A cyber mirror that reflects the real you. · 一面赛博镜子,照见真实的自己。
Mirror is an open-source, local-first, self-evolving personal AI agent. It learns your behavior, remembers your preferences and grows its own capabilities — the more you use it, the better it knows you.
| Typical AI assistant | Mirror | |
|---|---|---|
| Capability | Fixed at release | Self-evolving — grows with use |
| Memory | Chat history | Personality model: preferences, habits, thinking patterns |
| Privacy | Data uploaded to the cloud | Local-first, Ollama supported |
| Customization | Prompt tuning | Tool self-synthesis — learns missing skills on the spot |
| Openness | — | MIT, fully transparent |
Day 1 Day 7 Day 30
0 tools ──────────────→ learns weather/alarms ────→ sleep analysis / weekly reports / proactive suggestions
chat only starts to be useful irreplaceable
The Yunjue Agent paper proved the feasibility of "self-evolution". Mirror builds on it with three key differentiators:
- Not just tools — tools + memory + preferences + workflows, evolving across every dimension
- Personal — constructs a structured model of you, not a generic agent
- Minimal deployment — one
pip install, running in 5 minutes
# Install
pip install mirror-agent
# Initialize
mirror start
# Start from zero — Mirror knows nothing, but will gradually learn everything🚀 Mirror v0.1.0
model: gpt-4o
tools: 0
interactions: 0
EGL: ∞ (not yet evolved)
you: check tomorrow's weather in Hangzhou
Mirror: I don't have a weather tool — let me create one...
✓ new tool: get_weather synthesized
Hangzhou tomorrow 18–26°C, cloudy turning sunny ☁️→☀️
you: good day for a run?
Mirror: based on your preference (you like evening runs) and tomorrow's weather...
suggested window 17:00–18:00 — comfortable temperature, low wind.
┌─────────────────────────────────────────┐
│ Mirror Agent │
│ │
│ ┌─────────┐ ┌────────────────┐ │
│ │ Manager │──→│ Tool Developer │ │
│ │(dispatch)│ │ (synthesize) │ │
│ └────┬────┘ └────────────────┘ │
│ │ │
│ ┌────▼────┐ ┌────────────────┐ │
│ │Executor │──→│ Integrator │ │
│ │ (ReAct) │ │ (compose reply)│ │
│ └─────────┘ └────────────────┘ │
│ │
│ ┌──────────────────────────────────┐ │
│ │ Memory Layer │ │
│ │ • preference model • personality│ │
│ │ • memory │ │
│ └──────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────┐ │
│ │ Sensor Integration │ │
│ │ • Apple Health • Google Fit │ │
│ └──────────────────────────────────┘ │
└─────────────────────────────────────────┘
- v0.1.0 — core engine: agent + tool self-synthesis + sandbox
- v0.2.0 — LLM integrations (OpenAI / Anthropic / Ollama)
- v0.3.0 — memory layer + preference learning
- v0.4.0 — health data (Apple Health / Google Fit)
- v0.5.0 — web UI + chat interface
- v1.0.0 — stable API + full docs + skills marketplace
Mirror is at an early stage — bug reports, feature ideas, documentation and PRs are all welcome.
See CONTRIBUTING.md.
MIT License — use, modify and distribute freely.
AI should grow with you, not watch you.