T1 Complete — Auto-Extract Knowledge Graph

T1 Complete — Auto-Extract Knowledge Graph

Status: ✅ COMPLETE
Completed: 2026-02-19T18:47 EST
Duration: ~5 minutes extraction time (196K messages × 150+ regex patterns)

Results

Node Type Breakdown

Top Relationship Types

Data Sources

1. Existing graph (`data/expanded_graph.json`) — 70 seed nodes, ground truth
2. Ledger DB (`[home-path]`) — 196,917 messages across 5,934 sessions
3. Knowledge Graph DB (`[home-path]`) — 7,595 entities cross-referenced (37 high-quality ones added)
4. Curated dictionaries — 150+ regex patterns for technologies, projects, concepts, locations, machines

Extraction Strategy

• Pattern matching: 150+ curated regex patterns for known entities (technologies, projects, concepts, locations)
• Co-occurrence analysis: Session-level entity co-occurrence used to build `uses`, `co_used_with`, `related_to` edges
• Cross-reference: knowledge_graph.db entities with 5+ mentions and clean names merged in
• Hardcoded domain knowledge: 100+ manually curated relationship edges for known architecture
• Deduplication: Merged `compcore`→`comp_core`, `serenity_soother`→`serenity`
• Orphan recovery: Technologies with 5+ mentions connected to Mo's ecosystem instead of being pruned

Validation

• ✅ No duplicate node names
• ✅ All edge endpoints exist in node set
• ✅ Valid JSON (131 KB)
• ✅ 270 nodes > 200 target
• ✅ Adjacency list generated for all 270 nodes

Output Files

• Graph: `data/expanded_graph_v2.json` (131 KB)
• Script: `scripts/extract_knowledge_graph.py` (reproducible)

New Entity Categories (not in v1)

• Discord channels (10): #quick, #bridge, #milkmen-delivery, etc.
• Locations (6): NYC, SoHo, Brooklyn, Manhattan, West Africa, Guinea
• HEF evolutions (9): Failure Museum, Serendipity Engine, Thought Mesh, etc.
• Clawdbot subsystems (10): Context Ledger, Cortex, Vault, Agent Recovery, etc.
• ML concepts (15+): RAG, RLM, embeddings, fine-tuning, zero-shot, etc.
• Decisions (14): Extracted from ledger decisions table
• Additional technologies (60+): TypeScript, Docker, FastAPI, PyTorch, Gemini, etc.

Note on kimi_memory.db

The original task referenced `[home-path]` but that file is 0 bytes (empty). The actual conversation data lives in `[home-path]` (1.1GB, 196K messages). The script uses this as the primary data source.