Cerebella

# Tool List > **A cerebellum cluster that doesn't belong to a superintelligent brain, but to every independent individual.** > *L1 Self-Evolution Program · L4 Noah Civilization Step 1* [](https://github.com/gymaira1990-jpg/Cerebella) []() []() --- ## We Are Here Mainstream AI is stuck in an "arms race" of brute force: piling on parameters, burning GPU, expensive API calls, and increasingly expensive context windows. But for us - ordinary developers, creators, and individuals - this is an unsustainable path. **The future of personal AI is not in an all-knowing, all-powerful superintelligent brain, but in a cluster of local, specialized, and self-evolving cerebellums.** **Cerebella** is an **ecosystem**. It transforms your personal computer from a passive API client into an active, continuously evolving AI execution hub. --- ## Why "Cerebellum" Instead of "Brain"? Large models have "catastrophic forgetting" - learning new things may cause forgetting old ones. Our solution: **don't put everything into one brain.** | Component | Role | Hardware Load | |------|------|---------| | 🗣️ **Butler Brain (0.5B)** | Intent routing + casual conversation/work classification | 2 threads, 0 VRAM | | 🧠 **Expert Brain (1.5B)** | Knowledge retrieval + skill internalization + direct generation | 2 threads, 0 VRAM | | 🔌 **MCP Plugin Layer** | Hermes Agent native integration, zero invasion | — | > **NPU Mode (Recommended): zero CPU/GPU usage, power consumption < 5W.** > Supports Intel Core Ultra (≥12 TOPS) / AMD Ryzen AI 300+ / Hailo-10H. --- ## Core Workflow ``` User Input → Butler Brain (0.5B) Intent Classification ├── Casual Conversation → Local Interception, no Token waste └── Work → Expert Brain (1.5B) Retrieval / Internalization / Generation ├── Phase 1: File Path + Summary (not learned) ├── Phase 2: Complete Skill Content (LoRA fine-tuned) └── Phase 3: Zero Retrieval Direct Generation (parameter internalized) → Skill Harvest (/harvest) → Store in Memory → Generate Training Data → Next Round LoRA Fine-tuning ``` > See **[Architecture Document](ARCHITECTURE.md)** and **[Roadmap](ROADMAP.md)** --- ## V1 Golden Combination (Purely Take-and-Use) | Module | Selection | Decision | |------|------|------| | **Local Inference** | [Ollama](https://ollama.com) | ✅ Directly Take and Use | | **Memory Warehouse** | [SuperMemory](https://github.com/supermemoryai/supermemory) | ✅ Directly Take and Use | | **Intent Routing** | [ClawRouter](https://github.com/blockrun/clawrouter) | ✅ Directly Take and Use (can be rewritten in 50 lines of Python) | | **Workflow Orchestration** | Dify (lightweight transformation) | ✅ Reuse Orchestration Logic, strip Docker frontend | | **Interaction Interface** | Minimalist HTML single page (≤80 lines) | 🔧 Self-developed, abandon bloated Open WebUI | > **2026 Key Decisions** > ❌ Abandon Open WebUI (Docker too heavy, functionality redundant, 60 lines of HTML replacement) > ⚠️ Unsloth not integrated (GPU conflicts with games, V1 manual trigger training) > 🆕 NPU Mode added (zero CPU/GPU usage, inference power consumption < 5W) --- ## Four-Layer Memory System | Layer | Storage | Speed | |------|------|------| | 🔥 **Hot Memory** | ChromaDB vector database | millisecond-level | | 🌡️ **Warm Memory** | File system + index | second-level | | ❄️ **Cold Memory** | Compressed archive | on-demand | | 🗑️ **Forgotten Zone** | Automatic cleanup | — | **Knowledge won't be lost or infinitely expanded.** --- ## One-Click Deployment Experience ### CPU Mode (Default) ``` 2-4 threads, 0 VRAM usage. Works quietly in the background while playing 3A games. ``` ### GPU Mode (High-Performance) ``` About 2.5-4.5 GB VRAM usage, suitable for dedicated inference workstations. ``` ### 🆕 NPU Mode (Recommended) ``` Zero CPU/GPU usage, power consumption < 5W — Intel/AMD NPU native inference. Automatically detect hardware and recommend optimal mode at first startup. ``` > Modify `deployment_mode` in configuration file `config.yaml` to switch. --- ## Three-Stage Skill Internalization Cerebella's core capability - expert models don't answer all questions upon launch, but grow gradually with usage: | Stage | Capability | Trigger Condition | |------|------|---------| | 1️⃣ **Index Navigation** | Return file path + summary | Vector database has relevant documents | | 2️⃣ **Skill Scheduling** | Return complete content after LoRA fine-tuning | Accumulate ≥ 50 training data | | 3️⃣ **Knowledge Internalization** | Zero retrieval direct generation of answers | Same skill hits ≥ 30 times | > From "tell me where the file is" to "give me the answer directly" - automatic evolution, no manual management needed. --- ## Token Cost Transparency ``` /cost today — Daily report: interaction times, saved Tokens, actual spending /cost week — Weekly report: trend chart + classification statistics /cost all — Cumulative total report: net profit = cumulative savings - running cost ``` **The value of every API call is clearly visible.** --- ## Position in Noah System Cerebella is the **L1 layer implementation** of the **[Noah World Protocol](https://github.com/gymaira1990-jpg/noah-world-protocol)** - the first construction of a four-level recursive experience network. | Layer | Name | Status | |------|------|------| | 🧠 **L1** | Self-Memory (Cerebellum Cluster) | 🟢 **Cerebella ← We Are Here** | | 🏛️ **L2** | Local Civilization Library | ⏳ To Be Built | | 🌐 **L3** | Team Shared Network | ⏳ To Be Built | | 🦅 **L4** | Noah World | 🌟 Framework Ready | - [`noah-world-protocol`](https://github.com/gymaira1990-jpg/noah-world-protocol) — Noah World Protocol - [`babel-experiment`](https://github.com/gymaira1990-jpg/babel-experiment) — Babel Experiment (L4 Prototype) > **L1 Cerebellum Development → L4 Noah Civilization, this is a path.** --- ## Core Documents | Document | Description | |------|------| | **[Architecture Document](ARCHITECTURE.md)** | Complete technical architecture: multi-brain bus + task book system + HLC integration | | **[Roadmap](ROADMAP.md)** | Phase 1→2→3 implementation roadmap | | **[Implementation Plan v2.0](IMPLEMENTATION-v2.0.md)** | ⚠️ Replaced by v3.0, only for historical reference | | **White Paper v3.0** | See vault (internal) | | **HLC v5.0 Technical Specification** | See vault (internal) | --- ## Join Evolution - ⭐ **Star** — Witness the growth of the cerebellum cluster - 🐛 **Issue** — Raise ideas or feedback - 🔀 **Fork & PR** — Build together **Fundamental Purpose: Evolution · Development · History** We don't compete with superintelligent brains; we build a **cerebellum cluster** for every unique individual. Smarter and more cost-effective with use, and more understanding. *Cerebella — Not for the superintelligent brain, but for every independent individual.* *Created by GCAT · 2026*