MCP Memory Service

# Memory Quality System - Detailed Reference **Quick Summary for CLAUDE.md**: See main file for architecture overview. This file contains implementation details, configuration options, and troubleshooting. ## Complete Configuration Options ```bash # Quality System (Local-First Defaults) MCP_QUALITY_SYSTEM_ENABLED=true # Default: enabled MCP_QUALITY_AI_PROVIDER=local # local|groq|gemini|auto|none MCP_QUALITY_LOCAL_MODEL=nvidia-quality-classifier-deberta # Default v8.49.0+ MCP_QUALITY_LOCAL_DEVICE=auto # auto|cpu|cuda|mps|directml # Legacy model (backward compatible, not recommended) # MCP_QUALITY_LOCAL_MODEL=ms-marco-MiniLM-L-6-v2 # Quality-Boosted Search (Recommended with DeBERTa) MCP_QUALITY_BOOST_ENABLED=true # More accurate with DeBERTa MCP_QUALITY_BOOST_WEIGHT=0.3 # 0.3 = 30% quality, 70% semantic # Quality-Based Retention MCP_QUALITY_RETENTION_HIGH=365 # Days for quality ≥0.7 MCP_QUALITY_RETENTION_MEDIUM=180 # Days for 0.5-0.7 MCP_QUALITY_RETENTION_LOW_MIN=30 # Min days for <0.5 ``` ## MCP Tools - `rate_memory(content_hash, rating, feedback)` - Manual quality rating (-1/0/1) - `get_memory_quality(content_hash)` - Retrieve quality metrics - `analyze_quality_distribution(min_quality, max_quality)` - System-wide analytics - `retrieve_with_quality_boost(query, n_results, quality_weight)` - Quality-boosted search ## Migration from MS-MARCO to DeBERTa **Why Migrate:** - ✅ Eliminates self-matching bias (no query needed) - ✅ Uniform distribution (mean 0.60-0.70 vs 0.469) - ✅ Fewer false positives (<5% perfect scores vs 20%) - ✅ Absolute quality assessment vs relative ranking **Migration Guide**: See [docs/guides/memory-quality-guide.md](../../docs/guides/memory-quality-guide.md#migration-from-ms-marco-to-deberta) ## Success Metrics (Phase 1 - v8.48.3) **Achieved:** - ✅ <100ms search latency with quality boost (45ms avg, +17% overhead) - ✅ $0 monthly cost (local SLM default) - ✅ 75% local SLM usage (3,570 of 4,762 memories) - ✅ 95% quality score coverage **Challenges:** - ⚠️ Average score 0.469 (target: 0.6+) - ⚠️ Self-matching bias ~25% - ⚠️ Quality boost minimal ranking improvement (0-3%) **Next Phase**: See [Issue #268](https://github.com/doobidoo/mcp-memory-service/issues/268) ## Troubleshooting See [docs/guides/memory-quality-guide.md](../../docs/guides/memory-quality-guide.md) for: - Model download issues - Performance tuning - Quality score interpretation - User feedback integration