Reproduce “MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization” 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

Jingxuan Wu, Yuzhe Yang, Yiqiao Huang, Chengzhi Liu, Qingni Wang, Chengxuan Qian, Shutong Wu, Jiawei Zhang and 1 more

MemFold builds a question-relevant textual memory from long conversational history, compresses it into 256 continuous vectors, and trains a reader on its own responses using task rewards plus token-level guidance from a frozen textual-memory reader. On PersonaMem-128K, Qwen2.5-3B-Instruct reaches 88.4% accuracy versus MemGen’s 66.1%, but memory construction still reads the history.

Summary by alphaXiv.

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