MARM-Systems MCP — 使用场景, 安装 & 实时演示

为什么要用

核心特性

18 tools covering sessions, notebooks, memory recall, context bridging
Semantic search via vector embeddings; auto-classifies content (code/project/book/general)
Cross-client memory: Claude, Qwen, Gemini, any MCP client shares the same store
Three transports: STDIO, HTTP (localhost:8001), WebSocket beta
Local SQLite storage — nothing leaves your machine by default

实时演示

实际使用效果

marm-systems.replay ▶ 就绪

0/0

安装

选择你的客户端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "marm-systems": {
      "command": "uvx",
      "args": [
        "MARM-Systems"
      ],
      "_inferred": true
    }
  }
}

打开 Claude Desktop → Settings → Developer → Edit Config。保存后重启应用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "marm-systems": {
      "command": "uvx",
      "args": [
        "MARM-Systems"
      ],
      "_inferred": true
    }
  }
}

Cursor 使用与 Claude Desktop 相同的 mcpServers 格式。项目级配置优先于全局。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "marm-systems": {
      "command": "uvx",
      "args": [
        "MARM-Systems"
      ],
      "_inferred": true
    }
  }
}

点击 Cline 侧栏中的 MCP Servers 图标，然后选 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "marm-systems": {
      "command": "uvx",
      "args": [
        "MARM-Systems"
      ],
      "_inferred": true
    }
  }
}

格式与 Claude Desktop 相同。重启 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "marm-systems",
      "command": "uvx",
      "args": [
        "MARM-Systems"
      ]
    }
  ]
}

Continue 使用服务器对象数组，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "marm-systems": {
      "command": {
        "path": "uvx",
        "args": [
          "MARM-Systems"
        ]
      }
    }
  }
}

加入 context_servers。Zed 保存后热重载。

claude mcp add marm-systems -- uvx MARM-Systems

一行命令搞定。用 claude mcp list 验证，claude mcp remove 卸载。

使用场景

实战用法： MARM-Systems

Build persistent memory across sessions on a long project

👤 Anyone working a multi-week effort with an AI ⏱ ~15 min beginner

何时使用： You keep re-explaining the same project background every session.

前置条件

MARM-Systems installed and running — docker pull lyellr88/marm-mcp-server && docker run -d -p 8001:8001 lyellr88/marm-mcp-server

步骤

Start a session tagged to the project

marm_start with project tag 'dataplatform-migration'. Log that we're migrating from Redshift to Snowflake, deadline end of Q2.✓ 已复制

→ Session started; initial entry saved
Drop context as you work

marm_contextual_log: 'Decided to use Fivetran for CDC replication, evaluated Airbyte but config overhead too high.'✓ 已复制

→ Auto-classified and stored
Next session, recall

marm_smart_recall 'what did we decide about CDC tooling?'✓ 已复制

→ Relevant past decision surfaces

结果： Session N+1 starts with all of session 1-N context accessible by query, not re-type.

注意事项

Dumping every chat into memory clogs recall — Use marm_contextual_log for decisions and milestones, not every exchange
Vector search misses on jargon — Tag entries explicitly with project names for keyword fallback

搭配使用： drift

Share memory across a team using multiple AI assistants

👤 Teams where different members use Claude / Qwen / Gemini ⏱ ~45 min advanced

何时使用： Knowledge trapped in one person's AI shouldn't be.

步骤

Run MARM as a shared HTTP server

Deploy MARM-Systems on a team server; each teammate points their MCP client at http://marm.team.internal:8001/mcp.✓ 已复制

→ All clients connect
Log shared context

marm_contextual_log: 'Database backup runbook is at /runbooks/db-backup.md; last updated 2026-04-10.'✓ 已复制

→ Everyone can recall it
Recall from any client

From a teammate's Qwen session: marm_smart_recall 'how do I restore the DB'?✓ 已复制

→ Same answer surfaces

结果： Cross-AI shared institutional memory.

