diff --git a/gc.c b/gc.c
index 0fccc17a32..49a4cc21b6 100644
--- a/gc.c
+++ b/gc.c
@@ -2541,6 +2541,13 @@ newobj_init(VALUE klass, VALUE flags, int wb_protected, rb_objspace_t *objspace,
     // TODO: make it atomic, or ractor local
     objspace->total_allocated_objects++;
 
+#if THREAD_TRACE_MEMORY_ALLOCATIONS
+    rb_thread_t *th = ruby_threadptr_for_trace_memory_allocations();
+    if (th) {
+	ATOMIC_SIZE_INC(th->memory_allocations.total_allocated_objects);
+    }
+#endif
+
 #if RGENGC_PROFILE
     if (wb_protected) {
         objspace->profile.total_generated_normal_object_count++;
@@ -12127,6 +12134,19 @@ objspace_malloc_increase_body(rb_objspace_t *objspace, void *mem, size_t new_siz
 #endif
     }
 
+#if THREAD_TRACE_MEMORY_ALLOCATIONS
+    rb_thread_t *th = ruby_threadptr_for_trace_memory_allocations();
+    if (th) {
+	if (new_size > old_size) {
+	    ATOMIC_SIZE_ADD(th->memory_allocations.total_malloc_bytes, new_size - old_size);
+	}
+
+	if (type == MEMOP_TYPE_MALLOC) {
+	    ATOMIC_SIZE_INC(th->memory_allocations.total_mallocs);
+	}
+    }
+#endif
+
     if (type == MEMOP_TYPE_MALLOC) {
       retry:
         if (malloc_increase > malloc_limit && ruby_native_thread_p() && !dont_gc_val()) {
diff --git a/thread.c b/thread.c
index 56a5644552..176509e1be 100644
--- a/thread.c
+++ b/thread.c
@@ -5289,6 +5289,55 @@ Init_Thread_Mutex(void)
     rb_native_mutex_initialize(&th->interrupt_lock);
 }
 
+#if THREAD_TRACE_MEMORY_ALLOCATIONS
+rb_thread_t *
+ruby_threadptr_for_trace_memory_allocations(void)
+{
+    // The order of this checks is important due
+    // to how Ruby VM is initialized
+    if (GET_VM()->thread_trace_memory_allocations && GET_EC() != NULL) {
+        return GET_THREAD();
+    }
+
+    return NULL;
+}
+
+static VALUE
+rb_thread_s_trace_memory_allocations(VALUE _)
+{
+    return GET_THREAD()->vm->thread_trace_memory_allocations ? Qtrue : Qfalse;
+}
+
+static VALUE
+rb_thread_s_trace_memory_allocations_set(VALUE self, VALUE val)
+{
+    GET_THREAD()->vm->thread_trace_memory_allocations = RTEST(val);
+    return val;
+}
+
+static VALUE
+rb_thread_memory_allocations(VALUE self)
+{
+    rb_thread_t *th = rb_thread_ptr(self);
+
+    if (!th->vm->thread_trace_memory_allocations) {
+        return Qnil;
+    }
+
+    VALUE ret = rb_hash_new();
+
+	VALUE total_allocated_objects = ID2SYM(rb_intern_const("total_allocated_objects"));
+	VALUE total_malloc_bytes = ID2SYM(rb_intern_const("total_malloc_bytes"));
+	VALUE total_mallocs = ID2SYM(rb_intern_const("total_mallocs"));
+
+	rb_hash_aset(ret, total_allocated_objects, SIZET2NUM(th->memory_allocations.total_allocated_objects));
+	rb_hash_aset(ret, total_malloc_bytes, SIZET2NUM(th->memory_allocations.total_malloc_bytes));
+	rb_hash_aset(ret, total_mallocs, SIZET2NUM(th->memory_allocations.total_mallocs));
+
+    return ret;
+}
+#endif
+
 /*
  *  Document-class: ThreadError
  *
@@ -5371,6 +5420,12 @@ Init_Thread(void)
     rb_define_method(rb_cThread, "to_s", rb_thread_to_s, 0);
     rb_define_alias(rb_cThread, "inspect", "to_s");
 
+#if THREAD_TRACE_MEMORY_ALLOCATIONS
+    rb_define_singleton_method(rb_cThread, "trace_memory_allocations", rb_thread_s_trace_memory_allocations, 0);
+    rb_define_singleton_method(rb_cThread, "trace_memory_allocations=", rb_thread_s_trace_memory_allocations_set, 1);
+    rb_define_method(rb_cThread, "memory_allocations", rb_thread_memory_allocations, 0);
+#endif
+
     rb_vm_register_special_exception(ruby_error_stream_closed, rb_eIOError,
                                      "stream closed in another thread");
 
diff --git a/vm_core.h b/vm_core.h
index 4f6e07d818..893b578cc8 100644
--- a/vm_core.h
+++ b/vm_core.h
@@ -114,6 +114,13 @@ extern int ruby_assert_critical_section_entered;
 # define VM_INSN_INFO_TABLE_IMPL 2
 #endif
 
+/*
+ * track a per thread memory allocations
+ */
+#ifndef THREAD_TRACE_MEMORY_ALLOCATIONS
+# define THREAD_TRACE_MEMORY_ALLOCATIONS 1
+#endif
+
 #if defined(NSIG_MAX)           /* POSIX issue 8 */
 # undef NSIG
 # define NSIG NSIG_MAX
@@ -661,6 +668,7 @@ typedef struct rb_vm_struct {
     unsigned int thread_abort_on_exception: 1;
     unsigned int thread_report_on_exception: 1;
     unsigned int thread_ignore_deadlock: 1;
+    unsigned int thread_trace_memory_allocations: 1;
 
     /* object management */
     VALUE mark_object_ary;
@@ -1040,6 +1048,14 @@ typedef struct rb_thread_struct {
 
     struct rb_waiting_list *join_list;
 
+#if THREAD_TRACE_MEMORY_ALLOCATIONS
+    struct {
+        size_t total_allocated_objects;
+        size_t total_malloc_bytes;
+        size_t total_mallocs;
+    } memory_allocations;
+#endif
+
     union {
         struct {
             VALUE proc;
@@ -1967,6 +1983,7 @@ void rb_threadptr_interrupt(rb_thread_t *th);
 void rb_threadptr_unlock_all_locking_mutexes(rb_thread_t *th);
 void rb_threadptr_pending_interrupt_clear(rb_thread_t *th);
 void rb_threadptr_pending_interrupt_enque(rb_thread_t *th, VALUE v);
+rb_thread_t *ruby_threadptr_for_trace_memory_allocations(void);
 VALUE rb_ec_get_errinfo(const rb_execution_context_t *ec);
 void rb_ec_error_print(rb_execution_context_t * volatile ec, volatile VALUE errinfo);
 void rb_execution_context_update(rb_execution_context_t *ec);