codebase-memory-mcp: Giving Claude Code & Codex a Queryable Code Knowledge Graph

01 Why AI Still Struggles to Understand Large Codebases

Current AI programming tools understand codebases in a fragmented way. They rely on repeatedly reading files and running grep to piece together the big picture, which consumes massive tokens and often misses critical dependencies. The analogy: asking someone to map an entire maze while only seeing one small wall section at a time.

codebase-memory-mcp takes a different approach: let the AI draw the complete maze map first, then simply query the map.

02 What Is It?

codebase-memory-mcp is a tool that uses MCP (Model Context Protocol) to index an entire codebase into a queryable knowledge graph. Your codebase becomes a "living map": functions are nodes, call relationships are edges. The AI no longer needs to blindly read files — it can directly "query the map" to answer: which functions call this one? What is the blast radius of a change here? What is the overall project architecture?

It also includes an interactive 3D knowledge graph visualization . Nodes glow and are clickable to view code and dependencies; call chains and blast radii are visible at a glance. The visualization runs at localhost:9749.

03 Four Core Advantages

Extremely lightweight. Written in pure C, zero dependencies, single-file binary. Runs locally; code never leaves your machine.

Comprehensive capabilities. Supports openCypher graph query language, covers 20+ semantic relation types, supports 158 programming languages.

Leading precision. In head-to-head tests against CodeGraph and similar tools, it leads across extraction precision, call graph completeness, and query capability.

Zero-friction integration. Claude Code / Codex reads the README and auto-installs/configures; the install script detects existing AI coding tools and completes MCP registration automatically.

04 Benchmark Comparison: Data Doesn't Lie

Using the OpenClaw memory plugin project as a benchmark:

Extraction precision: codebase-memory-mcp accurately identifies inter-function call relationships, including implicit dependencies; competing tools frequently miss cross-file indirect calls.

Call graph completeness: Its graph contains noticeably more nodes and edges, indicating a more "complete" understanding of the codebase.

Query capability: openCypher support enables complex structural questions, e.g., "find all functions that call function A but lack test coverage."

On a large AI project with ~18,000 nodes , it stably completes indexing and querying, with indexing speed remaining within acceptable bounds.

05 Five Practical Scenarios

Rapid onboarding to unfamiliar projects. Provide the project path; within minutes the AI can tell you the number of entry points, overall layer architecture, and how the main data flow operates.

Root cause analysis. Stop guessing at bugs — directly query the call chain to pinpoint the source.

Refactoring impact assessment. Before changing code, ask the AI to analyze the blast radius — how many files and functions will be affected — so you know the scope before you start.

Code health checks. Dead function detection to find "zombie" code; PR impact analysis to predict change scope before merging.

Cross-language pattern transfer. Example: porting GraphRAG core logic to another project in a different language; the AI uses the knowledge graph to produce a complete implementation plan and checklist.

06 Project Link & Quick Start

Repository: https://github.com/DeusData/codebase-memory-mcp Installation is straightforward: Claude Code or Codex reads the README and handles the rest. If you already use AI coding tools, it's essentially "plug and play."

The core value of codebase-memory-mcp: giving Claude Code and Codex a true "code brain" that perceives project structure.