diff --git a/TracyLua.hpp b/TracyLua.hpp
index ed23abf5..52d9e4e4 100644
--- a/TracyLua.hpp
+++ b/TracyLua.hpp
@@ -182,8 +182,9 @@ static inline void SendLuaCallstack( lua_State* L, uint32_t depth )
     auto& tail = token->get_tail_index();
     auto item = token->enqueue_begin<tracy::moodycamel::CanAlloc>( magic );
     MemWrite( &item->hdr.type, QueueType::CallstackAlloc );
-    MemWrite( &item->callstack.ptr, (uint64_t)ptr );
-    MemWrite( &item->callstack.thread, GetThreadHandle() );
+    MemWrite( &item->callstackAlloc.ptr, (uint64_t)ptr );
+    MemWrite( &item->callstackAlloc.nativePtr, (uint64_t)Callstack( depth ) );
+    MemWrite( &item->callstackAlloc.thread, GetThreadHandle() );
     tail.store( magic + 1, std::memory_order_release );
 }
 
diff --git a/client/TracyProfiler.cpp b/client/TracyProfiler.cpp
index 3b4d847b..5128cc67 100644
--- a/client/TracyProfiler.cpp
+++ b/client/TracyProfiler.cpp
@@ -1325,10 +1325,15 @@ static void FreeAssociatedMemory( const QueueItem& item )
         tracy_free( (void*)ptr );
         break;
     case QueueType::Callstack:
-    case QueueType::CallstackAlloc:
         ptr = MemRead<uint64_t>( &item.callstack.ptr );
         tracy_free( (void*)ptr );
         break;
+    case QueueType::CallstackAlloc:
+        ptr = MemRead<uint64_t>( &item.callstackAlloc.ptr );
+        tracy_free( (void*)ptr );
+        ptr = MemRead<uint64_t>( &item.callstackAlloc.nativePtr );
+        tracy_free( (void*)ptr );
+        break;
     default:
         assert( false );
         break;
@@ -1404,9 +1409,12 @@ Profiler::DequeueStatus Profiler::Dequeue( moodycamel::ConsumerToken& token )
                     tracy_free( (void*)ptr );
                     break;
                 case QueueType::CallstackAlloc:
-                    ptr = MemRead<uint64_t>( &item->callstack.ptr );
+                    ptr = MemRead<uint64_t>( &item->callstackAlloc.ptr );
                     SendCallstackAlloc( ptr );
                     tracy_free( (void*)ptr );
+                    ptr = MemRead<uint64_t>( &item->callstackAlloc.nativePtr );
+                    SendCallstackPayload( ptr );
+                    tracy_free( (void*)ptr );
                     break;
                 default:
                     assert( false );
diff --git a/common/TracyProtocol.hpp b/common/TracyProtocol.hpp
index 2c450005..2a437d27 100644
--- a/common/TracyProtocol.hpp
+++ b/common/TracyProtocol.hpp
@@ -9,7 +9,7 @@
 namespace tracy
 {
 
-enum : uint32_t { ProtocolVersion = 4 };
+enum : uint32_t { ProtocolVersion = 5 };
 
 using lz4sz_t = uint32_t;
 
diff --git a/common/TracyQueue.hpp b/common/TracyQueue.hpp
index bcc15efe..148d47b9 100644
--- a/common/TracyQueue.hpp
+++ b/common/TracyQueue.hpp
@@ -249,6 +249,13 @@ struct QueueCallstack
     uint64_t thread;
 };
 
+struct QueueCallstackAlloc
+{
+    uint64_t ptr;
+    uint64_t nativePtr;
+    uint64_t thread;
+};
+
 struct QueueCallstackFrameSize
 {
     uint64_t ptr;
@@ -312,6 +319,7 @@ struct QueueItem
         QueueMemFree memFree;
         QueueCallstackMemory callstackMemory;
         QueueCallstack callstack;
+        QueueCallstackAlloc callstackAlloc;
         QueueCallstackFrameSize callstackFrameSize;
         QueueCallstackFrame callstackFrame;
         QueueCrashReport crashReport;
@@ -331,7 +339,7 @@ static const size_t QueueDataSize[] = {
     sizeof( QueueHeader ) + sizeof( QueueZoneBegin ),       // allocated source location, callstack
     sizeof( QueueHeader ) + sizeof( QueueCallstackMemory ),
     sizeof( QueueHeader ) + sizeof( QueueCallstack ),
-    sizeof( QueueHeader ) + sizeof( QueueCallstack ),       // callstack alloc
+    sizeof( QueueHeader ) + sizeof( QueueCallstackAlloc ),
     // above items must be first
     sizeof( QueueHeader ),                                  // terminate
     sizeof( QueueHeader ),                                  // keep alive