Reproduce “MemT-1.0: Towards Memory-Native Evolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiyu Li, Zeyu Zhang, Linyun Liu, Yiwei Bu, Chenyang Xi, Xichong Zhang, Yi Yu, Yu Yu and 66 more
MemT-1.0 gives models a compact persistent state, updated token by token and read alongside ordinary attention. It extends Metis with token-level updates and Mixture-of-Storage. Across 11 benchmarks, Flash and Nano ranked first and second among parametric-memory methods, averaging 50.6 and 47.0. On harness-triggered SWE-bench Verified, Nano averaged 69.73% success, comparable to textual experience, not evidence of autonomous learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.