Repo Memory

Name: Repo Memory
Author: yubinkim444

@yubinkim444

代码仓库

关于 Repo Memory

A different approach from typical persistent-memory MCPs. Instead of a local

#memory #knowledge-base #ai-agents #claude-code #cursor #cline

基本信息

分类

记忆与知识

传输方式

stdio

发布者

yubinkim444

提交者

YooBin Kim

配置

使用下面的配置,将此服务器添加到你的 MCP 客户端。

{
  "mcpServers": {
    "repo-memory": {
      "command": "uvx",
      "args": [
        "repo-memory-mcp",
        "--repo",
        "/path/to/your/repo"
      ]
    }
  }
}

工具

未检测到工具

工具是从 README 中自动提取的。维护者可以在 ## Tools 标题下列出工具,即可填充这部分内容。

概览

What is Repo Memory?

Repo Memory is a shared, git-tracked working memory for AI agents that share a codebase. It stores structured facts, decisions, and gotchas in plain files under a .ai-memory/ directory, synchronized via git. It is designed for developers using AI coding assistants such as Claude, Cursor, or Cline.

How to use Repo Memory?

Install with pip install repo-memory or run via uvx. Initialize memory in your repo with repo-memory init, then use commands such as repo-memory add-fact, add-decision, or add-gotcha. For AI agent integration, add an MCP server entry in your client config pointing to repo-memory-mcp with the repo path.

Key features of Repo Memory

Git-tracked, append-only facts with evidence
No database, no SaaS, no daemon; works offline
Structured facts, decisions, and gotchas as plain files
Per-repo scope – memory stays local to the codebase
MCP server with 5 tools for read/write access
Automatic discovery hint in CLAUDE.md or AGENTS.md

Use cases of Repo Memory

Sharing known facts across AI agent sessions on the same repository
Recording architecture decisions and non-obvious conventions
Avoiding repeated grep or re-discovery of codebase structure
Multi-user or multi-tool collaboration on a shared codebase

FAQ from Repo Memory

How is Repo Memory different from `CLAUDE.md` or `.cursorrules`?

CLAUDE.md is for rules that a human writes; Repo Memory is for facts that an agent verifies – locations, behaviors, and decisions discovered during work.

What are the runtime requirements?

Python 3.10 or later. The only dependency is the official MCP SDK.

Where does Repo Memory store data?

In a .ai-memory/ directory inside the repository. The data is plain files (facts.jsonl, decisions/, gotchas.md) and is synchronized via normal git commits and pushes.

What transport does the MCP server use?

It uses stdio transport. The server is launched via repo-memory-mcp with a --repo argument or the REPO_MEMORY_ROOT environment variable.

How do multiple agents avoid conflicting facts?

Facts are append-only; stale entries remain. Readers check verified_at to re-verify if needed. Suspicious facts are filtered through normal code review via git.