On October 20, Memory Tensor officially announced the open-source release of its Memory Native Model MemT 1.0 series. This innovative series comprises two distinct models: MemT-1.0-Flash and MemT-1.0-Nano. These models boast parameter scales of 288 billion (288B) and 36 billion (36B), respectively, with active parameters amounting to 17 billion (17B) and 3 billion (3B).
A key feature of these models is their integration of sustainable and updatable internal memory mechanisms into the computational process. This allows them to automatically recall historical information and task-related experiences when faced with subsequent, related tasks. Consequently, there's no need to repeatedly input the original historical texts, streamlining the workflow and enhancing efficiency.
Looking ahead, Memory Tensor is committed to further exploring the realms of autonomous experience acquisition and cross-task memory integration. The goal is to gradually construct agents that can evolve through memory-native progression and learn from interactions. This endeavor aims to establish a solid foundation for continuous learning and recursive self-improvement, paving the way for more advanced and adaptive AI systems.