MemTensor open-sources MemT-1.0, first memory-native model boosting SWE-bench by 3+ points

机器之心 · wechat · 2026-10-20

MemTensor released and open-sourced MemT-1.0, billed as the first memory-native foundation model that encodes history into continuously updatable internal memory states instead of relying on external memory stores or context re-injection. It ships in two sizes (288B/17B-active Flash and 36B/3B-active Nano), the largest open parameterized-memory models to date.

Architecture: parameterized memory runs in parallel with attention; memory combines local dynamic memory with a hypermemory providing projection/gating, updated via gradients from answer prediction without fine-tuning the base model.
Training: teacher-student KL distillation where the teacher sees full context and the student only sees memory states.
Results: SWE-bench Verified rises from 66.07% to 69.73%; top scores on 11 memory benchmarks (50.6/47.0 vs 38.1 for Temp-LoRA); at 128K-token history, end-to-end inference is 1.86x faster than full-context with 1/66 the persistent storage, widening to 3.99x at 256K.

The team frames memory nativity as an architectural prerequisite for recursive self-improvement (RSI) and long-lived agents that get smarter—and cheaper—with use.

Original post →

MemTensor open-sources MemT-1.0, first memory-native model boosting SWE-bench by 3+ points

More from Models