diff --git a/mace/core/device_context.cc b/mace/core/device_context.cc
index 88a965fa2635da79dda3f3158b084c8ec8f41b11..4f5ff8b10e700be29a697a7cf92a583777222f7f 100644
--- a/mace/core/device_context.cc
+++ b/mace/core/device_context.cc
@@ -58,16 +58,16 @@ GPUContext::GPUContext(const std::string &storage_path,
 
 GPUContext::~GPUContext() = default;
 
-KVStorage *GPUContext::opencl_binary_storage() {
-  return opencl_binary_storage_.get();
+std::shared_ptr<KVStorage> GPUContext::opencl_binary_storage() {
+  return opencl_binary_storage_;
 }
 
-KVStorage *GPUContext::opencl_cache_storage() {
-  return opencl_cache_storage_.get();
+std::shared_ptr<KVStorage> GPUContext::opencl_cache_storage() {
+  return opencl_cache_storage_;
 }
 
-Tuner<uint32_t> *GPUContext::opencl_tuner() {
-  return opencl_tuner_.get();
+std::shared_ptr<Tuner<uint32_t>> GPUContext::opencl_tuner() {
+  return opencl_tuner_;
 }
 
 }  // namespace mace
diff --git a/mace/core/device_context.h b/mace/core/device_context.h
index 21d076730f25b3070c0f680de5b6370d860612f9..ea1e6590cfbe1afba0eb7d7df2cdfa57d168710c 100644
--- a/mace/core/device_context.h
+++ b/mace/core/device_context.h
@@ -32,15 +32,15 @@ class GPUContext {
              const std::string &opencl_parameter_path = "");
   ~GPUContext();
 
-  KVStorage *opencl_binary_storage();
-  KVStorage *opencl_cache_storage();
-  Tuner<uint32_t> *opencl_tuner();
+  std::shared_ptr<KVStorage> opencl_binary_storage();
+  std::shared_ptr<KVStorage> opencl_cache_storage();
+  std::shared_ptr<Tuner<uint32_t>> opencl_tuner();
 
  private:
   std::unique_ptr<KVStorageFactory> storage_factory_;
-  std::unique_ptr<Tuner<uint32_t>> opencl_tuner_;
-  std::unique_ptr<KVStorage> opencl_binary_storage_;
-  std::unique_ptr<KVStorage> opencl_cache_storage_;
+  std::shared_ptr<Tuner<uint32_t>> opencl_tuner_;
+  std::shared_ptr<KVStorage> opencl_binary_storage_;
+  std::shared_ptr<KVStorage> opencl_cache_storage_;
 };
 
 }  // namespace mace
diff --git a/mace/core/file_storage.cc b/mace/core/file_storage.cc
index 7c1fb35b3ebac4df42d2aac451ccbbbf0b4de464..6e6af1e23697bc7411c32b1b5b2ba36cee07888e 100644
--- a/mace/core/file_storage.cc
+++ b/mace/core/file_storage.cc
@@ -32,7 +32,7 @@ class FileStorageFactory::Impl {
  public:
   explicit Impl(const std::string &path);
 
-  std::unique_ptr<KVStorage> CreateStorage(const std::string &name);
+  std::shared_ptr<KVStorage> CreateStorage(const std::string &name);
 
  private:
   std::string path_;
@@ -40,10 +40,9 @@ class FileStorageFactory::Impl {
 
 FileStorageFactory::Impl::Impl(const std::string &path): path_(path) {}
 
-std::unique_ptr<KVStorage> FileStorageFactory::Impl::CreateStorage(
+std::shared_ptr<KVStorage> FileStorageFactory::Impl::CreateStorage(
     const std::string &name) {
-  return std::move(std::unique_ptr<KVStorage>(
-      new FileStorage(path_ + "/" + name)));
+  return std::shared_ptr<KVStorage>(new FileStorage(path_ + "/" + name));
 }
 
 FileStorageFactory::FileStorageFactory(const std::string &path):
@@ -51,7 +50,7 @@ FileStorageFactory::FileStorageFactory(const std::string &path):
 
 FileStorageFactory::~FileStorageFactory() = default;
 
