# Marrow

An MCP server that gives AI coding agents memory that survives between sessions, through persistent task tracking, versioned docs and semantic code search over a tree-sitter-based index.

A background daemon watches source files in real time, extracts structural skeletons with language-aware grammars, generates vector embeddings, and keeps a semantic index in sync, so agents can navigate code by meaning instead of by filename.

It exposes 23 structured MCP tools and a REST API, letting a heavier model handle architectural work and save state that a lighter local model can pick up later through a session-context call and a semantic task backlog, supporting handoff between different agents.

It is composed of three Python packages, a FastAPI/FastMCP server, a background worker, and a shared schema, storing vectors and metadata in LanceDB and tasks and artifacts as Markdown.

Website: https://github.com/desikai-lab/Marrow