diff --git a/doc/source/whatsnew/v3.0.0.rst b/doc/source/whatsnew/v3.0.0.rst
index 781b3b2282a87..723bb6a12060c 100644
--- a/doc/source/whatsnew/v3.0.0.rst
+++ b/doc/source/whatsnew/v3.0.0.rst
@@ -332,6 +332,7 @@ Performance improvements
 - Performance improvement in :meth:`Index.take` when ``indices`` is a full range indexer from zero to length of index (:issue:`56806`)
 - Performance improvement in :meth:`Index.to_frame` returning a :class:`RangeIndex` columns of a :class:`Index` when possible. (:issue:`58018`)
 - Performance improvement in :meth:`MultiIndex.equals` for equal length indexes (:issue:`56990`)
+- Performance improvement in :meth:`MultiIndex.memory_usage` to ignore the index engine when it isn't already cached. (:issue:`58385`)
 - Performance improvement in :meth:`RangeIndex.__getitem__` with a boolean mask or integers returning a :class:`RangeIndex` instead of a :class:`Index` when possible. (:issue:`57588`)
 - Performance improvement in :meth:`RangeIndex.append` when appending the same index (:issue:`57252`)
 - Performance improvement in :meth:`RangeIndex.argmin` and :meth:`RangeIndex.argmax` (:issue:`57823`)
diff --git a/pandas/core/indexes/base.py b/pandas/core/indexes/base.py
index 5654111132b5e..e08b585920779 100644
--- a/pandas/core/indexes/base.py
+++ b/pandas/core/indexes/base.py
@@ -4863,8 +4863,9 @@ def _from_join_target(self, result: np.ndarray) -> ArrayLike:
     def memory_usage(self, deep: bool = False) -> int:
         result = self._memory_usage(deep=deep)
 
-        # include our engine hashtable
-        result += self._engine.sizeof(deep=deep)
+        # include our engine hashtable, only if it's already cached
+        if "_engine" in self._cache:
+            result += self._engine.sizeof(deep=deep)
         return result
 
     @final
diff --git a/pandas/core/indexes/multi.py b/pandas/core/indexes/multi.py
index 21ce9b759f2df..c8e16fad00d5b 100644
--- a/pandas/core/indexes/multi.py
+++ b/pandas/core/indexes/multi.py
@@ -1391,8 +1391,9 @@ def _nbytes(self, deep: bool = False) -> int:
         names_nbytes = sum(getsizeof(i, objsize) for i in self.names)
         result = level_nbytes + label_nbytes + names_nbytes
 
-        # include our engine hashtable
-        result += self._engine.sizeof(deep=deep)
+        # include our engine hashtable, only if it's already cached
+        if "_engine" in self._cache:
+            result += self._engine.sizeof(deep=deep)
         return result
 
     # --------------------------------------------------------------------
diff --git a/pandas/tests/indexes/test_old_base.py b/pandas/tests/indexes/test_old_base.py
index 9b4470021cc1d..b929616c814ee 100644
--- a/pandas/tests/indexes/test_old_base.py
+++ b/pandas/tests/indexes/test_old_base.py
@@ -326,6 +326,30 @@ def test_memory_usage(self, index):
         if index.inferred_type == "object":
             assert result3 > result2
 
+    def test_memory_usage_doesnt_trigger_engine(self, index):
+        index._cache.clear()
+        assert "_engine" not in index._cache
+
+        res_without_engine = index.memory_usage()
+        assert "_engine" not in index._cache
+
+        # explicitly load and cache the engine
+        _ = index._engine
+        assert "_engine" in index._cache
+
+        res_with_engine = index.memory_usage()
+
+        # the empty engine doesn't affect the result even when initialized with values,
+        # because engine.sizeof() doesn't consider the content of engine.values
+        assert res_with_engine == res_without_engine
+
+        if len(index) == 0:
+            assert res_without_engine == 0
+            assert res_with_engine == 0
+        else:
+            assert res_without_engine > 0
+            assert res_with_engine > 0
+
     def test_argsort(self, index):
         if isinstance(index, CategoricalIndex):
             pytest.skip(f"{type(self).__name__} separately tested")