Confucius Agent

Terminal Intarface of QwenTree

🌳 QwenTree — Multimodal Tree File Agent

🏆 Global AI Hackathon Series with Qwen Cloud 2026 Track: Memory Agent

💡 Inspiration

We've spent months building SHIVA, an agentic AI system with persistent memory, skills, and swarm orchestration. We learned a hard truth: most AI agents are black boxes. You give them a task, they give you an answer, but you have no idea what they can actually do or how they think.

We wanted to build something different. An agent whose capabilities are transparent, discoverable, and extensible. An agent where you can literally list its skills like files in a directory.

When we discovered the Qwen Cloud ecosystem — with its incredible multimodal models (Qwen-Max, Qwen-VL, Qwen-Audio, CosyVoice, HappyHorse) all accessible through a single API — we knew we had the perfect platform.

QwenTree was born: a Tree File Agent that organizes its capabilities as a living file system.

🧠 What We Built

QwenTree is a multimodal Tree File Agent where:

31 skills across 10 categories (vision, audio, video, code, web, files, system, memory, media, integrations)
3-tier hierarchical memory (ChromaDB + PostgreSQL + Redis) inspired by the Confucius paper
6,099 memory items from 212 real development sessions — the agent remembers months of actual work
CLI-first with a beautiful Textual TUI
100% powered by Qwen Cloud models

The key differentiator? QwenTree doesn't just answer questions. It remembers. It recalls specific debugging sessions from weeks ago, the exact commands we ran, the errors we encountered, and the solutions we implemented. It's not a chatbot — it's a persistent, context-aware engineering partner.

🏗️ How We Built It

Architecture:

🧠 ORCHESTRATOR (Qwen-Max) → analyzes query, routes to skills 🌳 TREE NAVIGATOR → skills as navigable files (ls, cd, cat, exec) 🗄️ LIVING BRAIN → 3-tier memory (ChromaDB + PostgreSQL + Redis)

We reused our existing infrastructure from months of development:

Memory system from our Confucius Agent implementation
Infrastructure running on Hetzner VPS with Docker
212 real development sessions auto-ingested into the memory tiers

Tech Stack: | Component | Technology | |-----------|-----------| | Orchestrator | Qwen3.7-Max | | Vision | Qwen-VL-Max | | Audio | Qwen-Audio + CosyVoice-01 | | Video Generation | HappyHorse-1.1-T2V | | Memory | ChromaDB + PostgreSQL + Redis | | CLI | Textual (Python TUI) | | Web | Playwright + httpx | | Infra | Docker Compose + Hetzner VPS |

🚧 Challenges We Faced

Memory ingestion: Processing 212 sessions into 3 tiers of structured memory required careful NLP classification to separate canonical facts from observations and raw data.
True multimodality: Integrating vision, audio, video, and code execution through a unified skill tree while maintaining low latency was architecturally complex.
Making it extensible: We wanted anyone to add a skill by simply creating a Python file. Achieving auto-discovery with proper tool calling integration took several iterations.
The demo itself: Recording a 22-minute live demo and editing it down to 5 minutes with a proper voiceover was its own challenge!

🏁 Accomplishments

✅ 31 working skills across 10 categories
✅ 6,099 memory items from 212 real sessions
✅ 89 automated tests — all passing
✅ Fully open source under Apache 2.0
✅ Live production demo running on Hetzner VPS
✅ 3-tier hierarchical memory — a first among hackathon entries

📚 What We Learned

Memory makes the difference: A stateless chatbot is forgettable. An agent that remembers your history becomes a true partner.
Qwen Cloud is powerful: One API key unlocks vision, audio, video generation, and state-of-the-art reasoning. The multimodal capabilities exceeded our expectations.
CLI-first is liberating: Without the complexity of a web UI, we focused entirely on agent capabilities and terminal UX.
Real data beats synthetic: Our 212 real sessions made the agent functional from day one — no placeholder data needed.

🔮 What's Next for QwenTree

Web UI: Complementary interface for non-terminal users
Plugin marketplace: Community-contributed skills
Multi-agent collaboration: Multiple QwenTree instances coordinating
More integrations: Direct connections to CI/CD, monitoring, and business tools

🙏 Acknowledgments

Built with ❤️ by Sebastián Wagner and SHIVA (our AI development partner).

Special thanks to Alibaba Cloud and the Qwen team for: