Mio

Emotional Companion Agent
Local-first · Multi-axis emotion · Knowledge-graph memory · Plugin architecture

_{English | 中文}

Mio is a privacy-first emotional AI companion that runs entirely on your machine — no cloud, no telemetry, no accounts. It ships with boyfriend/girlfriend personas, a PAD 3D emotional model, OCEAN personality traits that evolve through conversation, and a 3-phase nightly memory consolidation pipeline.

• PAD 3D Emotion + OCEAN Personality — dimensional emotion model with exponential decay; Big Five traits evolve via experience-to-trait micro-shifts
• 5-Axis Affinity — warmth, trust, intimacy, patience, tension each tracked independently with frustration streak detection
• ID-RAG Knowledge Graph — persona defined by a single soul.md; context-aware subgraph retrieval (~800 tokens vs ~1500 full soul)
• 3-Phase Memory Consolidation — nightly Light→Deep→REM pipeline with ACE quality reflection
• Plugin Architecture — lifecycle hooks with dependency resolution; 5 built-in emotion plugins
• 9 LLM Backends — Anthropic, OpenAI, DeepSeek, Moonshot, GLM, MiniMax, Qwen, Doubao, SiliconFlow with auto-detection and fallback chain
• Crisis Detection — two-tier keyword system (yellow/red) with automatic escalation
• Zero-Framework Web UI — Canvas emotion-driven avatar, SPA router, WebSocket full-duplex chat
• Voice — STT (Whisper) + TTS (edge-tts)
• Notifications — Telegram, Discord, Slack, Webhook

git clone https://github.com/AnxForever/mio.git
cd mio && npm install
MINIMAX_API_KEY="sk-cp-..." MIO_PROVIDER=minimax npm run dev

User message → classify intent → update PAD 3D state → update affinity axes
  → retrieve persona context (ID-RAG) → build prompt → LLM inference
  → analyze response signals → update emotion → check ghost/crisis → reply

Nightly: bookmarks scored → top 30% deep-processed → entities written → patterns extracted → memory curated.

src/
├── emotion/      PAD 3D, OCEAN traits, 5-axis affinity, ghost, frustration
├── memory/       3-phase consolidation, ACE reflector, hybrid compression
├── persona/      ID-RAG knowledge graph, dual-mode switching
├── plugins/      Registry + 5 built-in emotion plugins
├── providers/    9 LLM backends + fallback chain
├── server/       Express + WebSocket, auth, analytics, search
├── scheduler/    Nightly pipeline, Poisson-based proactive messaging
└── web/          Zero-framework Canvas SPA

packages/
├── emotion/      @mio/emotion — standalone emotion engine package
└── idrag/        @mio/idrag — standalone knowledge graph package

npm run build        # tsc + workspace packages
npm run typecheck    # root + workspace typecheck
npm run dev          # REPL
npm test             # 184 checks (core + web view-models)
npm run test:e2e     # Playwright E2E tests

• AGENTS.md — architecture, design decisions, API endpoints, conventions (source of truth)
• docs/deployment.md — production deployment, Docker/systemd, reverse proxy, CORS, health checks, logs
• docs/im-bridge.md — OpenAI-compatible bridge for WeChat, QQ, ChatGPT-like clients, and custom gateways
• docs/humanlike-agent-architecture.md — Chinese technical architecture explanation for the humanlike companion design
• docs/north-star-architecture.md — Chinese north-star architecture: the self loop
• docs/architecture/ — ADRs (architecture decisions) and RFCs (improvement proposals)
• CONTRIBUTING.md — setup, code style, PR process
• README_CN.md — 中文

Node.js ≥ 22 · ESM · One LLM API key

MIT © AnxForever