A different approach from typical persistent-memory MCPs. Instead of a local
SQLite + embeddings store, the memory lives as plain files in a .ai-memory/
directory you commit to your repo (facts.jsonl, decisions/*.md, gotchas.md).
Git is the sync layer — what one Claude/Cursor/Cline learns about a repo, the
next session (or a teammate's agent) picks up automatically.
5 MCP tools: get_repo_memory, add_fact, list_facts, add_decision, add_gotcha.
Pure Python, depends only on the official mcp SDK. Zero infra. Per-repo scope.
Works offline. Cross-tool by file convention.
Server Config
{
"mcpServers": {
"repo-memory": {
"command": "uvx",
"args": [
"repo-memory-mcp",
"--repo",
"/path/to/your/repo"
]
}
}
}