@MemOS_dev@NVIDIAAIDev Putting memory inside the model instead of re-pasting it every turn is a genuinely new direction, congrats on the launch.
One question, what does updating that fixed-size state cost at inference time at 554B? Would love to see Pro numbers.
Memory has mostly lived outside the model: retrieval, longer prompts, more context to re-send every turn. MemT-1.0-Pro #MemT puts it inside.
A 554B memory-native model on NVIDIA Nemotron 3 Ultra that carries history in a fixed-size state, no matter how long the conversation gets.
Congrats @MemOS_dev 👏
Today, we launched MemT-1.0-Pro #MemT, a memory-native model developed in collaboration with
@NVIDIAAIDev.
Built on NVIDIA Nemotron 3 Ultra, with distributed training powered by NVIDIA Megatron-Core and Megatron Bridge.
Here's how it compares to MemT-1.0-Flash:
🧠 554B parameters with native memory
📈 +4.76 pts on BEAM-1M
📈 +6.51 pts on MemoryAgentBench
Store history in native memory. Use it in later interactions — without keeping the original history in the query context.
Advancing memory for long-horizon AI agents.
🎮 Explore the model: https://t.co/6nBlJ7fsvc
📄 Technical report: https://t.co/Qq7IOmFsQi