Reproduce “REMORY: Learning Residual Memory for Context Compaction” 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
Hanchen Xia, Baoyou Chen, Yutang Ge, Naihao Deng, Senqiao Yang, Zilong Dong, Weihao Yuan, Siyu Zhu
REMORY supplements a readable summary with bounded learned continuous memory tokens to help a language model continue after context compaction. It conditions the tokens on the summary and appends them after it. On 92 SummHay queries, summary plus REMORY reaches a 43.39 joint score using 5,024 input positions, versus 43.41 with raw context. It does not restore full history.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.