-std::unique_ptr<KVStorage> FileStorageFactory::CreateStorage(
+std::shared_ptr<KVStorage> FileStorageFactory::CreateStorage(
     const std::string &name) {
   return impl_->CreateStorage(name);
 }
diff --git a/mace/core/file_storage.h b/mace/core/file_storage.h
index c4efe8c3565229b371a99f59fc71345c041577bf..7d15449f87af39cbdf59adcd41828574a3927d1b 100644
--- a/mace/core/file_storage.h
+++ b/mace/core/file_storage.h
@@ -41,7 +41,7 @@ class KVStorage {
 
 class KVStorageFactory {
  public:
-  virtual std::unique_ptr<KVStorage> CreateStorage(const std::string &name) = 0;
+  virtual std::shared_ptr<KVStorage> CreateStorage(const std::string &name) = 0;
 
   virtual ~KVStorageFactory() {}
 };
@@ -53,7 +53,7 @@ class FileStorageFactory : public KVStorageFactory {
 
   ~FileStorageFactory();
 
-  std::unique_ptr<KVStorage> CreateStorage(const std::string &name) override;
+  std::shared_ptr<KVStorage> CreateStorage(const std::string &name) override;
 
  private:
   class Impl;
diff --git a/mace/core/runtime/opencl/gpu_device.cc b/mace/core/runtime/opencl/gpu_device.cc
index 1c85273e372a1de1c7e6214c29fe58d8b0a3a063..112a94bff7dd3ee0c24e56aae1431b24b258ce50 100644
--- a/mace/core/runtime/opencl/gpu_device.cc
+++ b/mace/core/runtime/opencl/gpu_device.cc
@@ -18,11 +18,11 @@
 
 namespace mace {
 
-GPUDevice::GPUDevice(Tuner<uint32_t> *tuner,
-                     KVStorage *opencl_cache_storage,
+GPUDevice::GPUDevice(std::shared_ptr<Tuner<uint32_t>> tuner,
+                     std::shared_ptr<KVStorage> opencl_cache_storage,
                      const GPUPriorityHint priority,
                      const GPUPerfHint perf,
-                     KVStorage *opencl_binary_storage,
+                     std::shared_ptr<KVStorage> opencl_binary_storage,
                      const int num_threads,
                      CPUAffinityPolicy cpu_affinity_policy,
                      bool use_gemmlowp) :
diff --git a/mace/core/runtime/opencl/gpu_device.h b/mace/core/runtime/opencl/gpu_device.h
index 64a2d5e34f5fa69b9d456782d4538adf6dca4edd..1d36461b219ce5b28e4efa7ce6f769613eb92634 100644
--- a/mace/core/runtime/opencl/gpu_device.h
+++ b/mace/core/runtime/opencl/gpu_device.h
@@ -25,11 +25,11 @@ namespace mace {
 
 class GPUDevice : public CPUDevice {
  public:
-  GPUDevice(Tuner<uint32_t> *tuner,
-            KVStorage *opencl_cache_storage = nullptr,
+  GPUDevice(std::shared_ptr<Tuner<uint32_t>> tuner,
+            std::shared_ptr<KVStorage> opencl_cache_storage = nullptr,
             const GPUPriorityHint priority = GPUPriorityHint::PRIORITY_LOW,
             const GPUPerfHint perf = GPUPerfHint::PERF_NORMAL,
-            KVStorage *opencl_binary_storage = nullptr,
+            std::shared_ptr<KVStorage> opencl_binary_storage = nullptr,
             const int num_threads = -1,
             CPUAffinityPolicy cpu_affinity_policy = AFFINITY_NONE,
             bool use_gemmlowp = false);
diff --git a/mace/core/runtime/opencl/opencl_runtime.cc b/mace/core/runtime/opencl/opencl_runtime.cc
index 97da0e45b3b8f5430e5a487eb74b26e68d3f0df5..18840b7107619adf94aca7ae739caa3358d33fd3 100644
--- a/mace/core/runtime/opencl/opencl_runtime.cc
+++ b/mace/core/runtime/opencl/opencl_runtime.cc
@@ -273,11 +273,11 @@ void OpenCLProfilingTimer::ClearTiming() {
 }
 
