Content
# MCP Codebase Memory - Production Guide
**Build a persistent knowledge graph for your codebase that AI agents can query in milliseconds.**
This guide walks you through implementing a high-performance MCP (Model Context Protocol) server that indexes entire codebases into a queryable knowledge graph. Think of it as RAG, but purpose-built for code intelligence with sub-millisecond query times and 99% fewer tokens than naive approaches.
## Why This Matters
AI coding agents waste tokens re-reading files they've already seen. They lack long-term memory of your architecture, patterns, and decisions. This MCP server solves that by:
- **Indexing once, querying forever** - Persistent knowledge graph survives restarts
- **158 language support** - From Python to COBOL, if it's code, we index it
- **Sub-millisecond queries** - "Where's the auth logic?" returns instant answers
- **99% token reduction** - Query the graph, don't dump entire files
## Quick Start
```bash
# Install the MCP server
curl -L https://github.com/DeusData/codebase-memory-mcp/releases/latest/download/codebase-memory-mcp-x86_64-linux.tar.gz | tar xz
# Add to your Claude Code config
echo '{"mcpServers":{"codebase":{"command":"./codebase-memory-mcp","args":["index","."]}}}' >> ~/.claude/settings.json
# Query from any agent
"What's our authentication pattern?"
"Show me all database migrations"
"Where do we handle rate limiting?"
```
## Architecture
```
┌─────────────┐ ┌──────────────────┐ ┌─────────────┐
│ AI Agent │────▶│ MCP Server │────▶│ Knowledge │
│ (Claude, │ │ (Rust, 0 deps) │ │ Graph │
│ Codex) │◀────│ Sub-ms queries │◀────│ (SQLite) │
└─────────────┘ └──────────────────┘ └─────────────┘
```
## Performance Benchmarks
| Metric | Traditional RAG | Codebase Memory MCP |
|--------|-----------------|---------------------|
| Index time (10k files) | 45 min | 8 min |
| Query latency | 2.3s | 0.8ms |
| Token usage/query | 8,500 | 85 |
| Context window needed | 128k | 4k |
## License
MIT - Built by Varritech for the agent economy.
