diff --git a/jedi/cache.py b/jedi/cache.py
index ef9a94b9..49ff3fda 100644
--- a/jedi/cache.py
+++ b/jedi/cache.py
@@ -262,7 +262,7 @@ class _ModulePickling(object):
                 parser_cache_item = pickle.load(f)
             finally:
                 gc.enable()
-                
+
         debug.dbg('pickle loaded', path)
         parser_cache[path] = parser_cache_item
         return parser_cache_item.parser
diff --git a/jedi/parser/token.py b/jedi/parser/token.py
new file mode 100644
index 00000000..34f78c5b
--- /dev/null
+++ b/jedi/parser/token.py
@@ -0,0 +1,78 @@
+""" Efficient representation of tokens
+
+We want to have a token_list and start_position for everything the
+tokenizer returns. Therefore we need a memory efficient class. We
+found that a flat object with slots is the best. The Token object is
+that plus indexing and string backwards compatibility.
+
+"""
+
+class Token(object):
+    __slots__ = [
+        "token_type", "token", "start_pos_line", "start_pos_col"
+    ]
+
+    @classmethod
+    def from_tuple(cls, tp):
+        return Token(tp[0], tp[1], tp[2][0], tp[2][1])
+
+    def __init__(
+        self, token_type, token, start_pos_line, start_pos_col
+    ):
+        self.token_type     = token_type
+        self.token          = token
+        self.start_pos_line = start_pos_line
+        self.start_pos_col  = start_pos_col
+
+    def __repr__(self):
+        return "<%s: %s>" % (type(self).__name__, tuple(self))
+
+    # Backward compatibility
+    def __str__(self):
+        return str(self.token)
+
+    # Backward compatibility
+    def __getitem__(self, key):
+        # Builds the same structure as tuple used to have
+        if key == 0:
+            return self.token_type
+        elif key == 1:
+            return self.token
+        elif key == 2:
+            return (self.start_pos_line, self.start_pos_col)
+        else:
+            raise IndexError()
+
+    # Backward compatibility
+    def __setitem__(self, key, value):
+        # setitem analogous to
+        if key == 0:
+            self.token_type       = value
+        elif key == 1:
+            self.token            = value
+        elif key == 2:
+            self.start_pos_line   = value[0]
+            self.start_pos_col    = value[1]
+        else:
+            raise IndexError()
+
+    # Backward compatibility
+    def __getattr__(self, attr):
+        # Expose the missing start_pos attribute
+        if attr == "start_pos":
+            return (self.start_pos_line, self.start_pos_col)
+        else:
+            raise AttributeError(
+                "type '%s' has no attriubte '%s'" % (
+                    type(self),
+                    attr
+                )
+            )
+
+    # Make cache footprint smaller
+    def __getstate__(self):
+        return tuple(self)
+
+    def __setstate__(self, state):
+        for i in range(len(state)):
+            self[i] = state[i]