diff --git a/conllu/models.py b/conllu/models.py
index 25e5b3b..1b73591 100644
--- a/conllu/models.py
+++ b/conllu/models.py
@@ -95,11 +95,11 @@ def copy(self) -> 'TokenList':
         return TokenList(tokens_copy, self.metadata, self.default_fields)
 
     def extend(self, iterable: T.Union['TokenList', T.Iterable[Token]]) -> None:
+        if not hasattr(self, "metadata"):
+            self.metadata = Metadata()
         if not isinstance(iterable, TokenList):
             iterable = TokenList(iterable)
-
         super(TokenList, self).extend(iterable)
-
         self.metadata.update(iterable.metadata)
 
     def _dict_to_token_and_set_defaults(self, token: T.Union[dict, Token]) -> Token:
@@ -187,7 +187,7 @@ def _create_tree(head_to_token_mapping: T.Dict[int, T.List[Token]], id_: int = 0
         return root
 
     def filter(self, **kwargs: T.Any) -> 'TokenList':
-        tokens: T.Iterable[Token] = self.copy()
+        tokens = self.copy()
 
         for query, value in kwargs.items():
             filtered_tokens = []
@@ -198,9 +198,9 @@ def filter(self, **kwargs: T.Any) -> 'TokenList':
                     if traverse_dict(token, query) == value:
                         filtered_tokens.append(token)
 
-            tokens = filtered_tokens
+            tokens[:] = filtered_tokens
 
-        return TokenList(tokens)
+        return tokens
 
 
 _T = T.TypeVar("_T")
@@ -353,6 +353,8 @@ def copy(self) -> 'SentenceList':
         return SentenceList(sentences_copy, self.metadata)
 
     def extend(self, iterable: T.Union['SentenceList', T.Iterable[TokenList]]) -> None:
+        if not hasattr(self, "metadata"):
+            self.metadata = Metadata()
         if not isinstance(iterable, SentenceList):
             iterable = SentenceList(iterable)
 
diff --git a/tests/test_models.py b/tests/test_models.py
index 838f6d9..f132dad 100644
--- a/tests/test_models.py
+++ b/tests/test_models.py
@@ -1,3 +1,5 @@
+import pickle
+import tempfile
 import unittest
 from textwrap import dedent
 
@@ -252,10 +254,22 @@ def test_no_root_nodes(self):
 
 
 class TestSerialize(unittest.TestCase):
+
     def test_serialize_on_tokenlist(self):
         tokenlist = TokenList([{"id": 1}])
         self.assertEqual(tokenlist.serialize(), serialize(tokenlist))
 
+    def test_pickling_tokenlist(self):
+        tokenlist = TokenList([
+            {"id": 1, "form": "a", "field": "x"},
+            {"id": 2, "form": "dog", "field": "x"},
+        ], metadata={"text": "a dog"})
+        sink = tempfile.NamedTemporaryFile("wb", suffix=".pkl", delete=False)
+        pickle.dump(tokenlist, sink)
+        sink.close()
+        with open(sink.name, "rb") as source:
+            tokenlist_copy = pickle.load(source)
+        self.assertEqual(tokenlist, tokenlist_copy)
 
 class TestFilter(unittest.TestCase):
     def test_basic_filtering(self):
@@ -280,6 +294,16 @@ def test_basic_filtering(self):
             tokenlist
         )
 
+    def test_metadata_propagates(self):
+        tokenlist = TokenList([
+            {"id": 1, "form": "a", "field": "x"},
+            {"id": 2, "form": "dog", "field": "x"},
+        ], metadata={"text": "a dog"})
+        self.assertEqual(
+            tokenlist.filter().metadata,
+            tokenlist.metadata,
+        )
+
     def test_and_filtering(self):
         tokenlist = TokenList([
             {"id": 1, "form": "a", "field": "x"},
@@ -521,6 +545,13 @@ def test_init_nonlist_raises(self):
         with self.assertRaises(ParseException):
             SentenceList((1, 2, 4))
 
+    def test_extend_with_no_metadata(self):
+        sl = SentenceList([TokenList([{"id": 1}])])
+        del sl.metadata
+        sl.extend(sl)
+        self.assertTrue(hasattr(sl, "metadata"))
+        self.assertIsNotNone(sl.metadata)
+
     def test_equals(self):
         tokenlists = [TokenList([{"id": 1}])]
         self.assertEqual(SentenceList(tokenlists), SentenceList(tokenlists))