Mnemosyne-Ollama

Ask your codebase questions using local LLMs. Zero config, zero cloud, zero new dependencies.

Bridges Ollama to Mnemosyne via MCP -- the same 6-signal hybrid retrieval that powers Mnemosyne's Claude Code integration, now available with any tool-calling Ollama model.

Install

pip install mnemosyne-ollama

This installs everything: the Ollama bridge, the MCP server, and the Mnemosyne retrieval engine.

Quick Start

cd /your/project
mnemosyne-ollama "how does authentication work"

That's it. Auto-detects your Ollama model, indexes if needed, searches with ranked hybrid retrieval, and returns an answer with file paths and line numbers.

Requirements

• Python 3.11+
• Ollama running locally with a tool-calling model

Supported Models

Any Ollama model with function/tool-calling support:

• qwen2.5, qwen3
• llama3.1, llama3.2, llama3.3, llama4
• gemma3, gemma4
• phi4
• mistral-nemo
• command-r

If no --model is specified, the first installed tool-capable model is used automatically.

Usage

Single query

mnemosyne-ollama "how does the rate limiter work"
mnemosyne-ollama "find all database queries" --model qwen2.5 --budget 12000
mnemosyne-ollama "explain the auth flow" -v   # verbose: shows tool calls

Interactive mode

mnemosyne-ollama
> how does the auth middleware work?
[searches, responds with code citations]
> what about rate limiting?
[follows up with conversation context]
> ^C

Python library

from mnemosyne_ollama import run

result = await run("how does auth work", model="qwen2.5", budget=8000)
print(result.response)

CLI Reference

mnemosyne-ollama [QUERY] [OPTIONS]

positional:
  query                    Question about the codebase (omit for interactive)

options:
  -m, --model MODEL        Ollama model (auto-detected if omitted)
  -b, --budget INT         Token budget for search results (default: 8000)
  -r, --project-root PATH  Project root directory (default: cwd)
  --ollama-url URL         Ollama URL (default: OLLAMA_HOST env or localhost:11434)
  -v, --verbose            Print tool calls to stderr
  --version                Show version

How It Works

mnemosyne-ollama
  |
  | 1. Spawns mnemosyne-mcp as subprocess (stdio)
  | 2. Discovers tools: search, index, stats
  | 3. Sends query + tools to Ollama /api/chat
  |
Ollama (local model)
  |
  | 4. Model calls search tool with your question
  |
mnemosyne-mcp
  |
  | 5. 6-signal hybrid retrieval (BM25 + TF-IDF + symbols + usage + prefetch + RRF)
  | 6. AST-aware compression, budget-cut to token limit
  |
  | 7. Results fed back to model
  | 8. Model generates answer with file citations

Everything runs locally. No API keys, no cloud, no data leaves your machine.

Configuration

Mnemosyne search settings are configured via .mnemosyne/config.toml in your project root (created on first index). See the Mnemosyne documentation for details.

The --budget flag overrides the configured default per query.

Trademarks

Ollama, Qwen, Llama, Gemma, Phi, Mistral, and Command-R are trademarks of their respective owners. mnemosyne-ollama is an independent project and is not endorsed by or affiliated with any of these companies.

License

AGPL-3.0 -- commercial licensing available from Cast Net Technology.

Copyright 2026 Cast Rock Innovation L.L.C.