diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.cc b/paddle/fluid/inference/analysis/ir_pass_manager.cc
index 26bca9b1e54ecb854b944e7b1311d4877f2c6796..a4e263e2f464c4021b049093c49ddaecb056284f 100644
--- a/paddle/fluid/inference/analysis/ir_pass_manager.cc
+++ b/paddle/fluid/inference/analysis/ir_pass_manager.cc
@@ -114,13 +114,25 @@ void IRPassManager::CreatePasses(Argument *argument,
               "When you are in TRT INT8 mode, and load model from "
               "memory, you should set optim_cache_dir using "
               "config.SetOptimCacheDir()"));
-      PADDLE_ENFORCE_EQ(
-          !(model_from_memory && use_static_engine), true,
-          platform::errors::PreconditionNotMet(
-              "When you are using Paddle-TRT, and also using load model "
-              "from memory, you should set the use_static to false."));
+      if (model_from_memory && use_static_engine) {
+        PADDLE_ENFORCE_EQ(
+            optim_cache_dir.empty(), false,
+            platform::errors::PreconditionNotMet(
+                "When you are using Paddle-TRT, and using load model "
+                "from memory, and also set the use_static to true. "
+                "you must set optim_cache_dir using "
+                "config.SetOptimCacheDir()."));
+      }
 
       if (!optim_cache_dir.empty()) {
+        if (!PathExists(optim_cache_dir)) {
+          PADDLE_ENFORCE_NE(
+              MKDIR(optim_cache_dir.c_str()), -1,
+              platform::errors::PreconditionNotMet(
+                  "Can not create optimize cache directory: %s, Make sure you "
+                  "have permission to write",
+                  optim_cache_dir));
+        }
         pass->Set("model_opt_cache_dir", new std::string(optim_cache_dir));
       } else if (use_static_engine || enable_int8) {
         std::string model_opt_cache_dir =
diff --git a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
index a450ebdf8919637f6753306e61c2467171cd9654..75111701f1f388d81d273fdc14847ead6772e581 100644
--- a/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
+++ b/paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
@@ -250,7 +250,6 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
   auto predictor_id = Get<int>("predictor_id");
 
   // Get "" when there is no cached calibration table data.
-  bool load_from_memory = Get<bool>("model_from_memory");
   std::string calibration_data = "";
   if (enable_int8 && use_calib_mode) {
     calibration_data = GetTrtCalibTableData(
@@ -323,8 +322,7 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
       graph->Has(framework::ir::kEmbEltwiseLayernormPass) &&
       graph->Has(framework::ir::kMultiheadMatmulPass));
 
-  bool need_serialize = (use_static_engine && !load_from_memory);
-  if (need_serialize) {
+  if (use_static_engine) {
     trt_engine_serialized_data = GetTrtEngineSerializedData(
         Get<std::string>("model_opt_cache_dir"), engine_key);
     // we can load the engine info serialized before from the disk.
@@ -352,7 +350,7 @@ void TensorRtSubgraphPass::CreateTensorRTOp(
           std::vector<std::string>(input_names.begin(), input_names.end()),
           param_set, output_mapping, trt_engine);
 
-  if (need_serialize) {
+  if (use_static_engine) {
     nvinfer1::IHostMemory *serialized_engine_data = trt_engine->Serialize();
     trt_engine_serialized_data =
         std::string((const char *)serialized_engine_data->data(),
diff --git a/paddle/fluid/inference/tests/api/trt_dynamic_shape_test.cc b/paddle/fluid/inference/tests/api/trt_dynamic_shape_test.cc
index 552aefac9b6da590d3684b50d827a0b78b0f8817..55ee2082e69593e6a4226c7cc2b11916f7c0f814 100644
--- a/paddle/fluid/inference/tests/api/trt_dynamic_shape_test.cc
+++ b/paddle/fluid/inference/tests/api/trt_dynamic_shape_test.cc
@@ -21,17 +21,32 @@ limitations under the License. */
 namespace paddle {
 namespace inference {
 
-void TestDynamic(bool with_dynamic = true) {
+void TestDynamic(bool with_dynamic = true, bool delete_cache = true,
+                 bool delete_conv_bn = false) {
   std::string model_dir =
       FLAGS_infer_model + "/conv_bn_swish_split_gelu/conv_bn_swish_split_gelu";
+
+  std::string opt_cache_dir = model_dir + "/my_cache";
+  if (delete_cache) {
+    delete_cache_files(opt_cache_dir);
+  }
+
   AnalysisConfig config;
   config.EnableUseGpu(100, 0);
-  config.SetModel(model_dir + "/model", model_dir + "/params");
+  std::string buffer_prog, buffer_param;
+  ReadBinaryFile(model_dir + "/model", &buffer_prog);
+  ReadBinaryFile(model_dir + "/params", &buffer_param);
+  config.SetModelBuffer(&buffer_prog[0], buffer_prog.size(), &buffer_param[0],
+                        buffer_param.size());
+  config.SetOptimCacheDir(opt_cache_dir);
+
   config.SwitchUseFeedFetchOps(false);
   // Set the input's min, max, opt shape
-
   config.EnableTensorRtEngine(1 << 30, 1, 1,
-                              AnalysisConfig::Precision::kFloat32, false, true);
+                              AnalysisConfig::Precision::kFloat32, true, true);
+  if (delete_conv_bn) {
+    config.pass_builder()->DeletePass("conv_bn_fuse_pass");
+  }
   if (with_dynamic) {
     std::map<std::string, std::vector<int>> min_input_shape = {
         {"image", {1, 1, 3, 3}}};
@@ -130,6 +145,12 @@ void TestDynamic2() {
 
 TEST(AnalysisPredictor, trt_dynamic) { TestDynamic(true); }
 TEST(AnalysisPredictor, trt_static) { TestDynamic(false); }
+TEST(AnalysisPredictor, trt_memory_serialize) {
+  // serailize
+  TestDynamic(false, true, true);
+  // deserailize
+  TestDynamic(false, false, true);
+}
 TEST(AnalysisPredictor, trt_dynamic2) { TestDynamic2(); }
 
 }  // namespace inference
diff --git a/paddle/fluid/inference/tests/api/trt_test_helper.h b/paddle/fluid/inference/tests/api/trt_test_helper.h
index 1abde733581218bed5c9058ca505ec8c6d965f1e..aaa285b2fc2c9e991ac6c4433a5dc82dcb353ca4 100644
--- a/paddle/fluid/inference/tests/api/trt_test_helper.h
+++ b/paddle/fluid/inference/tests/api/trt_test_helper.h
@@ -148,6 +148,7 @@ void delete_cache_files(std::string path) {
       remove(file_rm.c_str());
     }
   }
+  remove(path.c_str());
 }
 
 }  // namespace inference