MCP Standards

# MCP Standards v2: Simple Personal Memory **Philosophy**: Make Claude remember MY preferences automatically. Zero config, zero manual steps. ## What We're Building A simple MCP server that: 1. ✅ Auto-detects corrections ("use uv not pip") 2. ✅ Stores in AgentDB (semantic vector memory) 3. ✅ Retrieves before responding (automatic context) 4. ✅ Parses config files → minimal CLAUDE.md 5. ✅ ONE skill to enable it all ## Architecture (Simple) ``` ┌─────────────────────────────────────────┐ │ Claude Desktop/Code │ │ │ │ ┌───────────────────────────────────┐ │ │ │ MCP: personal-memory │ │ │ │ │ │ │ │ ┌─────────────────────────────┐ │ │ │ │ │ AgentDB Client │ │ │ │ │ │ (npx agentdb) │ │ │ │ │ │ - Vector storage │ │ │ │ │ │ - Semantic search │ │ │ │ │ │ - <10ms startup │ │ │ │ │ └─────────────────────────────┘ │ │ │ │ │ │ │ │ ┌─────────────────────────────┐ │ │ │ │ │ Auto Memory Hook │ │ │ │ │ │ - Detects corrections │ │ │ │ │ │ - Auto-stores patterns │ │ │ │ │ │ - Zero manual calls │ │ │ │ │ └─────────────────────────────┘ │ │ │ │ │ │ │ │ ┌─────────────────────────────┐ │ │ │ │ │ Config Parser (existing) │ │ │ │ │ │ - Reads .editorconfig │ │ │ │ │ │ - Minimal CLAUDE.md │ │ │ │ │ └─────────────────────────────┘ │ │ │ └───────────────────────────────────┘ │ └─────────────────────────────────────────┘ ``` ## Implementation ### Phase 1: AgentDB Integration (Days 1-3) - Install AgentDB via npm - Create simple Python client wrapper - Test store/retrieve operations - Performance benchmark (<10ms startup, <1ms search) ### Phase 2: Auto-Detection (Days 4-6) - Hook into MCP tool execution - Detect correction patterns - Auto-store in AgentDB (no manual calls) - Test detection accuracy ### Phase 3: Skill Creation (Day 7) - Create `remember-preferences.md` skill - Auto-query AgentDB before responding - Test memory retrieval in conversations ### Phase 4: Polish (Days 8-10) - Simplify codebase (remove complex layers) - Optimize for minimal CLAUDE.md output - End-to-end testing - Documentation ## What We're NOT Building ❌ Custom vector database (use AgentDB) ❌ Complex confidence scoring (trust AgentDB) ❌ Manual promotion system (auto-detect everything) ❌ Event-driven architecture (overkill) ❌ Hybrid memory tiers (AgentDB handles it) ## Success Criteria 1. **Zero Manual Steps**: User corrects Claude → automatically remembered 2. **Fast**: <1ms to retrieve preferences 3. **Accurate**: Semantic matching (>80% recall) 4. **Minimal CLAUDE.md**: <1000 tokens, not 23K 5. **Works Cross-Project**: Personal knowledge graph ## Files to Create (New) ``` /src/mcp_standards/agentdb_client.py # AgentDB wrapper (100 LOC) /src/mcp_standards/hooks/auto_memory.py # Correction detector (150 LOC) /.claude/skills/remember-preferences.md # The skill (50 LOC) /docs/SIMPLE_V2_PLAN.md # This file ``` ## Files to Keep (Simplified) ``` /src/mcp_standards/standards.py # Config parser ✅ /src/mcp_standards/server.py # MCP server ✅ /src/mcp_standards/export.py # Markdown export ✅ ``` ## Files to Remove (Complexity) ``` /src/intelligence/context/ # Over-engineered /src/intelligence/reasoning/ # Use AgentDB RL /src/mcp_standards/intelligence/temporal_graph.py /docs/architecture/v2-*.md # Old complex plans ``` ## Timeline - **Week 1**: Core system working (auto-detect + store) - **Week 2**: Polish + test + document - **Week 3**: Use it myself, iterate ## Inspiration - **Gunnar's approach**: Simple side projects that solve personal problems - **Context engineering**: Minimal CLAUDE.md, dynamic loading - **AgentDB**: 150x faster semantic search, built-in learning ## Next Steps 1. ✅ Create this plan 2. Install AgentDB: `npm install agentdb` 3. Build AgentDB client wrapper 4. Create correction detection hook 5. Build remember-preferences skill 6. Test end-to-end