diff --git a/paddle/gserver/layers/ConvBaseProjection.cpp b/paddle/gserver/layers/ConvBaseProjection.cpp
index 08f36c516cfdadd42e9333c1c5a7a247df1f263e..19efed7b52ee07a5c509d069c286ccc3b21602f4 100644
--- a/paddle/gserver/layers/ConvBaseProjection.cpp
+++ b/paddle/gserver/layers/ConvBaseProjection.cpp
@@ -17,7 +17,7 @@ limitations under the License. */
 
 namespace paddle {
 
-ThreadLocalD<std::vector<MemoryHandle *>> ConvBaseProjection::convMem_;
+ThreadLocalD<std::vector<MemoryHandlePtr>> ConvBaseProjection::convMem_;
 
 ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config,
                                        ParameterPtr parameter,
@@ -175,18 +175,18 @@ void ConvBaseProjection::reshape(int batchSize) {
 }
 
 void *ConvBaseProjection::getSpaceBytes(size_t size) {
-  std::vector<MemoryHandle *> &convMem = *convMem_;
+  std::vector<MemoryHandlePtr> &convMem = *convMem_;
   if (convMem.empty()) {
     int numDevices = hl_get_device_count();
     convMem.resize(numDevices);
   }
 
   int devId = hl_get_device();
-  MemoryHandle **localMem = &(convMem[devId]);
-  if (NULL == *localMem || size > (*localMem)->getAllocSize()) {
-    *localMem = new GpuMemoryHandle(size);
+  MemoryHandlePtr localMem = convMem[devId];
+  if (NULL == localMem || size > localMem->getAllocSize()) {
+    localMem = std::make_shared<GpuMemoryHandle>(size);
   }
-  return (*localMem)->getBuf();
+  return localMem->getBuf();
 }
 
 ConvBaseProjection::~ConvBaseProjection() {
diff --git a/paddle/gserver/layers/ConvBaseProjection.h b/paddle/gserver/layers/ConvBaseProjection.h
index ebdb57845bb36ac607b1e4c8e02f9d20b6e82a36..bb7ffa627b745f45b0f210cdb58ef87d6990af73 100644
--- a/paddle/gserver/layers/ConvBaseProjection.h
+++ b/paddle/gserver/layers/ConvBaseProjection.h
@@ -105,7 +105,7 @@ protected:
   bool bias_;
 
   std::unique_ptr<Weight> weight_;
-  static ThreadLocalD<std::vector<MemoryHandle*>> convMem_;
+  static ThreadLocalD<std::vector<MemoryHandlePtr>> convMem_;
 };
 
 }  // namespace paddle