 OpenCLRuntime::OpenCLRuntime(
-    KVStorage *cache_storage,
+    std::shared_ptr<KVStorage> cache_storage,
     const GPUPriorityHint priority_hint,
     const GPUPerfHint perf_hint,
-    KVStorage *precompiled_binary_storage,
-    Tuner<uint32_t> *tuner):
+    std::shared_ptr<KVStorage> precompiled_binary_storage,
+    std::shared_ptr<Tuner<uint32_t>> tuner):
     cache_storage_(cache_storage),
     precompiled_binary_storage_(precompiled_binary_storage),
     tuner_(tuner),
@@ -460,7 +460,7 @@ cl::Device &OpenCLRuntime::device() { return *device_; }
 
 cl::CommandQueue &OpenCLRuntime::command_queue() { return *command_queue_; }
 
-Tuner<uint32_t> *OpenCLRuntime::tuner() { return tuner_; }
+Tuner<uint32_t> *OpenCLRuntime::tuner() { return tuner_.get(); }
 
 uint64_t OpenCLRuntime::device_global_mem_cache_size() const {
   return device_gloabl_mem_cache_size_;
diff --git a/mace/core/runtime/opencl/opencl_runtime.h b/mace/core/runtime/opencl/opencl_runtime.h
index 8480848944b2aba302dcc6007f92b26d5081085d..3d182a9e0def86d6ecf73cd5140751b0d1702d31 100644
--- a/mace/core/runtime/opencl/opencl_runtime.h
+++ b/mace/core/runtime/opencl/opencl_runtime.h
@@ -65,11 +65,11 @@ const std::string OpenCLErrorToString(cl_int error);
 class OpenCLRuntime {
  public:
   OpenCLRuntime(
-      KVStorage *cache_storage = nullptr,
+      std::shared_ptr<KVStorage> cache_storage = nullptr,
       const GPUPriorityHint priority_hint = GPUPriorityHint::PRIORITY_NORMAL,
       const GPUPerfHint perf_hint = GPUPerfHint::PERF_NORMAL,
-      KVStorage *precompiled_binary_storage = nullptr,
-      Tuner<uint32_t> *tuner = nullptr);
+      std::shared_ptr<KVStorage> precompiled_binary_storage = nullptr,
+      std::shared_ptr<Tuner<uint32_t>> tuner = nullptr);
   ~OpenCLRuntime();
   OpenCLRuntime(const OpenCLRuntime &) = delete;
   OpenCLRuntime &operator=(const OpenCLRuntime &) = delete;
@@ -126,9 +126,9 @@ class OpenCLRuntime {
   OpenCLVersion ParseDeviceVersion(const std::string &device_version);
 
  private:
-  KVStorage *cache_storage_;
-  KVStorage *precompiled_binary_storage_;
-  Tuner<uint32_t> *tuner_;
+  std::shared_ptr<KVStorage> cache_storage_;
+  std::shared_ptr<KVStorage> precompiled_binary_storage_;
+  std::shared_ptr<Tuner<uint32_t>> tuner_;
   bool is_opencl_avaliable_;
   bool is_profiling_enabled_;
   OpenCLVersion opencl_version_;
diff --git a/mace/examples/android/macelibrary/src/main/cpp/image_classify.cc b/mace/examples/android/macelibrary/src/main/cpp/image_classify.cc
index 9928686c82b3162550d2e99ffd01d632a4b99259..ae5bbc78f58df7ebbcf24abe49118c03960afa18 100755
--- a/mace/examples/android/macelibrary/src/main/cpp/image_classify.cc
+++ b/mace/examples/android/macelibrary/src/main/cpp/image_classify.cc
@@ -67,7 +67,6 @@ mace::DeviceType ParseDeviceType(const std::string &device) {
 }
 
 MaceContext& GetMaceContext() {
-  // stay for the app's life time, only initialize once
   static auto *mace_context = new MaceContext;
 
   return *mace_context;
diff --git a/mace/public/mace.h b/mace/public/mace.h
index db50a58eebfef013f4a6dee925a7ae87aa95bead..313e1afb551fbffe6bc4c798296d89dcc08171fe 100644
--- a/mace/public/mace.h
+++ b/mace/public/mace.h
@@ -99,10 +99,8 @@ enum MaceStatus {
 
 /// \brief GPU context contain the status used for GPU device.
 ///
-/// The life cycle of GPUContext object is the same as MaceEngines use it.
-/// Just use one GPUContext for all MaceEngines, which will speed up the
-/// initialization procedure. There are some data in common between different
-/// MaceEngines using GPU, use one GPUContext could avoid duplication.
+/// There are some data in common between different MaceEngines using GPU,
+/// use one GPUContext could avoid duplication.
 class GPUContext;
 
 /// \brief GPUContext builder.