Project Memory MCP

# Claude Assistant Configuration for project-memory-mcp ## Project Overview This is an MCP (Model Context Protocol) server that acts as a **pure prompt provider** for AI-driven project memory management. It never touches files directly - only returns instructions for Claude to execute. ## Key Architecture Rules 1. **Pure Prompt Provider** - MCP server ONLY returns prompt text - NEVER reads/writes project files (except .project-memory/prompts/) - NEVER executes git commands - NEVER parses JSON or processes data 2. **Prompt Size Limit - CRITICAL** - Each .md prompt file MUST be ≤ 200 lines - This prevents context bloat - Validate during init and warn if exceeded 3. **Project-Specific Prompts** - Only ONE hardcoded prompt: init - All other prompts (parse-tasks, review, sync) are created during init - Prompts are customized per project based on language/framework ## Development Workflow ### Before any task, check complexity: - Simple: Direct implementation - Complex: Break down into subtasks, iterate with testing ### For each implementation: 1. **Research**: Check spec.md, understand requirements 2. **Plan**: Break down if complex 3. **Implement**: Write focused code 4. **Test**: Run build and tests 5. **Verify**: Ensure TypeScript compiles, tests pass ### Definition of Done: - ✅ `npm run build` succeeds - ✅ All TypeScript type checks pass - ✅ `npm test` passes - ✅ No new warnings or errors ## Tech Stack - **Language**: TypeScript (ES2022, Node16 modules) - **MCP SDK**: @modelcontextprotocol/sdk - **Testing**: Vitest - **Build**: tsc ## File Structure ``` src/ ├── index.ts # Main MCP server ├── prompts/ │ └── init-prompt.ts # ONLY hardcoded prompt ├── schemas/ │ └── task-schema.ts # Task JSON schema └── utils/ └── prompt-loader.ts # Prompt composition & validation ``` ## Key Commands ```bash npm run build # Compile TypeScript npm run watch # Watch mode npm test # Run tests npm run dev # Dev mode with tsx ``` ## Important Constraints 1. **NO file I/O** in main server logic (except reading prompts) 2. **NO git operations** in server code 3. **NO data processing** - just return prompts 4. **200-line limit** for all prompt files 5. **User approval required** - all prompts must instruct Claude to use AskUserQuestion ## Security Rules - ✅ No hardcoded credentials - ✅ No sensitive data in prompts - ✅ No file path traversal vulnerabilities - ✅ Validate prompt file paths ## Testing Requirements - Test prompt length validation - Test prompt composition - Test edge cases (empty files, missing prompts) - NO need to test actual file operations (Claude does those) ## Documentation Rule **NO MASSIVE .MD FILES. EVER.** - Keep documentation in code (docstrings, comments) - README.md for overview - spec.md for detailed specification - CLAUDE.md for project instructions (this file) - NO other markdown files unless absolutely necessary --- When working on this project, always remember: **We build prompts, not file processors.**