Reproduce “ENGRAM: From Agent Experience to Weights through Verified Consolidation” 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
Ruiyi Yang, Zechen Li, Lihuan Li, Du Yin, Hao Xue, Flora D. Salim
ENGRAM trains adapters on skill-linked agent experience, admitting skills through student mastery or successful teacher demonstrations. It then tests each trained skill against retrieved experience before releasing records from retrieval. On a fixed-checkpoint HotpotQA evaluation, verified release retained 269 of 600 records for retrieval and scored 70.9% token F1, versus 71.1% with all records retrievable; archived records were not deleted.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.