Dasix/mcp-shared-memory

MCP Shared Memory Server

An experimental Model Context Protocol (MCP) server that provides shared memory functionality across multiple Claude Code instances through a centralized Git-backed storage system.

Overview

This MCP server enables different Claude Code instances to share and manage memory in a consistent way by connecting to a single centralized server. It uses a Git repository as the backing store for shared memory, allowing for version control, collaboration, and persistence.

Features

• Centralized Memory Storage: Single server instance manages memory for multiple clients
• Git-Backed Persistence: All shared memory stored in a Git repository
• Project-Specific Configuration: Each client identified by unique auth token
• Template System: Markdown files support conditional logic based on project flags
• Network-First Design: Optimized for HTTP/SSE transport for remote access
• Flexible Architecture: Supports both centralized and local deployment models

Installation

npm install mcp-shared-memory

Configuration

The server is configured through environment variables:

• GIT_REPO_URL - URL of the Git repository for shared memory storage
• GIT_REPO_PATH - Local path to the Git repository clone
• PROJECT_META_PATH - Directory for storing project metadata
• SERVER_PORT - HTTP/SSE server port (default: 3000)
• SERVER_HOST - Host to bind to (default: 0.0.0.0)

Usage

Starting the Server

# Development mode with hot reload
npm run dev:server

# Production mode
npm run start:server

Claude Code Client Configuration

Configure each Claude Code instance to connect to the server:

{
  "mcp-servers": {
    "shared-memory": {
      "url": "https://shared-memory.internal.network:3000",
      "transport": "sse",
      "auth": {
        "type": "token",
        "token": "${SHARED_MEMORY_TOKEN}"
      }
    }
  }
}

Note: Each client must have a unique token - this is how the server identifies different projects/clients.

Available Tools

• get_available_flags - Returns available flags and their check instructions
• set_project_flags - Sets flags for the client's project
• get_instruction - Reads an instruction file with templates resolved
• get_instruction_template - Reads the raw template file
• set_instruction_template - Writes a template file to shared memory
• list_instructions - Lists available instruction templates

Template System

The server supports a Handlebars-inspired template syntax in markdown files:

# Deployment Instructions

{{#if argocd}}
## ArgoCD Deployment
Follow these steps for ArgoCD deployment...
{{/if}}

{{#if kubernetes}}
## Kubernetes Deployment
Apply the manifests with kubectl...
{{/if}}

{{> partials/common-footer}}

Features

• Conditionals: {{#if flag}}...{{/if}}, {{#unless flag}}...{{/unless}}
• Partials: {{> path/to/partial}}
• Comments: {{! This is a comment }}
• Escaping: \{{ to output literal braces

Development

# Install dependencies
npm install

# Run tests
npm test

# Run linting
npm run lint

# Format code
npm run format

Architecture

The server consists of several key components:

1. MCP Server Core - Handles protocol communication and client sessions
2. Git Repository Manager - Manages Git operations and file locking
3. Template Engine - Processes markdown templates with conditional logic
4. Project Metadata Store - Stores client-specific configuration
5. Tool Handlers - Implements MCP tool endpoints

Security

• Each client authenticated via unique API token
• Tokens serve as primary keys for project metadata
• File path sanitization prevents directory traversal
• Template validation before saving

License

ISC

Contributing

This is an experimental project. Contributions and feedback are welcome!