wtf is agent-knowledge?

Agent-knowledge, nicknamed Janet in its short description, gives an AI coding agent a knowledge base that improves the more you use it. It turns project documents, decisions, meeting notes, and past conversations into a portable wiki written in plain Markdown, stored right inside your project's Git repository. You can ask the agent a question and it answers with citations pointing back to where the information came from. When you add a new source, it folds that information into what the project already knows instead of just filing it away separately. A built in health check finds outdated claims, contradictions between different notes, and pages nothing else links to, so the wiki does not quietly become unreliable over time. Because everything is stored as plain Markdown files tracked by Git, the knowledge base is readable without any special software, easy to review alongside code changes, and can move between different AI agents rather than being locked into one tool. The project follows an open standard called the Open Knowledge Format, created by Google, rather than inventing its own private database or hidden memory system. New information is always added on top of the old rather than silently overwriting history, so you can trace how a piece of knowledge changed over time. The system is organized as a set of skills. Some are triggered automatically by the agent when a task calls for them, such as reading in a new source or answering a question from the knowledge base. Others you trigger yourself by name, like setting up a brand new knowledge base in a project, running a health check on an existing one, or viewing the whole knowledge base as an interactive graph. It installs into agents such as Claude Code, Cursor, and Codex through a shared skills installer, or as a Claude Code plugin. The project is licensed under MIT, with one vendored specification file licensed separately under Apache-2.0.