Mnemex

Local semantic code search for Claude Code. Index your codebase once, search it with natural language.

Install

Why this exists

Claude Code's built-in search (grep/glob) works fine for exact matches. But when you're trying to find "where do we handle auth tokens" or "error retry logic" — good luck.

mnemex fixes that. It chunks your code using tree-sitter (so it actually understands functions/classes, not just lines), generates embeddings via OpenRouter, and stores everything locally in LanceDB.

The search combines keyword matching with vector similarity. Works surprisingly well for finding stuff you kinda-sorta remember but can't grep for.

Quick start

That's it. Changed some files? Just search again — it auto-reindexes modified files before searching.

Embedding Model Benchmark

Run your own benchmark with mnemex benchmark. Here are results on real code search tasks:

Summary:
- 🏆 Best Quality: voyage-code-3 (175% NDCG)
- ⚡ Fastest: voyage-3.5 (1.2s)
- 💰 Cheapest: all-minilm-l6-v2 (local, free)

Embedding providers

mnemex supports three embedding providers:

OpenRouter (cloud, default)

Ollama (local, free)

Recommended Ollama models:
- nomic-embed-text — best quality, 768d, 274MB
- mxbai-embed-large — large context, 1024d, 670MB
- all-minilm — fastest, 384d, 46MB

Custom endpoint (local server)

View available models:

Using with Claude Code

Run it as an MCP server:

Then Claude Code can use these tools:
- search_code — semantic search (auto-indexes changes)
- index_codebase — manual full reindex
- get_status — check what's indexed
- clear_index — start fresh

Grep replacement (mnemex rg)

mnemex ships a drop-in rg replacement that Claude Code's built-in Grep tool will call. It runs real ripgrep and mnemex semantic search in parallel, then merges the results…