MCP Standards

# Quick Reference: Memory Systems Research **Last Updated**: 2025-10-20 --- ## 🎯 One-Minute Summary Research completed on 4 memory systems for v2: 1. **AgentDB**: 2-3ms retrieval, SQLite-based, 20 MCP tools 2. **ReasoningBank**: +34% effectiveness, Bayesian learning, self-improving 3. **Claude Skills**: Progressive disclosure, automatic generation, composable 4. **Context Engineering**: 20K+ token savings, 5-layer architecture **Recommendation**: Implement all four in 3 phases over 12 weeks. --- ## 📊 Key Numbers | Technology | Primary Metric | Benchmark | |------------|---------------|-----------| | AgentDB | Retrieval Speed | 2-3ms @ 100K patterns | | ReasoningBank | Effectiveness Gain | +34.2% | | Skills | Token Overhead | Few dozen per skill | | Context Eng | Token Savings | 20,000+ tokens | --- ## 🚀 Quick Start Commands ### AgentDB Setup ```bash npm install agentdb sqlite3 sqlite-vec npx agentdb benchmark --quick ``` ### ReasoningBank Import ```bash npm install agentic-flow ``` ```javascript import * as reasoningbank from 'agentic-flow/reasoningbank'; ``` ### Context Optimization ```bash # Minimize CLAUDE.md to <5K tokens # Create /prime commands: /prime-bug # Bug investigation /prime-feature # Feature development /prime-refactor # Refactoring tasks /prime-research # Research and analysis ``` ### Skills Framework ```yaml # SKILL.md format --- name: unique-skill-identifier description: Purpose of skill --- # Skill Name [Instructions here] ``` --- ## 📋 Implementation Checklist ### Phase 1: Foundation (Weeks 1-3) ✅ HIGH PRIORITY - [ ] Install AgentDB + dependencies - [ ] Set up SQLite + sqlite-vec - [ ] Implement 20 MCP memory tools - [ ] Minimize CLAUDE.md to <5K tokens - [ ] Create 4 /prime commands - [ ] Measure baseline performance - [ ] Achieve 20K+ token reduction ### Phase 2: Learning (Weeks 4-7) ✅ HIGH PRIORITY - [ ] Import agentic-flow/reasoningbank - [ ] Build STORE→EMBED→QUERY→RANK→LEARN pipeline - [ ] Configure 6 thinking modes - [ ] Set up Bayesian confidence updates - [ ] Implement 5-layer context architecture - [ ] Create 50-100 seed patterns - [ ] Test pattern learning effectiveness ### Phase 3: Skills (Weeks 8-12) ⚠️ MEDIUM PRIORITY - [ ] Adopt SKILL.md format - [ ] Build skill-creator - [ ] Implement progressive disclosure - [ ] Enable skill composition - [ ] Test cross-platform compatibility - [ ] Create skill validation framework --- ## ⚡ Performance Targets ### Retrieval Performance - **Target**: <5ms @ 100K patterns - **Measure**: P50, P95, P99 latency - **Baseline**: AgentDB 2-3ms ### Learning Effectiveness - **Target**: +30% task success rate - **Measure**: Before/after comparison - **Baseline**: ReasoningBank +34.2% ### Token Efficiency - **Target**: 20,000+ reduction - **Measure**: Average tokens per task - **Baseline**: Context engineering proven ### Pattern Quality - **Target**: 84% confidence @ 20 uses - **Measure**: Convergence rate - **Baseline**: Bayesian updates validated --- ## 🔧 Technology Stack ### Core Dependencies ```json { "agentdb": "latest", "agentic-flow": "latest", "sqlite3": "^5.1.0", "sqlite-vec": "^0.1.0", "claude-flow": "@alpha" } ``` ### MCP Servers - **Required**: `claude-flow` (orchestration) - **Recommended**: `agentdb` (20 memory tools) - **Optional**: `ruv-swarm`, `flow-nexus` ### System Requirements - Node.js 18+ - 2GB RAM (4GB recommended) - 1GB disk for patterns - SQLite 3.x with vec extension --- ## ⚠️ Top Risks ### 1. Performance at Scale (>1M vectors) - **Risk**: sqlite-vec brute-force degrades - **Mitigation**: Limit <1M, monitor ANN development ### 2. Pattern Quality - **Risk**: Poor patterns reduce effectiveness - **Mitigation**: Seed library, validation framework ### 3. Context Complexity - **Risk**: 5-layer architecture hard to debug - **Mitigation**: Comprehensive logging, gradual rollout --- ## 📚 5-Minute Deep Dive Sections ### AgentDB Architecture - **Foundation**: SQLite (transactional) + DuckDB (analytics) - **Vector Search**: sqlite-vec extension, cosine similarity - **Graph Search**: HNSW multi-level indexing (when