wtf is research-git?

research-git, also called rgit, is a companion tool to Git for people who work with AI coding agents such as Claude Code, Codex, Gemini CLI, and opencode. Regular Git remembers commits. research-git remembers ideas. When you try an experiment or build a feature, it can be saved as a reusable Feature Capsule that an agent can bring back later, adapted to whatever the code looks like by then. This is not the same as rolling back to an old commit or restoring a deleted branch. Instead, the agent reads the saved capsule and reimplements the idea on top of today's code, leaving you a normal diff to review. Each Feature Capsule stores why the change was made, the relevant code snippets, its settings or parameters, anything it depends on, the results or metrics it produced, and a guide the agent can follow to bring it back. These capsules are kept in a small folder called .rgit next to your repository, separate from your normal git history. Every run started through research-git also saves an exact snapshot of the code at that moment, so a past result can always be reproduced precisely. Setup is a short command line install with pip, followed by an install step that wires research-git into your agent tools, and an init step that creates the local store in your project. After that, you can simply talk to your agent, asking it to save an idea, or later asking it to bring a specific idea back. There are also terminal commands for running experiments, reviewing what has been captured, and comparing different variants against each other. The tool is aimed at situations where someone tries several versions of one thing, such as prompt structures, caching strategies, or machine learning training setups, and later wants to pull the best one forward onto current code. Teams can also share the same capsule graph over a read-only connection, so a teammate's own agent session can regenerate an idea using their own subscription. research-git is released under the MIT license.