Magnolia
An AI agent that keeps what computational work teaches: what was run, what failed, what was learnt.
Magnolia is a persistent-memory AI agent framework for computational research, with scientific tooling, a reviewed memory and provenance by construction.
The project
Built for research that keeps moving.
Magnolia helps a researcher retain and reuse the operational knowledge produced during computational work: which commands ran, which choices were made, what failed and why. It is used daily on the team’s computational chemistry projects.
It was released publicly in September 2026. It is research software under active development, not a finished product.
How it works
From a research task to an inspectable workflow.
Scientific tool integration
Typed wrappers for P2Rank, HADDOCK3, gnina, GROMACS, ORCA, Gaussian and xTB, on a workstation or a Slurm cluster.
A memory that is reviewed
Learnings are staged, corroborated across sessions, and promoted into version-controlled rules only with explicit agreement.
Provenance by construction
Runs, commands, job identifiers and outcomes are recorded as they happen, in files you can read, diff and archive.
Project facts
- Licence
- MIT with a Non-Military Clause
- Lead developer
- Tao Jiang
- Status
- Public since September 2026