
VectorClaw is a security-hardened MyVector MySQL skill for Hermes that gives your AI agent persistent, intelligent, self-sufficient memory. v5.0.0 makes MyVector the single source of truth — replacing external tools (Mem0, Hancho) with native auto-extraction and knowledge graph reasoning, all in MySQL.
What’s New in v5.0.0
Auto-Extraction Hook — Replaces Mem0
v5.0.0 adds a local LLM-powered extraction hook that automatically pulls atomic facts from conversation text and inserts them into MyVector. No external service needed:
- Uses local qwen3.5:4b with structured JSON prompt
- Extracts: core_fact, confidence, entities, linked_to, tags, memory_type, importance
- Auto-dedup on insert: Jaccard similarity check, merges if >50% overlap
- Source tracking: auto-extracted memories marked with source=’auto’
- Human verification flag for promoting accurate auto-facts
- Quality logging to extraction_log for empirical tuning
Memory Relations + Knowledge Graph — Replaces Hancho
Native MySQL knowledge graph with graph traversal during retrieval:
- memory_relations table: fact_id, related_fact_id, relation_type, confidence
- Relation types: mentions, implies, contradicts, same_entity, related_to
- Auto-discovery during extraction: finds existing memories sharing entities
- Consolidation pass: periodic scanning for contradictions and new edges
- Hub insight derivation: identifies high-degree facts (3+ connections) as important
- memory_graph_1hop view: pre-computed 1-hop traversal for fast retrieval
Source Tracking + Human Verification
Every memory now tracks its provenance:
| Source | Description | Initial Confidence |
|---|---|---|
| manual | Written explicitly by agent | 0.9 |
| auto | Extracted by local LLM hook | 0.6-0.7 |
| consolidation | Derived from consolidation pass | 0.7 |
| import | Imported from external system | 0.5 |
Auto-extracted facts start at lower confidence and get promoted when verified by grounding checks during conversation.
Memory Architecture (v5.0.0)
v5.0.0 completes the consolidation started in v4.0.0. The full memory stack:
- MEMORY.md — Always-in-context narrative (relationships, lessons, emotional context)
- MyVector (MySQL) — Structured profiles, facts, preferences, dimensional tags, knowledge graph edges, extraction logs
- ChromaDB — Semantic search across workspace files (skills, projects, logs)
- Retrieval Gate v3 — Smart triggering: explicit signals, entity SQL pre-filter, depth mode, keyword Jaccard, embedding delta, graph expansion
- Grounding Protocol — Literal message quoting, delta detection, entity self-critique, memory source check
- Auto-Extraction — Local LLM extracts facts from every conversation turn
- Knowledge Graph — Native MySQL relations with contradiction detection
Deprecation Plan
v5.0.0 is designed to make external memory tools unnecessary:
| External Tool | Replacement | Migration |
|---|---|---|
| Mem0 | auto-extract.py (local LLM) | Run 7-10 days parallel, compare quality, then retire |
| Hancho | memory_relations + hancho-consolidate.py | Already native — retire after validation |
How It Compares
| Feature | VectorClaw v5 | Mem0 | Supermemory |
|---|---|---|---|
| Storage | Self-hosted MySQL (MyVector) — you own everything | Cloud-hosted | Cloud-hosted |
| Auto-extraction | ✅ Local LLM (qwen3.5:4b) + regex fallback | ✅ Managed NLP | ✅ Managed NLP |
| Dedup on insert | ✅ Jaccard similarity with merge | ❌ | ❌ |
| Source tracking | ✅ manual/auto/consolidation/import | ❌ | ❌ |
| Knowledge graph | ✅ Native MySQL edges + traversal view | ❌ | ❌ |
| Contradiction detection | ✅ Polarity-based consolidation pass | ❌ | ❌ |
| Post-conversation analysis | ✅ HindSight consolidation | ❌ | ❌ |
| Multi-dimensional tags | ✅ HoloGraphic (emotion, context, urgency, people) | ❌ | ❌ |
| Quality logging | ✅ extraction_log with per-run metrics | ❌ | ❌ |
| Self-hosted | ✅ Full control, no API limits, no vendor lock-in | ❌ Cloud dependency | ❌ Cloud dependency |
| Security | ✅ 20/20 audit score, least-privilege, DDL blocked | Basic | Basic |
| Cost | Free (self-hosted) | Free tier + paid | Free tier + paid |
Requirements
- Docker (for MyVector container)
- Hermes agent
- Ollama with qwen3.5:4b (for auto-extraction)
- ~500MB RAM for the MyVector container
Get It
VectorClaw is available as a free Hermes skill on ClawHub:
🔗 ClawHub: https://clawhub.ai/paradoxfuzzle/custom-mysql
Install with: hermes skills install custom-mysql
🐼 Remember everything. Reason about it. Grow from it. Own it all.