cursor-memory

Persistent, searchable memory for Cursor AI — You control what your AI remembers.

The Problem • Quick Demo • Installation • Commands • How It Works • Troubleshooting

😤 The Problem

You just spent an hour with Cursor AI figuring out the right architecture. You made decisions, weighed trade-offs, landed on a solution.

Then you open a new chat.

❌ "What database did we choose last week?"
   → "I don't have access to previous conversations."

❌ "Continue the migration plan from yesterday."
   → "Could you provide context about the migration?"

❌ "Why did we pick EFS over EBS again?"
   → "I don't have information about previous decisions."

Every new chat, your AI has amnesia. Every decision you made, every context you built — gone.

The workarounds make it worse:

✨ See It in Action

🎯 Why cursor-memory

You decide what gets saved. Type /memo when something matters — AI creates a structured memo, tags it, and stores it locally. Next time you need it, it's there.

📦 Installation

Prerequisites

• Node.js — Node 20, 22, or 24 LTS recommended (download)
  • Native modules ship prebuilt binaries for these versions, no build tools required.
• C++ compiler (optional) — only needed as a fallback for non-LTS Node versions (Xcode CLI on Mac, build-essential on Ubuntu, VS Build Tools on Windows).

⚡ 2-minute setup

# 1. Install globally
npm install -g cursor-memory

# 2. Setup — downloads model, configures Cursor automatically
cursor-memory setup

# 3. Restart Cursor — done 🎉

The CLI handles everything:

• 📥 Downloads the embedding model
• ⚙️ Configures MCP server for Cursor
• 📋 Sets up AI behavior rules

🤖 Choose your model

All models support 100+ languages and run fully offline after download.

💬 Commands

In Cursor Chat

Three commands. That's it.

AI detects when your question refers to past context and automatically searches your memories — no command needed. /recall is available as a manual fallback.

cursor-memory setup      # First-time setup or switch model
cursor-memory status     # Check MCP, rules, model, database health
cursor-memory reset      # Clear all data and start fresh
cursor-memory -v         # Show version
cursor-memory --help     # Show all commands

🔬 How It Works

💾 Architecture

💾 Save & Search Flow

📁 Scope Isolation

🧪 Tech Stack

🐛 Troubleshooting

Restart Cursor completely (quit and reopen — not just reload window).

cursor-memory status        # check system health

If auto-config failed, manually add MCP server in Cursor:

Cursor → Settings → MCP → Add server, or edit your MCP config file:

{
  "mcpServers": {
    "cursor-memory": {
      "command": "npx",
      "args": ["-y", "cursor-memory"]
    }
  }
}

cursor-memory requires Node 20, 22, or 24 LTS — these are the versions that ship with prebuilt native binaries (better-sqlite3). Other versions may fail to install.

Check your version:

node -v

If you're on a non-LTS version (18, 19, 21, etc.), install an LTS via nvm:

nvm install 22
nvm use 22
npm install -g cursor-memory
cursor-memory setup

If you switched Node versions after installing, just reinstall:

npm install -g cursor-memory
cursor-memory setup

Run cursor-memory setup again to reinstall rules.

If that doesn't work, manually add the rule via Cursor → Settings → Rules → create a new User rule and paste the following:

## cursor-memory MCP

### Auto-recall
BEFORE answering, ask yourself: "Does the user expect me to know something from a previous chat?"
If YES → call search_memory from cursor-memory MCP immediately. Do NOT answer first.
If UNSURE → answer normally, do NOT search.

Signs of past context (any language):
- References to previous decisions ("what did we choose", "as we discussed")
- Continuation requests ("continue the plan", "pick up where we left off")
- "We/our" referring to past work, not general questions
- Temporal cues: "last time", "before", "already", "remember", "yesterday"

### Auto-save awareness
After a substantive conversation, assess whether it produced knowledge worth preserving:

SUGGEST SAVING when:
- A decision was reached (chose X over Y, with reasoning)
- A plan, strategy, or approach was agreed upon
- A problem was analyzed and a solution was identified
- A comparison or evaluation was completed with a conclusion
- Important context, constraints, or requirements were established
- Knowledge was shared that would be useful to recall in future sessions

Do NOT suggest when:
- Quick Q&A with a generic/textbook answer
- Still exploring — no conclusion or decision yet
- User already said /memo in this conversation

How to suggest: at the END of your response, briefly ask:
"This seems worth remembering. Want me to /memo this?"
Do NOT auto-save without user confirmation.

### Commands
/memo → save to memory. With content: save directly. Without content: summarize conversation then save.
/recall → search via search_memory
/forget → delete via delete_memory

• Try more specific terms
• Similarity threshold is 0.2 — very broad queries may not match
• Consider upgrading to a larger model for better recall accuracy

🛠️ Development

git clone https://github.com/tranhuucanh/cursor-memory.git
cd cursor-memory
npm install
npm run build       # build once
npm run dev         # watch mode
node dist/cli.js setup
node dist/index.js

🤝 Contributing

1. 🍴 Fork the repository
2. 🌿 Create feature branch: git checkout -b feature/your-feature
3. 💾 Commit: git commit -m 'feat: your feature'
4. 🚀 Push: git push origin feature/your-feature
5. 🔁 Open a Pull Request

📄 License

MIT — see LICENSE.

🙏 Acknowledgments

• Model Context Protocol SDK
• HuggingFace Transformers.js
• sqlite-vec by Alex Garcia
• Multilingual E5 by Microsoft

Built with ❤️ for developers who are tired of repeating themselves to AI

⬆ Back to top