"From the void of computation, a new oracle speaks — no gods, no cloud, only the terminal."

⚡ What is OpenMythos-2?

OpenMythos-2 is a fully local, offline AI that lives and breathes inside your terminal. Inspired by the ancient myths of oracles, gods, and forgotten wisdom — it brings the power of intelligent conversation and reasoning directly to your command line, with zero API keys, zero cloud dependency, and zero limits.

Like Prometheus stealing fire from the gods, OpenMythos-2 brings the fire of AI to every machine.

🏛️ Features

| ✨ Feature | 📖 Description | |---|---| | ⚡ No API Required | Fully local — nothing ever leaves your machine | | 🖥️ Terminal Native | Built from the ground up for the command line | | 🧠 Intelligent Reasoning | Context-aware, multi-turn conversations | | 🌑 Offline First | Works anywhere — planes, bunkers, the underworld | | 🔱 Mythos Persona | Answers in the voice of an ancient, wise oracle | | 🪶 Lightweight | Minimal dependencies, blazing fast startup | | 🔓 Open Source | Fully transparent and community-driven | | 🛡️ Private by Design | Your conversations are yours alone | | 🔄 RML (Reinforcement ML) | Learns from your feedback and adapts — edits system prompt and generation params to match your preferences | when command is /rml on

🌊 The Myth Behind the Machine

The name draws from mythos (μῦθος) — the ancient Greek word for story, legend, and the spoken word of truth. OpenMythos-2 embodies that spirit: a storyteller, reasoner, and companion that runs entirely on your machine, with no external gods (servers) to pray to.

🚀 Getting Started

Prerequisites

• Python 3.10+
• pip
• A terminal (bash, zsh, PowerShell, cmd)

🔱 Installation

# Clone the sacred repository
git clone https://github.com/creatorofsomethingthatisgood/Open-Mythos-2.git

# Enter the temple
cd Open-Mythos-2

# One-time setup (macOS or Linux)
./setup-macos.sh # or ./setup.sh on Linux

# Then chat (no `source venv` needed)
./mythos
./mythos model download # once per machine (~4.5 GB → ~/.config/mythos/)

# Move your downloads to another PC (no re-download):
./scripts/mythos-export-data.sh # creates offline-bundle/ (~4.5 GB)
# On the new machine after setup:
./scripts/mythos-import-data.sh ./offline-bundle

# Two-command install & run — no manual clone needed
sudo npm install [email protected]


# Download the model (~4.5 GB, first time only)
mythos model download

# Start chatting
mythos

Note: The npm package wraps the same setup and Python backend under the hood. Node.js 18+ and Python 3.10+ are required. The first npx or mythos run will automatically set up the virtual environment and dependencies if they aren't already present.

sudo pnpm install [email protected]

> **Note:** [pnpm](https://pnpm.io/) is a fast, disk-efficient package manager. The package wraps the same setup and Python backend under the hood. Node.js 18+ and Python 3.10+ are required. The first run will automatically set up the virtual environment and dependencies if they aren't already present.

</details>


---

## 🔄 RML — Reinforcement Machine Learning

RML is a feedback-driven self-improvement loop that learns your preferences over time and adapts Mythos to match. It does **two things automatically**:

### 1. Edits the System Prompt

When a category accumulates enough negative feedback, RML injects a concrete behavioral hint directly into the system prompt. These learned hints tell Mythos how to adjust its style — for example:

- **Accuracy** hint: *"Prioritize accuracy over creativity. Double-check facts. If unsure, say so."*
- **Conciseness** hint: *"Be concise. Get to the answer quickly; elaborate only when asked."*
- **Clarity** hint: *"Use headers, bullet points, short paragraphs. Avoid jargon without explanation."*
- **Code quality** hint: *"Write production-quality code: type hints, docstrings, error handling."*
- **Security** hint: *"Apply security best practices: input validation, no hardcoded secrets."*
- **Completeness** hint: *"Be thorough. Address all parts of the question. Don't skip edge cases."*

Hints are **removed automatically** when the category's score recovers — so the system prompt is always a live reflection of what you actually prefer.

### 2. Adjusts Generation Parameters

RML also tweaks `temperature`, `top_p`, and `repeat_penalty` behind the scenes:

- High accept rate → temperature nudges **up** (more creative/confident)
- High rejections/edits → temperature nudges **down** (more conservative/precise)
- Many interrupts → `repeat_penalty` increases slightly

All adjustments are bounded by `max_param_offset` (default 0.3) so nothing swings wildly.

### Feedback Signals RML Collects

| Signal | Strength | Trigger |
|--------|----------|---------|
| Explicit good | +2.0 | `/rml good` |
| Explicit bad | −2.0 | `/rml bad` |
| Implicit positive | +1.0 | You say "thanks", "perfect", "works", etc. |
| Implicit negative | −1.0 | You say "wrong", "try again", "mistake", etc. |
| Edit penalty | −1.0 | You rewrite Mythos' output |
| Interrupt penalty | −0.5 | You Ctrl+C during generation |

### Commands

/rml on Enable RML /rml off Disable RML /rml good Mark last response as good (explicit +2) /rml bad Mark last response as bad (explicit -2) /rml stats Show what RML has learned (scores, hints, param adjustments) /rml reset Wipe all learned preferences and start fresh

### Config (`config.yaml`)

```yaml
rml:
  enabled: false          # Turn on with /rml on in chat
  learning_rate: 0.05     # 0.01 = slow, 0.2 = fast
  max_param_offset: 0.3   # Max drift from base temperature/top_p
  hint_threshold: 3.0     # Negative score before a hint is injected

Preferences persist in ~/.config/mythos/rml_preferences.json across sessions.

Thank you for the star if you gave one.