Structure is Context. 0-byte files govern AI — ₩0 infrastructure, ~200x token efficiency. OS-native constraint engine for LLM agents.
# Add to your Claude Code skills
git clone https://github.com/rhino-acoustic/NeuronFS⚠️ v4.2 (2026-03-31) — Auto-Evolution Pipeline Complete
Completed:
- Emit engine optimization: Dual-path sync (local
.gemini/+ global~/.gemini/), strength prefix batch rendering, Kanji micro-opcode double-expression removal for massive token savings- Auto-evolution:
--neuronize(Groq-powered correction log analysis → auto contra neuron generation) and--polarize(positive English neurons → negative/inhibitory auto-conversion)- Maintenance: Duplicate neuron (similar/EN-KR) auto-merge complete (420 neurons stabilized), full PII purge
In Progress: OS auto-start registration (L0), PII git-hook scanner, empty folder quarantine
Breaking:
brain_v4/excluded from git — users mustneuronfs --initto create their own brain.Full changelog: LIFECYCLE.md · LIFECYCLE_EN.md
mkdir replaces system prompts. Folders are neurons. Paths are sentences. Counter files are synaptic weights.
# Create a rule = create a folder
mkdir -p brain/brainstem/禁fallback
touch brain/brainstem/禁fallback/1.neuron
# Compile = auto-generate system prompts
neuronfs ./brain --emit cursor # → .cursorrules
neuronfs ./brain --emit claude # → CLAUDE.md
neuronfs ./brain --emit all # → All AI formats at once
| Before | NeuronFS | |--------|----------| | 1000-line prompts, manually edited | one folder | | Vector DB $70/mo | (folders = DB) | | Switch AI → migration | — 1 second | | Rule violation → wishful thinking | → | | Rules managed by humans | Correction → auto neuron growth |
No comments yet. Be the first to share your thoughts!
mkdircp -r brain/bomb.neuronOption A — Full engine (Go required)
git clone https://github.com/rhino-acoustic/NeuronFS.git
cd NeuronFS/runtime; go build -o ../neuronfs .
./neuronfs --init ./my_brain # Create brain with 7 regions
./neuronfs ./my_brain --emit all # Compile to .cursorrules / CLAUDE.md / GEMINI.md
./neuronfs ./my_brain --api # Dashboard at localhost:9090
./neuronfs ./my_brain --watch # Auto-recompile on changes
./neuronfs ./my_brain --fire cortex/frontend/禁console_log # +1 counter
./neuronfs ./my_brain --grow cortex/backend/禁raw_SQL # Create neuron
Go gives you: --init, --emit, --watch, --fire, --grow, --decay, --api, --supervisor, --neuronize, --polarize.
Option B — Live injection only (No Go needed)
# 1. Create a brain manually (just folders)
mkdir -p ~/NeuronFS/brain_v4/brainstem/禁fallback
echo. > ~/NeuronFS/brain_v4/brainstem/禁fallback/5.neuron
mkdir -p ~/NeuronFS/brain_v4/cortex/frontend/禁console_log
echo. > ~/NeuronFS/brain_v4/cortex/frontend/禁console_log/9.neuron
# 2. Set environment variables
export NEURONFS_BRAIN="$HOME/NeuronFS/brain_v4"
export NODE_OPTIONS="--require $HOME/NeuronFS/runtime/v4-hook.cjs"
# 3. Start your IDE — done
cursor .
No build step. The hook is a single .cjs file with zero dependencies.
Node.js is already included in every Electron-based IDE (VS Code, Cursor, Windsurf).
Daily driver since January 2026. MIT License.
| | Section | Description | |---|---|---| | 💡 | Core Structure | Folder = Neuron, Path = Sentence, Counter = Weight | | 🧬 | Brain Regions | 7 regions, priority cascade, hormone system | | ⚖️ | Governance | 3-Tier injection, bomb circuit breaker, harness | | 🧬 | Neuron Lifecycle | Birth → Reinforcement → Dormancy → Apoptosis | | 🏗️ | Architecture | Autonomous loop, CLI, MCP, multi-agent | | 📊 | Benchmarks | Performance, competitor comparison | | ⚠️ | Limitations | Honest assessment | | ❓ | FAQ | Expected questions and answers | | 📖 | Story | Why this exists |
Unix said "Everything is a file." We say: Everything is folders.