注意事项

Shared memory needs real OAuth — Don't rely on the dev hardcoded credentials — wire to your IdP before prod use
Sensitive data leaks into shared recall — Use project tags and scope queries by tag

Use notebooks as AI-queryable scratchpads

👤 Learners, researchers ⏱ ~5 min beginner

何时使用： You want a topic-scoped note area the AI can search later.

步骤

Create a notebook

marm_notebook_add name='rust-ownership' with entry: 'Move vs borrow: move transfers ownership, borrow lets you peek.'✓ 已复制

→ Notebook created
Add as you learn

marm_notebook_add 'rust-ownership': 'Mutable borrow is exclusive; only one at a time.'✓ 已复制

→ Entry appended
Recall later

marm_notebook_show 'rust-ownership'✓ 已复制

→ Full notebook

结果： A queryable learning journal.

组合

与其他 MCP 搭配，撬动十倍杠杆

marm-systems + drift

drift remembers code conventions; MARM remembers project decisions

Recall drift conventions AND MARM decisions for project 'dataplatform-migration' and summarize where they interact.✓ 已复制

marm-systems + claude-code-organizer

Move short-lived memories out of CC into MARM for long-term recall

Organizer listed my 12 biggest memories — move the 5 project-specific ones into MARM notebooks.✓ 已复制

工具

此 MCP 暴露的能力

工具	输入参数	何时调用	成本
marm_start	project?, tags?	Start of any working session	free (local)
marm_refresh	session_id?	Reload context mid-session	free
marm_smart_recall	query: str, top_k?: int	Semantic search past entries	free (local embeddings)
marm_contextual_log	text: str, tags?: str[]	Persist a decision/milestone/fact	free
marm_log_session	session_id	Review a past session	free
marm_notebook_add	name: str, entry: str	Topic-scoped notes	free
marm_notebook_use	name: str	Pin a topic for the session	free
marm_notebook_show	name: str	Read a notebook	free
marm_summary	scope: session\|notebook\|tag, id	Condense long history	free (local)
marm_context_bridge	from_session, to_session	Cross-project reasoning	free

成本与限制

运行它的成本

API 配额: None — local-only
每次调用 Token 数: Recall responses 500-2000 tokens typically
费用: Free, open source
提示: Rotate or archive old sessions periodically — SQLite handles scale but recall quality degrades with noise.

安全

权限、密钥、影响范围

凭据存储： Local SQLite; for shared HTTP mode use env-based OAuth secrets

数据出站： None by default (local storage); if you run shared, only within your network

切勿授予： Exposing port 8001 to the public internet without an auth layer

Dev mode ships hardcoded OAuth credentials — change before any shared deployment.
Memory can accumulate sensitive data; back up and encrypt the SQLite file.

故障排查

常见错误与修复

Cannot connect to localhost:8001

Container not running. docker ps to check; docker start marm-mcp-server.

验证： curl http://localhost:8001/docs

Embeddings model download fails

First run downloads the model; needs outbound net. After that, offline works.

验证： Check docker logs for HuggingFace download

Recall returns unrelated entries

Too few entries for good embeddings; add more, or pre-filter by tag.

验证： marm_smart_recall with tag filter

替代方案

MARM-Systems 对比其他方案

替代方案	何时用它替代	权衡
drift	You want code-specific memory (conventions, decisions)	Code-focused; less general
Letta (MemGPT)	You want a research-grade memory agent, not just a server	Heavier to run; opinionated architecture
CC native memories	Claude Code only; no cross-client need	No semantic search; bloats context

MARM-Systems

为什么要用

核心特性

实时演示

实际使用效果

安装

选择你的客户端

使用场景

实战用法： MARM-Systems

Build persistent memory across sessions on a long project

前置条件

步骤

注意事项

Share memory across a team using multiple AI assistants

步骤

注意事项

Use notebooks as AI-queryable scratchpads

步骤

组合

与其他 MCP 搭配，撬动十倍杠杆

工具

此 MCP 暴露的能力

成本与限制

运行它的成本

安全

权限、密钥、影响范围

故障排查

常见错误与修复

替代方案

MARM-Systems 对比其他方案

更多

资源