diff --git a/graph/fixes/fix-decay-filter-bypass-in-semantic-recall-cognitive-memory-96a6c1.md b/graph/fixes/fix-decay-filter-bypass-in-semantic-recall-cognitive-memory-96a6c1.md
new file mode 100644
index 00000000000..df00e74f731
--- /dev/null
+++ b/graph/fixes/fix-decay-filter-bypass-in-semantic-recall-cognitive-memory-96a6c1.md
@@ -0,0 +1,32 @@
+---
+id: 96a6c125-bb69-45af-b1b6-a9b48654fb7e
+type: fix
+title: "Fix: decay filter bypass in semantic_recall() — cognitive-memory"
+tags: [cognitive-memory, python, fix, decay, embeddings, semantic-recall]
+importance: 0.6
+confidence: 0.8
+created: "2026-03-03T05:07:32.495008+00:00"
+updated: "2026-03-03T05:07:32.495008+00:00"
+---
+
+## Problem
+
+`semantic_recall()` in `embeddings.py` scored all memories in `_embeddings.json` against the query vector with no decay filter. Memories with `decay_score < THRESHOLD_DORMANT` (0.05) — archived/dormant — could re-surface if their semantic similarity was high enough, bypassing the decay system's intent.
+
+The keyword path in `recall()` (`client.py`) already applied this filter correctly.
+
+## Root Cause
+
+`EmbeddingsMixin.semantic_recall()` only loaded the index (`_load_index()`), not the state (`_load_state()`). The decay threshold guard was missing entirely.
+
+## Solution
+
+Added `THRESHOLD_DORMANT` to `embeddings.py` imports from `common`. In `semantic_recall()`, loaded `_state.json` via `self._load_state()` and added a `decay_score < THRESHOLD_DORMANT` skip before computing cosine similarity.
+
+## Files Changed
+
+- `embeddings.py` — 6-line change: import + state load + filter
+
+## PR
+
+https://git.manticorum.com/cal/cognitive-memory/pulls/7