diff --git a/paddle/capi/gradient_machine.cpp b/paddle/capi/gradient_machine.cpp
index 482b51e8a8430863c3e13df2298f6979d3959461..1f0e033c5b7ab713f1747d58f4fcdaad001f0715 100644
--- a/paddle/capi/gradient_machine.cpp
+++ b/paddle/capi/gradient_machine.cpp
@@ -168,3 +168,13 @@ paddle_error paddle_gradient_machine_get_layer_output(
   out->args.push_back(layerOutput);
   return kPD_NO_ERROR;
 }
+
+paddle_error paddle_gradient_machine_release_layer_output(
+    paddle_gradient_machine machine) {
+  auto m = cast(machine);
+  if (m == nullptr || m->machine == nullptr) {
+    return kPD_NULLPTR;
+  }
+  m->machine->releaseOutput();
+  return kPD_NO_ERROR;
+}
diff --git a/paddle/capi/gradient_machine.h b/paddle/capi/gradient_machine.h
index 28eeb23e3bbdd4cc22a25c14170bf56c294f8cd7..7e37dea00b27b6cb955486b4210c1373decbcfa5 100644
--- a/paddle/capi/gradient_machine.h
+++ b/paddle/capi/gradient_machine.h
@@ -113,6 +113,14 @@ paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
                                          const char* layerName,
                                          paddle_arguments args);
 
+/**
+ * @brief Release the middle layer's output memory of the gradient machine.
+ * @param [in] gradient machine that have run a inference
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_release_layer_output(paddle_gradient_machine machine);
+
 #ifdef __cplusplus
 }
 #endif