available) - **Performance**: SIMD acceleration (AVX/NEON) - **Scale**: Thousands to hundreds of thousands of vectors ### ReasoningBank Pipeline ``` 1. STORE → Save experience as pattern (SQLite) 2. EMBED → Convert to 1024-dim vector (SHA-512) 3. QUERY → Semantic search (cosine, 2-3ms) 4. RANK → Multi-factor score (semantic, confidence, recency, diversity) 5. LEARN → Bayesian update (+20% success, -15% failure) ``` ### Six Thinking Modes 1. **Convergent**: Focused, analytical 2. **Divergent**: Creative, exploratory 3. **Lateral**: Indirect, innovative 4. **Systems**: Holistic, interconnected 5. **Critical**: Evaluative, questioning 6. **Adaptive**: Flexible, context-responsive ### 5-Layer Context Architecture 1. **Meta-Context**: Agent identity, tone, persona 2. **Operational Context**: Task, user intent, tools 3. **Domain Context**: Industry-specific knowledge 4. **Historical Context**: Condensed interaction memory 5. **Environmental Context**: System state, live data --- ## 🎯 Success Criteria ### Phase 1 (Week 3) - ✅ <5ms retrieval @ 100K patterns - ✅ 20,000+ token reduction achieved - ✅ 20 MCP tools functional - ✅ 4 /prime commands working ### Phase 2 (Week 7) - ✅ +30% effectiveness improvement - ✅ 84% confidence convergence - ✅ 50+ quality patterns seeded - ✅ 5-layer context operational ### Phase 3 (Week 12) - ✅ Skill-creator functional - ✅ Progressive disclosure working - ✅ 10+ skills validated - ✅ Cross-platform compatibility --- ## 🔗 Essential Links ### Documentation - [Comprehensive Analysis](/docs/research/memory-systems-analysis.md) - [Executive Summary](/docs/research/executive-summary.md) - [This Quick Reference](/docs/research/quick-reference.md) ### External Resources - [AgentDB](https://agentdb.ruv.io) - [ReasoningBank Paper](https://arxiv.org/abs/2509.25140) - [Claude Skills](https://www.anthropic.com/news/skills) - [Context Engineering Guide](https://github.com/coleam00/context-engineering-intro) - [Vector Benchmarks](https://www.letta.com/blog/benchmarking-ai-agent-memory) ### GitHub Repositories - [agentdb](https://github.com/ruvnet/agentdb) - [agentic-flow](https://github.com/ruvnet/agentic-flow) - [anthropics/skills](https://github.com/anthropics/skills) - [sqlite-vec](https://github.com/asg017/sqlite-vec) --- ## 💡 Pro Tips ### AgentDB - Start with brute-force, monitor ANN development - Benchmark early and often (`npx agentdb benchmark`) - Keep vectors <1M until HNSW available - Use chunked storage for memory efficiency ### ReasoningBank - Seed 50-100 quality patterns before production - Track confidence convergence rate - Learn from failures (-15% is valuable) - Use right thinking mode per task type ### Skills - Keep skills focused (single responsibility) - Test progressive disclosure thoroughly - Validate skill composition early - Monitor token overhead per skill ### Context Engineering - Measure token usage before/after changes - Use /prime commands for task-specific context - Implement context budgeting from day one - Quality > quantity always --- ## 🚦 Decision Framework ### Use AgentDB when: - ✅ Need <5ms retrieval - ✅ Working with <1M vectors - ✅ Want SQLite reliability - ✅ Need 20 MCP tools ### Use ReasoningBank when: - ✅ Need self-improving system - ✅ Want to learn from failures - ✅ Have domain-specific patterns - ✅ Need +30% effectiveness ### Use Skills when: - ✅ Need composable capabilities - ✅ Want automatic generation - ✅ Require cross-platform support - ✅ Need minimal token overhead ### Use Context Engineering when: - ✅ Context window approaching limit - ✅ Need 20K+ token reduction - ✅ Have task-specific context needs - ✅ Want quality over quantity --- ## 📞 Quick Support ### Issues - AgentDB: https://github.com/ruvnet/agentdb/issues - Agentic-Flow: https://github.com/ruvnet/agentic-flow/issues - Claude-Flow: https://github.com/ruvnet/claude-flow/issues ### Community - Discord: [rUv Community](https://discord.gg/ruv) - Twitter: [@rUvInc](https://twitter.com/rUvInc) --- **Last Updated**: 2025-10-20 **Next Review**: Phase 1 completion (Week 3) **Status**: ✅ Research Complete, Ready for Implementation