diff --git a/hindsight-api-slim/hindsight_api/engine/local_device.py b/hindsight-api-slim/hindsight_api/engine/local_device.py
index ab9704d5a..78bc609b0 100644
--- a/hindsight-api-slim/hindsight_api/engine/local_device.py
+++ b/hindsight-api-slim/hindsight_api/engine/local_device.py
@@ -163,10 +163,14 @@ def _empty_gpu_cache(device_type: str | None) -> None:
 def release_local_inference_memory(device_type: str | None = None) -> None:
     """Release transient heap (and GPU allocator) memory after a local inference batch.
 
-    Frees Python objects, returns freed native pages to the OS, and empties the GPU
-    allocator pool when the model ran on a GPU. Safe to call on every platform and
-    device; the pieces that don't apply are cheap no-ops.
+    Returns freed native pages to the OS, and empties the GPU allocator pool when
+    the model ran on a GPU. Python's normal reference counting already releases
+    the short-lived CPU inference buffers; a full cyclic-GC scan on every batch is
+    needlessly expensive on that hot path, so CPU callers leave cyclic GC to its
+    normal threshold-based schedule. GPU callers retain the existing full cleanup
+    because allocator release is part of the opt-in accelerator memory policy.
     """
-    gc.collect()
+    if device_type != "cpu":
+        gc.collect()
     _heap_trim()
     _empty_gpu_cache(device_type)
diff --git a/hindsight-api-slim/tests/test_local_device.py b/hindsight-api-slim/tests/test_local_device.py
index 874e02069..5ee536c5d 100644
--- a/hindsight-api-slim/tests/test_local_device.py
+++ b/hindsight-api-slim/tests/test_local_device.py
@@ -161,8 +161,14 @@ class TestReleaseLocalInferenceMemory:
         assert calls == ["gc", "trim"]
         assert log == ["cuda"]
 
-    def test_release_cpu_skips_empty_cache(self):
+    def test_release_cpu_skips_gc_and_empty_cache_but_trims_heap(self):
+        calls = []
         log = []
-        with patch.dict(sys.modules, {"torch": _fake_torch(empty_cache_log=log)}):
+        with (
+            patch.dict(sys.modules, {"torch": _fake_torch(empty_cache_log=log)}),
+            patch.object(local_device.gc, "collect", lambda: calls.append("gc")),
+            patch.object(local_device, "_heap_trim", lambda: calls.append("trim")),
+        ):
             release_local_inference_memory("cpu")
+        assert calls == ["trim"]
         assert log == []