GitHub Copilot memory: instructions, repo facts, and the gaps

How Copilot handles memory today

What GitHub Copilot forgets

Copilot memory is a strong default inside GitHub. It is not portable, and it is not the whole story.

• Everything outside GitHub. Work you do in a terminal agent, a chat product, or a local editor does not feed Copilot memory.
• The reasoning in a session that never produced a commit or a comment.
• Cross-repository context. Memory is tied to a repository, so two services that talk to each other hold separate facts.
• Anything you decide with another assistant, which stays in that assistant.

The fix with MemoryRouter

VS Code agent mode supports remote MCP servers. Connect MemoryRouter there and the same vault you use with Claude Code and Codex becomes available to Copilot.

1. 1

   Add the server in VS Code

   Use MCP server configuration in VS Code, pointing at the MemoryRouter remote endpoint. Remote HTTP servers with OAuth are supported; the exact file location depends on your VS Code version, and the menu path is command palette, then MCP: Add Server.

   mcp.json (VS Code)

   {
     "servers": {
       "memoryrouter": {
         "type": "http",
         "url": "https://mcp.memoryrouter.ai/mcp"
       }
     }
   }

2. 2

   Sign in when prompted

   VS Code opens the OAuth flow. Approve the connection and pick the vault.

3. 3

   Keep copilot-instructions.md lean

   Instructions stay the place for how your team works. Let the vault hold decisions and session context so the instruction file does not grow into an archive.

Related

GitHub Copilot memory FAQ