Kayori v2

Kayori v2 is an async, adapter-based AI companion built around LangGraph, platform runtimes, and an orchestrated message pipeline.

What Kayori Does

Kayori is an intelligent conversational agent that connects to multiple platforms (Discord, Telegram, Webhook) and provides:

• Natural Conversations - ReAct-based reasoning with context-aware responses using LangGraph
• Tool Execution - Real-world actions like weather lookup, Spotify control, reminders, and web search
• Emotional Intelligence - Mood analysis across 10 bounded emotion dimensions
• Memory - Short-term conversation history plus long-term episodic memory via pluggable backends
• Proactive Behavior - Scheduler-driven actions, inactivity-based history compaction, and internal runtime events
• Audio Support - Full speech-to-text (Whisper) and text-to-speech (EdgeTTS) pipeline

Think of it as a personal AI assistant that lives in one primary chat platform plus webhook, remembers conversations, understands emotions, and can take actions on your behalf.

Architecture Overview

Message Flow:

Input → Gateway BUS → Orchestrator → Agent → Output Sink → Response

Core Components:

• Gateway BUS - Central message bus decoupling input from processing
• Orchestrator - Routes inbound sources and coordinates the runtime turn flow
• Agent - LangGraph ReAct agent with tool access
• Output Sink - Routes responses (direct or multi-platform)
• Scheduler - Publishes delayed and recurring events into the bus

Features

• Primary Chat + Webhook: Run with either Discord or Telegram as the main chat interface, with webhook always enabled
• Audio Pipeline: Whisper STT + EdgeTTS
• Memory Systems: Short-term state + Episodic memory with in-memory, Redis, and Pinecone backends
• Mood Engine: Fast/long emotion model with classifier-driven deltas
• Tools: Reminder, Spotify, Tavily Search, calendar integrations
• Scheduler: Precise/fuzzy triggers, repeating tasks, and inactivity-based compaction timers

Quick Start

Prerequisites: Python >=3.13,<3.14 and API keys for the services you enable

Environment (.env):

API_KEY=your_groq_api_key
PRIMARY_CHAT_APP=discord
DISCORD_BOT_TOKEN=your_discord_token
DISCORD_USER_ID=your_user_id
TELEGRAM_BOT_TOKEN=your_telegram_token
WEBHOOK_BEARER_TOKEN=your_webhook_bearer_token

PRIMARY_CHAT_APP must be either discord or telegram. Webhook stays enabled in both modes, and the inactive chat app does not need valid credentials.

Install:

uv sync

Run:

python main.py

Core Runtime Docs

• Scheduler
• Mood Engine
• Orchestrator
• Episodic Memory
• Conversation Contraction
• Output Sink

Project Structure

src/
├── adapters/     # Input/output, bus, memory, state, scheduler backends
├── agent/        # LangGraph-based agent service and nodes
├── core/         # Orchestrator, mood, scheduler, episodic memory, contraction
├── templates/    # Prompt templates used by core services
├── tools/        # Built-in tools
└── shared_types/ # Models, protocols, trigger types, and envelope types

License

MIT. See LICENSE.