Graph + Vector Intelligence (Mega-Cube 17) -- Stage 1, Path A: The Edge Table

# Graph + Vector Intelligence (Mega-Cube 17) -- Stage 1, Path A: The Edge Table **Run:** mega-cube-17-graph-vector-intel **Generated:** 2026-04-04 **Method:** Evolution-Cube Stage 1 -- Divergent exploration (relational graph in Postgres) **Run Directory:** Desktop/evo-cube-output/mega-cube-17-graph-vector-intel/ Add a single `memory_edges` table to the existing Postgres schema. Edges connect `memory_turns` rows by relationship type: temporal (same session), referential (shared file), causal (correction/follow-up), and categorical (shared inscription). Vector search stays exactly as-is. Graph queries are standard SQL joins. No new services, no new databases, no new dependencies. The graph is a Postgres table. | Type | Source | Target | Weight | Metadata | |------|--------|--------|--------|----------| | `temporal` | Turn N in session | Turn N+1 in session | 1.0 (decays with gap) | `{session_id, gap_seconds}` | | `referential` | Turn that reads file X | Turn that wrote file X | cosine(embed_A, embed_B) | `{file_path, operation}` | | `causal` | Correction turn | Original turn | KARL correction score | `{correction_type, overlap}` | | `categorical` | Turn with inscription A | Turn with same inscription | inscription similarity | `{inscription, category}` | | `trajectory` | Trajectory start | Trajectory end | KARL composite reward | `{skill, domain, reward}` | - 332K turns, avg 5 turns/session = ~66K sessions - Temporal edges: ~266K (N-1 per session of N turns) - Referential edges: ~50K (file path overlap across sessions) - Causal edges: ~5K (KARL corrections + redo detections) - Categorical edges: ~20K (inscription-matched pairs, sampled) - Trajectory edges: ~121 (KARL trajectory start-to-end) - **Total: ~341K edges.** Fits in ~50MB with indexes. Weight decays with temporal gap: 1.0 for consecutive turns, ~0.5 for 1-hour gap, ~0.1 for 10-hour gap.