Canon Keeper MCP Server

An MCP (Model Context Protocol) server that persists learnings from Copilot conversations to copilot-instructions.md.

No API key required! Copilot extracts learnings using its own LLM, this tool handles deduplication and formatting.

Quick Install

pip install canon-keeper-mcp
python -m canon_keeper_mcp install

# Or specify a workspace
python -m canon_keeper_mcp install --workspace /path/to/project

The installer will:

1. ✅ Configure .vscode/mcp.json with the canon-keeper server
2. ✅ Enable MCP in .vscode/settings.json
3. ✅ Create/update .github/copilot-instructions.md with:
   • Memory Persistence Protocol (@History directive)
   • Best practices template
   • Session Learnings Log table

Features

• Zero Configuration: No API keys needed - uses Copilot's built-in LLM
• Smart Deduplication: Uses Jaccard similarity to avoid duplicate entries
• Simple Integration: Works with Copilot's existing capabilities
• Markdown Formatting: Returns properly formatted table rows for the Session Learnings Log

Usage

After installation, reload VS Code and use these trigger phrases in any Copilot chat:

User: @History save what we learned
User: save this learning  
User: remember this
User: add to memory

Copilot will:

1. Extract learnings from the conversation (using its own LLM)
2. Call the MCP tool to deduplicate against existing entries
3. Append new learnings to your copilot-instructions.md
4. Report what was saved/skipped

Manual Installation

If you prefer to install manually:

1. Install the package

pip install canon-keeper-mcp

2. Configure MCP server

Add to .vscode/mcp.json:

{
  "servers": {
    "canon-keeper": {
      "type": "stdio",
      "command": "/path/to/your/python",
      "args": ["-m", "canon_keeper_mcp"]
    }
  }
}

Note: Replace /path/to/your/python with the absolute path to your Python executable (e.g., C:/Users/You/project/.venv/Scripts/python.exe on Windows or /home/you/project/.venv/bin/python on Linux/Mac).

3. Enable MCP in VS Code

Add to .vscode/settings.json:

{
  "github.copilot.chat.modelContextProtocol.enabled": true
}

4. Add directive to copilot-instructions.md

Add this to your .github/copilot-instructions.md:

### Memory Persistence Protocol (@History)

**Rule:** When the user includes `@History`, `save this`, `remember this`, or `add to memory`:

1. Extract learnings from the conversation as an array of objects with: topic, decision, rationale
2. Read the current `copilot-instructions.md` file content
3. Call MCP tool `canon_keeper.extract_and_save_learnings` with:
   - `learnings`: Array of extracted learnings
   - `current_instructions`: The full content of copilot-instructions.md
4. If `markdown_to_append` is non-empty, append it to the Session Learnings Log table
5. Report what was saved and what was skipped as duplicates

**Trigger phrases:** `@History`, `save this`, `remember this`, `add to memory`

5. Add Session Learnings Log table

## Session Learnings Log

| Date | Topic | Decision | Rationale |
|------|-------|----------|----------|

MCP Tools

extract_and_save_learnings

Deduplicates learnings against existing entries and formats for markdown.

Input:

• learnings (array): Array of learning objects, each with:
  • topic (string): Short topic name
  • decision (string): The decision or learning
  • rationale (string): Why this decision was made
• current_instructions (string): Current copilot-instructions.md content

Example Input:

{
  "learnings": [
    {
      "topic": "MCP Config Format",
      "decision": "Use 'servers' key not 'mcpServers'",
      "rationale": "VS Code MCP expects this exact format"
    }
  ],
  "current_instructions": "... content of copilot-instructions.md ..."
}

Output:

{
  "status": "success",
  "message": "Found 1 new learning(s), skipped 0 duplicate(s).",
  "new_learnings": [...],
  "duplicates_skipped": [],
  "markdown_to_append": "| 2026-01-18 | MCP Config Format | Use 'servers' key not 'mcpServers' | VS Code MCP expects this exact format |",
  "target_section": "Session Learnings Log"
}

check_learning_exists

Check if a specific learning already exists.

Input:

• topic (string): Learning topic
• decision (string): The decision/learning
• current_instructions (string): Current copilot-instructions.md content

Output:

{
  "exists": true,
  "similar_to": "Existing Topic Name",
  "similarity_score": 0.85,
  "reason": "High similarity to existing entry"
}

How It Works

1. Copilot extracts learnings from your conversation using its built-in LLM
2. MCP tool deduplicates using Jaccard similarity (threshold: 0.6)
3. Tool formats new learnings as markdown table rows
4. Copilot appends the formatted rows to your copilot-instructions.md

This approach means:

• ✅ No API keys needed
• ✅ No external LLM calls
• ✅ Fast deduplication
• ✅ Works offline (except for Copilot itself)

License

MIT