Player not loading? Watch on YouTube
Codebase Memory MCP builds a local graph of a repository so a coding assistant can query relationships instead of reading whole files to reconstruct them. The video explains the indexing pipeline and MCP integration, rather than walking through installation commands.
The pipeline starts with Tree-sitter syntax parsing, then adds a Language Server Protocol hybrid resolution engine for relationships such as parameter binding and generic substitution. It also maps gRPC, tRPC and HTTP routes to handlers. Louvain community detection groups connected code into functional modules that may differ from the repository's folder layout.
SQLite stores the graph with write-ahead logging. A background watcher checks XXH3 content hashes and re-indexes modified files. The speaker describes automatic configuration changes for Claude Code, Cursor and Zed, plus a hook that augments grep searches with graph data. The supplied metadata lists a single static binary for macOS, Linux and Windows.
The speaker claims a structural query drops from up to 80,000 tokens to 500 and that indexing the 28-million-line Linux kernel takes three minutes. These are reported results, not independently verified benchmarks. The speaker also says searches and database entries stay on the local machine without external cloud infrastructure or APIs; this describes the graph engine, not the connected agent's model hosting.
