Reproduce “From Reflection to Co-Evolution: Self-Evolving Agents with Locally Verified Memory” 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

Xinxiang Yin, Xinglang Zhang, Haodong Chen, Zhiqing Cui, Weidong Tang, Yuxuan Wang, Junrui Wang, Jingyi Zhang and 2 more

Anchor tests candidate memory lessons by replaying their source decisions with and without each lesson, then retains them only when the action-preference shift is large enough and aligned with the decision’s role. With Qwen3.5-27B on ALFWorld, the full gate reaches 96.27% success versus 74.88% for magnitude-only admission, showing direction adds value in this setup, not that every memory helps.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.