diff --git a/.jules/bolt.md b/.jules/bolt.md
index 96ecba4..efe25c0 100644
--- a/.jules/bolt.md
+++ b/.jules/bolt.md
@@ -1,3 +1,7 @@
 ## 2024-02-18 - Pre-scaled Audio Feedback
 **Learning:** AudioFeedback volume scaling was applied during every playback, causing unnecessary numpy overhead and latency.
 **Action:** Pre-calculate scaled audio during `_load_and_cache` to minimize `_play_cached` latency (from ~1ms to ~0.04ms).
+
+## 2025-05-18 - Unlocked Garbage Collection
+**Learning:** `ParakeetManager` held the model lock during `gc.collect()`, forcing concurrent `transcribe` requests (which require the lock) to wait for full GC completion before they could even *start* reloading the model.
+**Action:** Release the lock before calling `gc.collect()`. This allows `transcribe` to acquire the lock and begin `load_model` (which releases GIL for C++ ops) in parallel with the garbage collection of the old model.
diff --git a/src/chirp/parakeet_manager.py b/src/chirp/parakeet_manager.py
index 54667ef..2cc1440 100644
--- a/src/chirp/parakeet_manager.py
+++ b/src/chirp/parakeet_manager.py
@@ -65,11 +65,15 @@ def _monitor_loop(self) -> None:
                 self._unload_model()
 
     def _unload_model(self) -> None:
+        should_collect = False
         with self._lock:
             if self._model is not None and (time.time() - self._last_access > self._timeout):
                 self._logger.info("Unloading Parakeet model to free memory.")
                 self._model = None
-                gc.collect()
+                should_collect = True
+
+        if should_collect:
+            gc.collect()
 
     def ensure_loaded(self):
         with self._lock: