scale models, allowing experience accumulation to directly participate in subsequent reasoning processes. It lays the architectural foundation for AI recursive self-improvement while helping deployers accumulate proprietary model assets, opening up new pathways for large model development.