MemTensor launches MemT-1.0-Pro, a 554B memory-native model built with NVIDIA
_akhaliq · x · 2026-11-23
MemOS team, in collaboration with NVIDIA AI Dev, launched MemT-1.0-Pro, a memory-native model built on NVIDIA Nemotron 3 Ultra with distributed training via Megatron-Core and Megatron Bridge.
- 554B parameters with native memory: history is stored in the model's memory and reused later without keeping the original context in the query
- Outperforms MemT-1.0-Flash by +4.76 pts on BEAM-1M and +6.51 pts on MemoryAgentBench
- Aimed at long-horizon AI agents; weights and technical report are available on Hugging Face
More from coding & agent
- User asks Anthropic to publish a full human-agent trace of solving a Millennium Prize problem — _arohan_ · 2026-11-23
- Team coordinates AI research agents with a shared research graph — hyperparticle · 2026-11-23
- Kimi K3.1 Pro on Cerebras hits 1375 TPS after 594 agent-run experiments — hyperparticle · 2026-11-23
- Formalizing CFSG at 3x RH scale: Lean is now the bottleneck for AI math — Vjeux · 2026-11-23
- Quote-gate + string checker: agent memory got reliable once the author stopped trusting it — capngreenjeans · 2026-11-23
- 'Permission is the enemy of autonomy': devs slam Codex approval loops over worktree cleanup — Baconbrix · 2026-11-23