store: Auto-edge implementation: design analysis findings and fixes

graph/insights/auto-edge-implementation-design-analysis-findings-and-fixes-ab3ffc.md

@@ -0,0 +1,27 @@
+---
+id: ab3ffc18-1707-42ff-ad7e-bffb16f6fdef
+type: insight
+title: "Auto-edge implementation: design analysis findings and fixes"
+tags: [cognitive-memory, edges, code-review, design-patterns, mcp-server, auto-edges]
+importance: 0.7
+confidence: 0.8
+created: "2026-03-01T04:12:45.563773+00:00"
+updated: "2026-03-01T04:12:45.563773+00:00"
+---
+
+# Auto-edge Implementation: Design Analysis Findings and Fixes
+
+## Issues Found and Fixed
+
+1. **Similarity threshold bypass** — keyword-only results lack a `"similarity"` key, so `if similarity is not None` always short-circuited the threshold check. Fixed by requiring tag overlap for the keyword-only path instead.
+
+2. **Backwards description arrow** — The `→` in the edge description could point the wrong direction when a heuristic reversed from/to. Fixed by building the description string after direction resolution.
+
+3. **Wasted slice slots** — `[:AUTO_EDGE_MAX]` slice happened before threshold filtering, consuming slots on results that would later be rejected. Fixed by filter-then-slice ordering.
+
+4. **Weak recall query** — Title-only recall gave weak signal. Fixed by concatenating tags into the query string.
+
+## Deferred Issues (Not Fixed This Session)
+
+- Git commit batching in `relate()` (issue #2)
+- Semantic recall surfacing dormant memories (issue #3)