diff --git a/docs/self-evolution.md b/docs/self-evolution.md
index abfcc34..7a2c233 100644
--- a/docs/self-evolution.md
+++ b/docs/self-evolution.md
@@ -9,7 +9,7 @@ Historical verification snapshot: `51be33361422e55e1f2f00c33a0e0f8c56132a91`
 (the post-#54 `main` revision, captured before this #55 documentation-only
 update). Snapshot date: 2026-09-04.
 
-Current repository test count at this snapshot: **202 unittest cases**.
+Current repository test count at this snapshot: **203 unittest cases**.
 
 ## Verified surface
 
diff --git a/main.py b/main.py
index a68ddf7..2d04134 100644
--- a/main.py
+++ b/main.py
@@ -1962,12 +1962,13 @@ def _detect_user_preferences(user_input: str) -> None:
     detected = {k: v for k, v in _user_preferences.items() if v}
     if detected:
         pref_str = "; ".join(f"{k}={v}" for k, v in detected.items())
-        # Only store if new or changed
-        recent = memory_bank.get_recent(3)
-        already = any(f"PREF: {pref_str}" in r for r in recent)
-        if not already:
-            learning_engine.submit("preference", f"PREF: {pref_str}", evidence_id=stable_hash(user_input),
-                                   confidence=0.5, metadata={"source": "preference_detection"})
+        # Each user turn is an independent observation.  Reusing a content hash
+        # (or suppressing a repeated message) prevents the evidence threshold
+        # from ever promoting a preference across a process restart.
+        learning_engine.submit(
+            "preference", f"PREF: {pref_str}", evidence_id=f"turn-{uuid.uuid4().hex}",
+            confidence=0.5, metadata={"source": "preference_detection"},
+        )
 
 
 def _build_preference_context() -> str:
diff --git a/tests/test_learning_dispatcher.py b/tests/test_learning_dispatcher.py
index e64bf16..f247666 100644
--- a/tests/test_learning_dispatcher.py
+++ b/tests/test_learning_dispatcher.py
@@ -177,6 +177,43 @@ def test_verified_preferences_hydrate_in_fresh_runtime_and_prompt(self):
             self.assertIn("naming_style=snake_case", completed.stdout)
             self.assertIn("Known user preferences", completed.stdout)
 
+    def test_detected_preference_promotes_and_survives_restart(self):
+        with tempfile.TemporaryDirectory(prefix="openkyrozen-detected-preference-") as directory:
+            root = Path(directory)
+            db_path = root / "state.sqlite3"
+            env = os.environ.copy()
+            env.update({
+                "HOME": str(root), "KYROZEN_DB_PATH": str(db_path),
+                "KYROZEN_DISABLE_VECTOR_INDEX": "1", "KYROZEN_WORKSPACE_ROOT": str(root),
+                "PYTHONPATH": str(Path(__file__).parents[1]),
+            })
+            observe = subprocess.run(
+                [sys.executable, "-c", (
+                    "import main; "
+                    f"main.configure_launch_context(project_path={str(root)!r}); "
+                    "[main.dispatch_learning_cycle(surface='cli', trigger='turn', max_features=1, "
+                    "user_input='Please use concise Python and snake_case names.', "
+                    "feature_names=('detect_user_preferences',)) for _ in range(2)]; "
+                    "print(main.learning_engine.status(100)[0]['status'])"
+                )],
+                cwd=Path(__file__).parents[1], env=env, capture_output=True, text=True, check=False,
+            )
+            self.assertEqual(observe.returncode, 0, observe.stderr)
+            self.assertIn("active", observe.stdout)
+
+            fresh = subprocess.run(
+                [sys.executable, "-c", (
+                    "import main; "
+                    f"main.configure_launch_context(project_path={str(root)!r}); "
+                    "print(main._build_preference_context())"
+                )],
+                cwd=Path(__file__).parents[1], env=env, capture_output=True, text=True, check=False,
+            )
+            self.assertEqual(fresh.returncode, 0, fresh.stderr)
+            self.assertIn("language=python", fresh.stdout)
+            self.assertIn("naming_style=snake_case", fresh.stdout)
+            self.assertIn("verbosity=concise", fresh.stdout)
+
     def test_feature_failure_is_recorded_without_stopping_the_cycle(self):
         names = main._LEARNING_FEATURE_ORDER[:2]
         registry = dict(main._LEARNING_FEATURE_REGISTRY)