NORDON — A Local Brain for Your AI Coding Assistant

Your AI coding assistant forgets everything between sessions. NORDON fixes that.

NORDON is a local Rust-powered ML engine that runs on your machine. It captures decisions, patterns, and failures from your Claude Code sessions, then automatically writes the most important knowledge into your project's CLAUDE.md file. When Claude Code starts a new session, it reads CLAUDE.md and instantly knows your project's history.

The result: Claude Code stops repeating mistakes, remembers architecture decisions, and gets smarter with every conversation.

How It Works (The Feedback Loop)

This is the core concept — understand this and you understand NORDON:

1. You code with Claude Code (normal usage)
      ↓
2. NORDON hooks capture every tool use (Edit, Write, Bash, Read)
      ↓
3. Rust ML engine scores each event (decisions > typo fixes)
      ↓
4. High-scoring events become typed memories (Decision, Failure, Constraint...)
      ↓
5. Knowledge Graph builds connections between related memories
      ↓
6. Reasoning engine produces inferences ("this looks like a past bug...")
      ↓
7. CLAUDE.md is auto-updated with your top knowledge
      ↓
8. Next session → Claude Code reads CLAUDE.md → knows your project
      ↓
9. Better decisions, fewer repeated mistakes → NORDON captures more → Loop repeats

You don't need to do anything. Install the extension, and the loop runs automatically.

What Happens When You Install

1. Extension auto-installs Claude Code hooks — no manual settings.json editing
2. Extension auto-starts the daemon — if CLI is installed
3. Extension auto-links your workspace repo — on first connect
4. Status bar shows ✓ NORDON · 5 memories

If the CLI isn't installed yet, the extension shows a Welcome panel with install instructions.

What You See in VS Code

Status Bar

✓ NORDON · 12 memories          (connected, 12 memories stored)
✓ NORDON · 12 memories 🔔2      (2 unread notifications)
⟳ NORDON                        (connecting...)
✗ NORDON                        (click to start daemon)

Sidebar Panel (Status & Plan)

✓ Daemon: Connected
✓ Hooks: Installed
☆ Plan: Free (click to upgrade)
  Version: v0.5.1
  Uptime: 2h 15m
  Health: B
  Repos linked: 1
  Memories: 12
  Events: 34
  Graph: 45 connections
── Brain ──
  ⚠ Similar to past failure: auth token bug     85%
  🛡 Rule applies: never use inline styles       90%

Knowledge Graph

Command Palette → "NORDON: Knowledge Graph"

Opens an interactive force-directed visualization:

• Nodes = memories (sized by importance, colored by type)
• Edges = connections (semantic similarity, causal links, temporal proximity)
• Hover = see memory title and type
• Click = open full memory detail

The graph starts empty and grows organically as you use Claude Code.

Injection Preview

Command Palette → "NORDON: Injection Preview"

Shows exactly what context Claude Code would get:

• Each memory as a card with type badge and importance
• Approve/Reject buttons per memory
• Total token count

What Gets Captured (and What Doesn't)

Captured (high score → becomes memory):

• "We decided to use REST because..." → Decision (score ~0.85)
• "Error: connection refused. Fixed by..." → Failure (score ~0.70)
• "Never use Co-Authored-By in commits" → Constraint (score ~0.90)
• "These files always change together" → Pattern (score ~0.75)

NOT captured (low score → discarded):

• "Read the file contents" → too boring (score ~0.30)
• "ok" → too short
• "sk_live_xxxxx..." → secret auto-redacted, never stored

The ML Engine (18 Modules, Built in Rust)

All ML runs locally. Zero API calls. Zero cloud. < 50ms retrieval.

Commands

Plans

Free plan works fully without an account. Upgrade from VS Code: Command Palette → "NORDON: Upgrade Plan"

Privacy & Security

• 100% local — all data in SQLite on your machine
• Zero cloud — nothing sent anywhere, ever
• Zero telemetry — no usage tracking
• Secret redaction — 31 patterns auto-detected (API keys, tokens, credentials)
• File blocking — 28 sensitive file types never stored (.env, credentials, SSH keys)
• Localhost only — daemon accepts connections only from 127.0.0.1

Technical Details

• Daemon: Rust binary (~8MB), port 7533
• ML Engine: 18 modules, 277 tests, < 50ms retrieval
• Extension: esbuild bundle (~22KB)
• Storage: SQLite with knowledge graph
• Version: v0.5.1

Links

• Website
• Documentation
• Getting Started
• ML Engine
• Pricing
• npm CLI

Built in Rust by SodaSoft