codebase-memory — indexed code knowledge graph over MCP

codebase_memory_mcp is the MCP server for the third-party codebase-memory-mcp binary: it indexes a repository into a persistent, local SQLite knowledge graph and answers structural queries in-process. Everything stays on the machine — no cloud calls, no API keys. Prerequisite: requires the codebase-memory-mcp binary on PATH. Install with: curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash (or npm install -g codebase-memory-mcp). If the server is absent, ask the user to install it per docs/codebase-memory.md; do not pretend to have graph access.

Core workflow

1. Index once: say "index this project" or run codebase-memory-mcp cli index_repository '{"path": "."}' before deep dives.
2. Prefer graph queries over file reads when the question is structural: get_architecture (one call for the whole repo), trace_path (call chains), search_graph (regex/label search), detect_changes (git-diff impact). Only open a file to confirm the findings you already have.
3. Context minimisation: one graph query replaces dozens of grep/read cycles (~99% fewer tokens). Never paste file contents to "explore"; query the graph and read only the cited file, then cite it in your answer.

Synergy with the AOS local stack

• memB: the graph holds structure, memB (memb-mcp) holds project memory and conversation semantics — query memB for "what we decided", the graph for "how the code is wired".
• Synapse: Synapse visualises the codebase spatially (synapse map <repo>); use the graph for precise symbol-level questions and Synapse for spatial orientation, then the source files for verbatim confirmation.