Shanghai team releases open-source memory-native LLM MemT-1.0

The significance here is architectural rather than scale: embedding persistent, updatable memory into the model computation, instead of treating context as something re-supplied each session, points at a different cost and capability profile for long-running agent-style workloads. The two-tier release, with a large Flash model and a much smaller Nano variant, suggests the team sees deployment breadth as part of the pitch, and open-sourcing lowers the barrier for others to test that claim. Whether memory-native design delivers measurable advantages in real multi-task use, and how the open community adopts or extends the weights, is the open question worth watching.