What is marrow?

Marrow is a self-hosted search tool for people who use AI assistants like Claude Code during software development. Over time, you accumulate hundreds or thousands of conversation sessions filled with decisions, debugging notes, and the reasoning behind your code. Marrow ingests all of those sessions into a database you control and lets you ask questions about them. Every answer it produces includes citations pointing back to the specific conversation turn it came from. The project is built in TypeScript and runs entirely on infrastructure you own, with no data sent to a shared backend. The tool combines two kinds of search. Semantic search matches by meaning, so you can find related discussions even when the wording differs. Lexical search matches by exact keywords. Marrow blends both approaches and then optionally reranks the top results for better relevance. When you ask a question, it synthesizes a response with numbered citations to the source passages. Embeddings, which are the numerical representations of text used for semantic search, come from a service called Voyage by default, using your own API keys. You can try Marrow without any API key using a built-in demo that sets up a small synthetic corpus. After installing dependencies, a few commands spin up a local database, seed it with sample data, run a search, and tear it all down. For real use, you need a Postgres database with the pgvector extension, your own API credentials, and a set of session transcripts to index. Marrow can read Claude Code sessions by default, and also supports exports from Claude.ai, ChatGPT, Telegram, and generic markdown directories. The project is intentionally scoped to be just an engine: it ingests your data, retrieves relevant passages, and cites its sources. It does not build profiles, generate recommendations, or add analytics layers. Configuration is handled through environment variables and a single config file. Marrow is the open-core version of a retrieval engine that its maintainer, Gnosis Labs, runs in production. The analytics layers built on top remain closed, but the engine itself is fully open under the Apache 2.0 license.