Sending source-derived notes to a hosted embedding endpoint is a non-starter in plenty of shops, and routing through Ollama removes that objection entirely. The trade is setup: Ollama has to be installed and running, and the package is built locally rather than installed from an index. Two tools is a narrow surface, but storing and retrieving is the whole job of a memory bank.
A knowledge store for code repositories. Notes about a codebase are embedded and kept in a local ChromaDB, and an assistant retrieves the relevant ones as context instead of re-deriving them each session.
- add_knowledge stores a piece of knowledge under a path with its content and metadata
- search_knowledge queries the store and returns matches, with a limit that defaults to 5
- Embeddings are generated by a local Ollama service, so the content never goes to a hosted embedding API
- The store is local and persistent, which makes it usable as a project memory bank rather than a per-session cache
- A clinerules template ships with the repository for wiring it into a project's development workflow
Python 3.8 or higher, chromadb, and a running Ollama service — `ollama serve` must be up before the server works. There is no package index step: build the wheel from a checkout with `python -m build` and install it, or use `pip install -e` for development. It is launched as `python -m code_knowledge_tool.mcp_tool`.
Build from source — clone the repository and build it, then point your client at the binary