Claude Code 的生产级向量内存 MCP 服务器。语义搜索、逐项目隔离、规则执行、团队共享（通过 git）。
概览
每个项目都有自己的大脑——决策、规则、架构笔记、冲刺目标——这些都存储在本地矢量数据库中，每次新会话开始时自动加载。 Claude 会在两次会话之间忘记一切。你最终会反复解释同样的内容，规则会被遗漏，过去的决策在上下文窗口满时丢失，切换机器意味着从头开始。这个 MCP 给了 Claude 真正的记忆——它能经受重启、上下文溢出和团队交接的考验。 把它当作每个项目的独立大脑。 当你在做某件事时，你会做出决策（“我们选择了PostgreSQL”），设定规则（“在提交前一定要运行测试”），并规划冲刺。通常，这些都是，或者分散在Claude每次都得重新阅读的.md文件中。
概述（AI 翻译）
Claude Code 的持久、可搜索项目内存。 每个项目都有自己的大脑——决策、规则、架构笔记、冲刺目标——这些都存储在本地矢量数据库中，每次开始新会话时自动加载。
这解决了什么问题？ Claude 在会话间会忘记一切。你会反复重复同样的内容，规则会被遗漏，过去的决策在上下文窗口满时丢失，切换机器意味着从头开始。这个 MCP 给了 Claude 真正的记忆——能经受重启、上下文溢出和团队交接的考验。
用通俗英语
这个项目到底做了什么？
它为每个项目赋予克劳德永久记忆。 把它当作每个项目的独立大脑。
当你在做某件事时，你会做出决策（“我们选择了PostgreSQL”），设定规则（“总是在提交前运行测试”），并规划冲刺。通常，当新会话开始时，这些都会被忘记，或者分散在需要Claude每次都重新阅读的.md文件中。
有了这个MCP：
- 当你做出决定时 → Claude 会自动存储
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
SMG Claude Memory MCP
Persistent, searchable project memory for Claude Code. Each project has its own brain — decisions, rules, architecture notes, sprint goals — all stored locally in a vector database and automatically loaded every time you start a new session.
What problem does this solve? Claude forgets everything between sessions. You end up re-explaining the same things over and over, rules get missed, past decisions get lost when your context window fills up, and switching machines means starting from scratch. This MCP gives Claude a real memory — one that survives restarts, context overflows, and team handoffs.
In Plain English
What does this project actually do?
It gives Claude permanent memory per project. Think of it as a separate brain for each of your projects.
When you’re working on something, you make decisions (“we chose PostgreSQL”), set rules (“always run tests before commits”), and plan sprints. Normally, all of that is forgotten when a new session starts, or gets scattered across .md files that Claude has to re-read every time.
With this MCP:
- When you make a decision → Claude stores it automatically
- When you ask about something → Claude searches memory first, then answers
- When you start a new session → All the important stuff loads automatically
- When you set a rule (“always do X”) → Claude won’t forget it again
Three things users should understand
-
Semantic search — Ask “what database did we pick?” and it finds the PostgreSQL decision, even if you never typed “PostgreSQL” in your query. It searches by meaning, not keywords.
-
Project isolation — Every project has its own separate memory database. Memories from
project-Anever leak intoproject-B. -
Team collaboration — You can move a project’s memory into the project folder, commit it to git, and your teammates get the same memory after
git pull.
The technical one-liner
“An MCP server for Claude Code that stores per-project memory in local DuckDB with vector embeddings. It remembers decisions, rules, and sprint notes, retrieves them via semantic search, and auto-loads full context at session start — so Claude never forgets anything.”
What you’ll feel day-to-day
- You won’t need to re-explain your project’s decisions every session
- Rules you’ve set stick — Claude can’t skip them (enforced by hooks)
- Context window overflows don’t lose important info (it’s in the DB)
- Switching machines?
git pull+ one command, and Claude picks up exactly where you left off - Onboarding a teammate’s Claude? Same thing — they get your whole project memory instantly
FAQ — Common Questions, Clear Answers
Q: When I run make_portable, does it move ALL projects’ memory to git, or just the one I’m working on?
Just the one project. Other projects are never touched.
Each project has its own separate .duckdb file:
~/.memory-mcp/安装
This server does not publish a one-line install command.
Open the repository installation guide 新手？先看 mcpskill.org 教程 →配置
{
"mcpServers": {
"memory": {
"command": "uv",
"args": ["run", "--directory", "/path/to/smg-claude-memory-mcp", "memory-mcp"]
}
}
}