AgentMind

Self-Learning Memory System for AI Agents

Give your AI agent memory. Let it learn your style. Make it better over time.

Quick Start • How It Works • Features • Why AgentMind • 中文文档

---

The Problem

Every time you start a new session with an AI coding agent, it forgets everything:

• ❌ Your coding style preferences
• ❌ The mistakes it already made (and you corrected)
• ❌ Your project conventions and patterns
• ❌ The workflows that work best for you

You end up repeating the same corrections, over and over.

The Solution

AgentMind observes how you work, learns your preferences, and automatically applies them in future sessions. Zero configuration. Install and forget.

Session 1:  You correct the agent → AgentMind observes
Session 2:  Agent remembers → No correction needed
Session 10: Agent anticipates → Works exactly how you like

Why AgentMind

Approach	Remembers?	Learns?	Cross-Session?	Zero Config?
AGENTS.md / CLAUDE.md	✅	❌ Manual	✅	❌ Manual
Chat history	❌ Lost	❌	❌	✅
Custom system prompts	✅	❌ Manual	✅	❌ Manual
AgentMind	✅	✅ Auto	✅	✅

AgentMind is the only solution that learns automatically from your behavior and applies it across sessions.

Quick Start

Install (30 seconds)

# In Claude Code
/plugin marketplace add Youhai020616/Agentmind
/plugin install agentmind@agentmind-marketplace

That's it. AgentMind starts learning immediately.

Or install from source

git clone https://github.com/Youhai020616/Agentmind.git
cd Agentmind && npm install
claude --plugin-dir .

How It Works

          You work with your AI agent
                    │
         ┌──────────▼──────────┐
         │   👁️  OBSERVE        │  Watches corrections, patterns, errors
         └──────────┬──────────┘
                    │
         ┌──────────▼──────────┐
         │   🧠 UNDERSTAND     │  Extracts preferences and patterns
         └──────────┬──────────┘
                    │
         ┌──────────▼──────────┐
         │   💾 MEMORIZE       │  Stores as "instincts" with confidence
         └──────────┬──────────┘
                    │
         ┌──────────▼──────────┐
         │   ⚡ APPLY          │  Auto-injects into future sessions
         └──────────┬──────────┘
                    │
         ┌──────────▼──────────┐
         │   ✅ VALIDATE       │  Tracks if it helped or not
         └──────────┬──────────┘
                    │
         ┌──────────▼──────────┐
         │   🔄 EVOLVE         │  Strengthens good patterns, drops bad
         └──────────┴──────────┘

Confidence System

Every learned pattern gets a confidence score:

Confidence = Frequency × 0.35 + Effectiveness × 0.40 + Human Feedback × 0.25

Confidence	Level	What happens
≥ 0.8	🟢 Core	Auto-applied every session
0.6–0.8	🔵 Strong	Suggested when relevant
0.4–0.6	🟡 Moderate	Available on request
0.2–0.4	⚪ Tentative	Observing silently
< 0.2	🔴 Deprecated	Queued for removal

Four-Level Evolution

Patterns grow smarter over time:

Level 0: Instinct    →  "Use semicolons in TypeScript"
Level 1: Pattern     →  "Always run tests before committing"
Level 2: Strategy    →  "When refactoring, start with types"
Level 3: Expert      →  "Full-stack TypeScript project workflow"

Features

🔍 Automatic Observation (Zero Config)

Observer	What it watches
Intent	When you correct the agent or express preferences
Execution	Tool call patterns and workflows
Evaluation	Error patterns and how they get resolved
Context	Project conventions and coding style

🎮 Commands

Command	Description
/agentmind:status	Learning status overview
/agentmind:instincts	Browse and manage learned patterns
/agentmind:reflect	Review pending learning outcomes
/agentmind:evolve	Trigger pattern evolution
/agentmind:dashboard	Visualize learning data

🔒 Privacy First

• All data stored locally (never sent to any server)
• Only metadata patterns stored (no actual code content)
• You control what's learned via commands
• Full transparency — inspect any stored instinct

Architecture

agentmind/
├── hooks/           # Event observers (intent, execution, evaluation)
├── agents/          # Evolution engine, learning analyst
├── commands/        # CLI commands (status, instincts, reflect, evolve)
├── scripts/lib/     # Core library
│   ├── storage.ts       # Local persistence
│   ├── confidence.ts    # Confidence scoring
│   ├── detector.ts      # Pattern detection
│   ├── evolution.ts     # Pattern evolution
│   └── types.ts         # Type definitions
└── scripts/         # Shell integration

Use Cases

For individual developers:

• Agent remembers your code style across sessions
• Fewer repeated corrections = faster coding

For teams:

• Export team conventions as shareable instincts
• Onboard new team members with learned patterns

For AI agent builders:

• Plug-in memory layer for any agent
• Modular: use storage, confidence, or detection independently

Development

# Prerequisites: Node.js 20+, Claude Code (latest)

git clone https://github.com/Youhai020616/Agentmind.git
cd Agentmind
npm install
npm test          # Run tests
npm run build     # Build

Roadmap

• Core learning engine
• Confidence scoring system
• Four-level evolution
• Claude Code plugin integration
• VS Code extension
• Cross-device sync
• Team sharing
• Multi-agent support
• OpenAI / Cursor integration

Contributing

Contributions welcome! Some areas where help is appreciated:

• New observers — Detect more types of user preferences
• Evolution strategies — Smarter pattern merging and promotion
• Integrations — VS Code, Cursor, OpenCode, other agents
• Testing — More edge cases and real-world scenarios

Star History

🔗 Ecosystem

Project	Description
stealth-cli	Anti-detection browser CLI powered by Camoufox
stealth-x	Stealth X/Twitter automation
dy-cli	Douyin/TikTok CLI
xiaohongshu	Xiaohongshu automation
freepost	AI social media management

License

MIT