Most code exploration workflows ask an LLM to read many files; codebase-memory-mcp replaces that with a structural knowledge graph so agents ask a single graph query instead of scanning hundreds of files. That design reduces token usage and tool calls dramatically while giving precise, import-aware call resolution across large repositories.
What Sets It Apart
- Graph-first indexing with vendored tree-sitter grammars (158 languages). This eliminates per-project language setup and lets the binary index diverse repos quickly (Linux kernel: full index in ~3 minutes).
- Hybrid LSP semantic resolver layered over tree-sitter. Type- and import-aware resolution for many languages (Python, TypeScript/JSX, Java, Rust, Go, C/C++, C#, PHP, Kotlin) means CALLS/IMPORTS edges are accurate across files and packages, so trace queries follow real runtime call chains.
- Local, single static binary with secured releases. No external APIs or language servers required; binaries are signed, SLSA-attested, and VirusTotal-scanned for reproducible local operation.
- Multi-signal search (bundled embeddings + BM25 + structural signals). Semantic search and structural queries let agents find symbols and code patterns without sending code to a remote LLM or embedding service.
Who It's For & Trade-offs
Great fit if you run LLM-driven coding agents or automated tooling that need precise, low-latency code discovery across medium-to-large repos, especially in polyglot environments. It's also useful for teams that want a committable graph artifact (.codebase-memory/graph.db.zst) to avoid repeated reindexes.
Look elsewhere if you need a full natural-language UI or built-in LLM — codebase-memory-mcp is a structural backend (an MCP server) and expects your agent to translate natural language into graph queries. Also, projects that demand per-project language servers for deep semantic analysis beyond the Hybrid LSP coverage may still prefer an IDE language server.
Where It Fits
Use this as the local code-intelligence backend for MCP-compatible agents (Claude Code, Gemini CLI, Codex CLI, VS Code, etc.). It complements—not replaces—LLMs: let the agent handle NL→query translation and use codebase-memory-mcp to return compact, high-fidelity structural context that dramatically reduces token and tool-call overhead.
