AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle

arXiv preprint, 2026.

Weitong Qian*, Beicheng Xu*, Zhongao Xie*, Bowen Fan, Guozheng Tang, Jiale Chen, Xinzhe Wu, Mingtian Yang, Chenyang Di, Jiajun Li, Lingching Tung, Peichao Lai, Yifei Xia, Ziyi Guo, Yanwei Xu, Yanzhao Qin, Shaoduo Gan, Xupeng Miao, Bin Cui

* Equal contribution.

PDF Code arXiv

Overview

AutoSci is a memory-centric agentic system for the full scientific research lifecycle, from literature understanding and idea generation to experimentation, manuscript writing, and rebuttal. It maintains structured, persistent memory across research projects and uses feedback to improve its research procedures over time.

The system combines four modules: SciMem for reusable scientific knowledge and project-level research artifacts, SciFlow for research workflow execution, SciDAG for multi-agent operators and reusable stage-specific templates, and SciEvolve for feedback-driven updates to memory organization, skills, and templates.

All publications