kuberwastaken/reference

Reference is a Python tool that solves a specific frustration for people who use multiple AI coding assistants. When you work with tools like Claude Code, Codex, or Cursor, each one operates in its own silo. Claude Code has no idea what you discussed in a Codex session yesterday, and Codex cannot see your Claude history. Reference bridges that gap by letting any of your AI tools search the past conversations and memory files from all the other tools you use. It works as a single server you register once in each AI coding tool. After setup, when you ask an agent to recall what you decided about a certain topic, it can search through transcripts and memory files from every other tool on your machine. The search runs entirely offline using a text-ranking approach called BM25, which finds the best matching snippets from past sessions without sending anything to a remote server. Setup is simple. You run one command to check that the tool can see your existing session files, then add a short configuration line to each coding tool you want to connect. Claude Code, Codex, Cursor, and VS Code are supported out of the box. If you use a different tool, you can add support for it with a few lines in a configuration file. The tool reads two kinds of files: session transcripts, which are the logs of your past AI conversations, and memory files like CLAUDE.md or AGENTS.md that store long-term notes about your projects. Both types are indexed and searchable together. The index updates automatically when files change, so it always reflects your current history without any manual steps. Privacy is a stated priority. Reference is read-only and local-only. It does not send your session history anywhere and does not modify your files. The only external dependency is the MCP library, keeping it lightweight. The project is at an early version and the author is actively seeking feedback.