MatrixOrigin Research
Stateful Infrastructure for Production AI Agents
We believe agents are becoming the new unit of computation in the AI era, while the infrastructure that keeps them running reliably and safely over long horizons is still early. MatrixOrigin's research targets production AI agents and the systems questions behind them: how to manage versionable state, how to build trustworthy business context, how to guarantee reliable task execution, and how to support long-term memory and continual evolution.
The work runs in close concert with MatrixOne Intelligence product development — testing new systems ideas against real systems, real workloads and real enterprise scenarios, and turning what holds up into product capability, open-source technology and industry practice.
Research Areas
Cloud-Native Database Architecture
Storage-compute separation, elastic scheduling, and hyper-converged engine design for next-generation HTAP workloads.
AI-Native Data Management
Multimodal data processing, vector indexing, hybrid retrieval, and RAG-optimized data pipelines for enterprise AI.
Git for Data
Zero-copy branching, snapshot isolation, and time-travel mechanisms that bring version control semantics to database systems.
Trustworthy AI Systems
Building secure, efficient, and reliable AI infrastructure — from model serving to agent memory and data governance.
Database Principles for Agentic AI
Apply what database and distributed systems have accumulated over decades — versioning, isolation, optimization, recovery, audit — to agent infrastructure. These mechanisms have been proven in data systems for a long time, and agents face the same class of problems on a different substrate.
Publications
Selected papers from the research team
ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents
NEWPeng Xu, Zuyu Zhang, Yuze Sun, Feng Tian, Long Wang, Chen Zhang
VLDB 2026INTENT-AS-A-TOOL Makes it Easy to Track Agentic Misalignment
NEWYutong Zhang, Jianshuo Dong, Peng Xu, Long Wang, Jie Zhang, Tianwei Zhang, Xiaoping Zhang, Han Qiu
EMNLP 2026Version Control System for Data with MatrixOne
NEWHongshen Gou, Feng Tian, Long Wang, Nan Deng, Peng Xu
arXiv 2026CLAP: Closed-Loop Training, Evaluation, and Release Control for Domain Agent Post-training
NEWFangfei Li, Chenyang Zhao, Long Wang, Feng Tian, Zhiyue Zheng, Lv Guo
CRAE 2026 · Best Poster AwardEdge Fusion of Intelligent Industrial Park Based on MatrixOne and Pravega
Jingwei Yan, Yitong Liu, Shuqi Qin, Yun Shen, Yu Teng, Hongwen Yang
IEEE BMSB 2023Academic Collaborators
Research Projects
Open-source projects with published research behind them
Astra
An agent runtime for enterprise work. Its context assembly mechanism, ContextPipe, brings database query execution to long-horizon context management — data sources catalogued like tables, assembly split into five phases, optimization made deterministic. On the Qutebrowser subset of SWE-bench Pro it cuts tokens by 31% and LLM calls by 23% against append-only assembly.
SiriusDB
A GPU-native SQL analytics engine. Plugs into existing databases like DuckDB via Substrait — achieving over 10× speedup at the same hardware rental cost. Co-developed with Prof. Xiangyao Yu and NVIDIA.
OpenShell
NVIDIA's open-source safe, private runtime for autonomous AI agents, written in Rust. It addresses the same class of problem as Astra: keeping agent execution controllable and auditable in production.
Collaborate With Us
We welcome research partnerships with universities and institutions working on database systems, AI infrastructure, and data management.
Get in Touch