| Concept | Biology | NeuronFS | OS Primitive |
|---------|---------|----------|-------------|
| Neuron | Cell body | Directory | mkdir |
| Rule | Firing pattern | Full path | Path string |
| Weight | Synaptic strength | Counter filename | N.neuron |
| Reward | Dopamine | Reward file | dopamineN.neuron |
| Kill | Apoptosis | bomb.neuron | touch |
| Sleep | Synaptic pruning | *.dormant | mv |
| Connection | Axon | .axon file | symlink |
Paths are natural language commands. Depth equals specificity:
brain/cortex/NAS_file_transfer/ → Category
brain/cortex/NAS_file_transfer/禁Copy-Item_UNC/ → Specific rule
brain/cortex/NAS_file_transfer/robocopy_large/ → Sub-context
禁 (1 char) = "NEVER_DO" (8 chars). 3-5x more meaning density in folder names:
| Hanja | Meaning | Example |
|-------|---------|---------|
| 禁 | Forbidden | 禁fallback |
| 必 | Required | 必auto_reference_KI |
| 推 | Recommended | 推robocopy_large_files |
| 警 | Alert | 警DB_delete_confirm_required |
.cursorrules is a static file you edit manually. NeuronFS is different:
AI makes mistake → correction → corrections.jsonl → mkdir (auto neuron growth)
AI does well → praise → dopamine.neuron (reward signal)
Same mistake 3x → bomb.neuron (entire output halted)
30 days unused → *.dormant (auto sleep)
**🔥 v4.2 Auto-Evolution:**
1. **auto-neuronize**: Groq LLM analyzes correction logs and auto-generates inhibitory (Contra) rules that prevent the same mistakes from ever recurring.
2. **auto-polarize**: Detects positive-form `use_X` rules and auto-converts them to strong Kanji micro-opcode inhibitory forms (`禁X_dependency`).
↓
Automatically reflected in next session's system prompt
7 brain regions layered via Brooks' Subsumption Architecture. Lower P always suppresses higher P.
brainstem(P0) > limbic(P1) > hippocampus(P2) > sensors(P3) > cortex(P4) > ego(P5) > prefrontal(P6)
| Region | Priority | Role | Examples |
|--------|----------|------|----------|
| brainstem | P0 | Absolute laws | 禁fallback, 禁SSOT_duplication |
| limbic | P1 | Emotion filter, hormones | dopamine_reward, adrenaline_emergency |
| hippocampus | P2 | Memory, session restore | error_patterns, KI_auto_reference |
| sensors | P3 | Environment constraints | NAS/禁Copy-Item, design/sandstone |
| cortex | P4 | Knowledge, skills (largest) | frontend/react/hooks, backend/supabase |
| ego | P5 | Tone, personality | concise_execution, korean_native |
| prefrontal | P6 | Goals, projects | current_sprint, long_term_direction |
dopamineN.neuron): Praise → positive weight boostadrenaline.neuron): "urgent" detected → lower P suppresses higher Pbomb.neuron): 3x repeated mistake → entire region output disabled (circuit breaker)16 .axon files connect 7 regions into a layered network:
brainstem/cascade_to_limbic.axon → "limbic" # bomb → block emotions
sensors/cascade_to_cortex.axon → "cortex" # env constraints filter knowledge
cortex/shortcut_to_hippocampus.axon → "hippocampus" # learning results → memory
| Tier | Scope | Tokens | When |
|------|-------|--------|------|
| Tier 1 | brainstem + TOP 5 rules | ~200 | Every session, auto |
| Tier 2 | Full region summary (GEMINI.md) | ~800 | System prompt |
| Tier 3 | Specific region _rules.md full | ~2000 | On-demand when task detected |
※ v4.2: GEMINI.md is now dual-synced to both workspace-local (.gemini/) and global user profile (~/.gemini/) to ensure all agent environments (including Antigravity extension) see the latest brain state immediately.
| bomb location | Result | |---------------|--------| | brainstem (P0) | Entire brain halted. GEMINI.md becomes empty | | cortex (P4) | brainstem~sensors only. Coding region blocked |
bomb doesn't remove the rule — it halts all output from that region. It's an emergency stop button.
Release: rm brain_v4/.../bomb.neuron — delete one file.
15-item automated verification script runs daily like CI:
Neurons are born, reinforced, audited, put to sleep, and killed. Full lifecycle: LIFECYCLE.md
Birth → Reinforcement → Maturation → Dormancy/Bomb → (Apoptosis or Revival)
| Stage | Trigger | Result |
|-------|---------|--------|
| Birth | Correction, Memory Observer, manual mkdir | Neuron folder + 1.neuron |
| Reinforcement | Repeated correction | Counter ↑ → higher placement in prompt |
| Reward | Praise | dopamine.neuron → positive weight |
| Dormancy | 30 days untouched | *.dormant → excluded