Project Memory MCP

A Model Context Protocol (MCP) server that provides AI-driven project memory management through structured prompts. This server acts as a pure prompt provider - it never touches your files directly. Instead, it returns instructions for Claude to execute using its standard tools.

What is This?

project-memory-mcp helps Claude manage your project context by:

• Parsing tasks from specs and implementation plans

• Reviewing code before commits

• Syncing project memory with commit history

• Maintaining project documentation (architecture, conventions, commands)

All operations are performed by Claude using its Read, Write, Edit, and Bash tools after getting your approval.

Key Features

✅ Pure prompt provider - No file access, only returns instructions ✅ Interface-agnostic - Works with Claude Desktop, Claude Code CLI, or custom clients ✅ No API costs - Uses your existing Claude subscription ✅ User approval required - Claude asks before making any changes ✅ Project-specific prompts - Customized during initialization for your stack ✅ 200-line limit - Prevents context bloat

Installation

Install globally from GitHub:

Setup

Configure MCP in Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

Configure MCP in Claude Code CLI

Add to your user config at ~/.claude.json:

This will make the mcp available in all your project.

Restart Claude

Restart Claude Desktop or Claude Code CLI to load the MCP server.

Usage

Initialize Project Memory

In your project directory, ask Claude:

Claude will:

1. Create .project-memory/ folder structure

2. Analyze your project (language, frameworks, structure, conventions)

3. Generate and customize project-specific prompts (base.md, parse-tasks.md, review.md, sync.md)

4. Populate initial documentation files (architecture.md, conventions.md, useful-commands.md) with detected project info

5. Add session-start checklist to CLAUDE.md (ensures project memory is loaded every session)

Important: Each prompt file is limited to ≤ 200 lines to prevent context bloat.

Parse Tasks from Specs

Add a spec file to .project-memory/specs/feature-name.md, then ask Claude:

Claude will:

1. Read the spec file

2. Extract tasks with IDs, descriptions, acceptance criteria, dependencies

3. Show you the parsed tasks

4. After approval, add them to .project-memory/tasks/tasks-active.json

Review Code Before Committing

Before committing, ask Claude:

Claude will:

1. Get git diff and git diff --cached

2. Read current tasks and architecture

3. Analyze code for issues

4. Propose task status updates

5. After approval, update project memory

Sync After Commits

After committing, ask Claude:

Claude will:

1. Get recent commit history

2. Determine completed tasks

3. Update commit log (last 20 commits)

4. Update architecture if changed

5. After approval, apply updates

Organize Existing CLAUDE.md

For existing projects with verbose CLAUDE.md files, ask Claude:

Claude will:

1. Read and analyze your CLAUDE.md

2. Identify sections: architecture, conventions, commands, tasks, specs

3. Show migration plan with line numbers

4. After approval, migrate content to .project-memory/ files

5. Replace verbose sections with minimal references in CLAUDE.md

Example migration:

• Architecture (75 lines) → .project-memory/architecture.md

• Conventions (60 lines) → .project-memory/conventions.md

• Commands (30 lines) → .project-memory/useful-commands.md

• Tasks → .project-memory/tasks/tasks-active.json

• Specs → .project-memory/specs/*.md

Result: CLAUDE.md stays clean with just references, detailed content lives in organized files.

Create Spec from Requirements

When you have a feature idea or requirements, ask Claude:

Claude will:

1. Clarify ambiguous requirements

2. Validate against existing codebase

3. Consider security, edge cases, and testing

4. Write spec to .project-memory/specs/

Refresh Prompts

When prompt templates have been updated, ask Claude:

Claude will:

1. Backup existing prompts

2. Compare with new templates

3. Preserve your customizations

4. Merge improvements into updated prompts

Proactive Prompting

After initialization, Claude will automatically:

• Session start: Read project memory files (tasks, architecture, conventions) before working

• When you provide a spec: Offer to parse tasks

• Before commits: Offer to review changes

• After commits: Offer to sync project memory

Project Structure

After initialization, your project will have:

Task Schema

Tasks follow this structure:

MCP Tools

The server exposes 7 tools (all return prompts only):

init

Initialize project memory system. Run once per project.

parse-tasks

Parse tasks from spec files or implementation plans. Checks existing tasks AND codebase to avoid duplicates.

review

Review uncommitted code changes against project context. Uses extended thinking for thorough analysis.

sync

Sync project memory with recent commits. Ensures CLAUDE.md never has outdated references.

organize

Organize existing CLAUDE.md into project-memory structure. Migrates architecture, conventions, commands, tasks, and specs from verbose CLAUDE.md to organized files.

create-spec

Create detailed specification from user requirements. Clarifies ambiguity, validates against codebase, considers security and edge cases.

refresh-prompts

Update project-specific prompts with latest template improvements while preserving customizations. Backs up existing prompts before changes.

How It Works

Architecture Principles

1. MCP is a pure prompt provider

   • Only returns text instructions

   • Never reads/writes project files

   • Never executes git commands

2. Claude does all the work

   • Uses standard tools (Read, Write, Edit, Bash)

   • User sees all operations

   • Requires approval via AskUserQuestion

3. Project-specific customization

   • Prompts tailored to your tech stack

   • Language-specific guidelines

   • Framework conventions

4. 200-line limit per prompt

   • Prevents context bloat

   • Keeps prompts focused

   • Enforced during init

Development

Testing

Run the test suite:

Tests cover:

• Prompt length validation (200-line limit)

• Prompt composition

• Edge cases

Contributing

Contributions welcome! Please:

1. Fork the repository

2. Create a feature branch

3. Make your changes

4. Add tests

5. Submit a pull request

License

MIT

Links

• Spec - Detailed specification

• MCP Documentation

• GitHub Repository

Support

For issues, questions, or feedback:

• GitHub Issues

Note: This MCP server requires Claude Desktop, Claude Code CLI, or another MCP-compatible client to function.