v2.1.0

Eggdrop AI v2.1.0

What's New

Gemini 3 Flash Preview via OpenRouter BYOK

Production now runs on Google Gemini 3 Flash Preview using OpenRouter's Bring Your Own Key feature with a Google Studio API key. This provides a significantly more capable model at near-zero cost.

Configurable Bot Name

Bot nickname is now set via BOT_NAME environment variable and injected into the system prompt at startup. No more hardcoded names — deploy the same codebase for any bot.

Smarter Trigger: Any Nickname Mention

The bot now responds whenever its nickname appears anywhere in a message (e.g. "hey soonyo what's up?"), not just @botname or botname: prefixes. Uses string match for security (no regex injection).

Bug Fixes

• Fixed message duplication bug causing repetitive responses
• Fixed identity confusion where bot would refer to itself in third person
• Fixed bot incorrectly prefixing responses with usernames

Upgrade Notes

• Add BOT_NAME=yournick to gateway/.env
• For Gemini BYOK: add your Google Studio API key in OpenRouter settings, add ~$5 credit balance, set MODEL=google/gemini-3-flash-preview

v2.0.0 - Vector Memory System

Eggdrop AI v2.0.0 - Vector Memory Release

🚀 Major Features

Vector Memory System

The bot now has a complete conversational memory using vector embeddings and semantic search:

• Full channel awareness: Stores ALL channel messages, not just direct mentions
• Semantic search: Finds relevant past messages using vector similarity (Xenova/all-MiniLM-L6-v2)
• Hybrid context: Combines recent messages + semantically similar messages
• Chronological ordering: Context presented in proper timeline for coherent recall
• Persistent storage: SQLite database with sqlite-vec extension for vector operations
• Configurable retention: Default 90-day message retention (configurable or unlimited)

New Endpoints

• POST /store - Passive message storage without LLM response (used for all channel messages)
• POST /chat - Enhanced to retrieve context from vector memory before generating responses

How It Works

1. All channel messages → stored in vector memory via /store endpoint
2. When mentioned → bot retrieves relevant context (recent + similar messages)
3. LLM response → generated with full conversational awareness
4. Bot remembers → facts, preferences, and conversations over time

🔧 Improvements

Performance & Reliability

• Increased API timeout: 30s → 90s (handles slow free tier models)
• Increased Eggdrop timeout: 45s → 100s
• Fixed message duplication in context (3x → 2x)
• Async message storage (doesn't block responses)

Configuration

New environment variables for vector memory:

• MEMORY_ENABLED - Enable/disable memory system (default: true)
• MEMORY_DB_PATH - Database file path
• MEMORY_TOP_K - Max similar messages to retrieve (default: 15)
• MEMORY_RECENT_COUNT - Recent messages to include (default: 5)
• MEMORY_RETENTION_DAYS - Message retention period (default: 90)
• DEBUG_LOG_REQUESTS - Log full context sent to LLM (for debugging)

Model Configuration

• Increased max_tokens: 100 → 300 tokens
• Updated production model: xiaomi/mimo-v2-flash:free
• System prompt improvements for memory usage

📦 Installation Notes

New Requirements

cd gateway
npm run setup  # Downloads sqlite-vec extension

On first run, the embedding model (~90MB) will be downloaded automatically. This takes 10-30 seconds and only happens once.

Upgrading from v1.0.0

1. Pull latest code:

   git pull origin main

2. Install new dependencies:

   cd gateway
   npm install
   npm run setup  # Download sqlite-vec extension

3. Update Eggdrop script:

   cp eggdrop/eggdrop-ai.tcl /path/to/eggdrop/scripts/

4. Rebuild and restart gateway:

   npm run build
   npm start  # or restart your systemd/pm2 service

5. Rehash Eggdrop:

   .rehash
   

Configuration Migration

The .env file has new optional variables. Your existing configuration will continue to work with defaults:

# Optional - vector memory is enabled by default
MEMORY_ENABLED=true
MEMORY_RETENTION_DAYS=90

📝 Documentation Updates

• Comprehensive README updates explaining vector memory architecture
• Updated CLAUDE.md with implementation details
• New troubleshooting sections for memory-related issues
• Enhanced testing instructions

🐛 Bug Fixes

• Fixed TypeScript compilation error with Pipeline type
• Fixed message duplication causing 3x context repetition
• Improved error handling for memory system failures
• Better timeout handling for slow API responses

🔍 Example Usage

Teaching the bot facts:

<user> @bot my favorite color is crimson
<bot> Noted.
... many messages later ...
<user> @bot what's my favorite color?
<bot> Crimson.

Context-aware conversations:

<alice> I'm going to the store
<bob> Can you get milk?
<alice> @bot what did bob ask me to get?
<bot> Milk.

The bot now maintains full conversational context and can recall information from anywhere in the channel history.

⚠️ Breaking Changes

None - this is a backward-compatible release. The memory system is opt-in via environment variables.

📊 Performance Impact

• Storage: ~1-2KB per message (text + 384-dim vector embedding)
• Memory: ~90MB for embedding model (loaded once on startup)
• Response time: +10-50ms for context retrieval (negligible)
• Startup time: +10-30 seconds on first run (model download)

🙏 Credits

Built with:

• OpenRouter - LLM API aggregation
• transformers.js - Embedding model
• sqlite-vec - Vector similarity search
• better-sqlite3 - SQLite driver

---

Full Changelog: v1.0.0...v2.0.0

v1.0.0 - Production-Ready Release

Eggdrop AI v1.0.0

First production-ready release of Eggdrop AI - a minimal, security-hardened LLM-powered IRC bot using OpenRouter.

✨ Features

• LLM Integration: OpenRouter API support with configurable models
• Default Model: qwen/qwen3-4b:free (fast, capable, and free)
• Dynamic Triggers: @botname or botname: - automatically uses your bot's nickname
• Rate Limiting: Per-user 10-second cooldown to prevent abuse
• Production Ready: PM2 and systemd deployment examples included

🔒 Security

• Helmet middleware: Security headers enabled
• Input validation: Comprehensive validation of all user input
• Size limits: 10KB request body, 50KB response size limits
• Timeout protection: 30-second API timeout with AbortController
• IRC injection prevention: Control character sanitization
• Localhost binding: Gateway only accessible locally (127.0.0.1)
• Fail-fast validation: API key validated on startup
• No regex injection: Uses string match instead of regex for security

📦 What's Included

• Gateway (152 lines TypeScript): Express server with helmet, validation, timeouts
• Eggdrop Script (127 lines Tcl): Rate-limited bot with sanitization and error handling
• Documentation: README.md, INSTALL.md, CLAUDE.md for complete setup guidance

🚀 Quick Start

# Clone and setup
git clone https://github.com/splinesreticulating/eggdrop-ai.git
cd eggdrop-ai/gateway
npm install
cp .env.example .env
# Add your OPENROUTER_API_KEY to .env

# Run gateway
npm start

# Install Eggdrop script
cp eggdrop/eggdrop-ai.tcl /path/to/eggdrop/scripts/
# Add "source scripts/eggdrop-ai.tcl" to eggdrop.conf
# .rehash in IRC

📊 Performance

• Minimal dependencies (express, helmet, dotenv)
• Efficient Tcl implementation
• 100 token response limit (configurable)
• Concise, maintainable codebase

🔧 Configuration

All configuration via environment variables and Tcl script variables - no code changes needed for most use cases.

See README.md for full documentation.