diff --git a/phiml/math/_tensors.py b/phiml/math/_tensors.py
index 9c5ac2a8..3b9e6a14 100644
--- a/phiml/math/_tensors.py
+++ b/phiml/math/_tensors.py
@@ -1951,6 +1951,57 @@ def assemble_tree(obj: PhiTreeNodeType, values: List[Tensor], attr_type=variable
         return obj
 
 
+def attr_paths(obj: PhiTreeNodeType, attr_type: Callable, root: str) -> List[str]:
+    if obj is None:
+        return []
+    elif isinstance(obj, Tensor):
+        return [root]
+    elif isinstance(obj, (tuple, list)):
+        paths = []
+        for i, item in enumerate(obj):
+            path = attr_paths(item, attr_type, f'{root}[{i}]')
+            paths.extend(path)
+        return paths
+    elif isinstance(obj, dict):
+        paths = []
+        for name, item in obj.items():
+            path = attr_paths(item, attr_type, f'{root}[{name}]')
+            paths.extend(path)
+        return paths
+    elif isinstance(obj, PhiTreeNode):
+        attributes = attr_type(obj)
+        paths = []
+        for attr in attributes:
+            path = attr_paths(getattr(obj, attr), attr_type, f'{root}.{attr}')
+            paths.extend(path)
+        return paths
+    else:  # native tensor?
+        try:
+            return [] if choose_backend(obj) == OBJECTS else [root]
+        except NoBackendFound:
+            return []
+
+
+def attr_paths_from_container(obj: PhiTreeNodeType, attr_type: Callable, root: str) -> List[str]:
+    if isinstance(obj, str) and obj == MISSING_TENSOR:
+        return []
+    elif isinstance(obj, str) and obj == NATIVE_TENSOR:
+        return [root]
+    elif obj is None:
+        return [root]
+    elif isinstance(obj, (tuple, list)):
+        return sum([attr_paths_from_container(v, attr_type, f'{root}[{i}]') for i, v in enumerate(obj)], [])
+    elif isinstance(obj, dict):
+        return sum([attr_paths_from_container(v, attr_type, f'{root}[{k}]') for k, v in obj.items()], [])
+    elif isinstance(obj, Tensor):
+        raise RuntimeError("Tensor found in container. This should have been set to None by disassemble_tree()")
+    elif isinstance(obj, PhiTreeNode):
+        attributes = attr_type(obj)
+        return sum([attr_paths_from_container(getattr(obj, k), attr_type, f'{root}.{k}') for k in attributes], [])
+    else:
+        return []
+
+
 def cached(t: TensorOrTree) -> TensorOrTree:
     from ._sparse import SparseCoordinateTensor, CompressedSparseMatrix, CompactSparseTensor
     assert isinstance(t, (Tensor, PhiTreeNode)), f"All arguments must be Tensors but got {type(t)}"
@@ -2802,17 +2853,57 @@ def specs_equal(spec1, spec2):
 
 
 def save_tree(file: str, obj):
-    tree, tensors = disassemble_tree(obj, cache=False, attr_type=all_attributes)
-    natives, _, specs = disassemble_tensors(tensors, expand=False)
-    np_natives = [choose_backend(n).numpy(n) for n in natives]
-    np.savez(file, *np_natives, tree=tree, specs=specs)
+    tree, tensors = disassemble_tree(obj, False, all_attributes)
+    paths = attr_paths(obj, all_attributes, 'root')
+    assert len(paths) == len(tensors)
+    natives = [t._natives() for t in tensors]
+    specs = [serialize_spec(t._spec_dict()) for t in tensors]
+    native_paths = [[f'{p}:{i}' for i in range(len(ns))] for p, ns in zip(paths, natives)]
+    all_natives = sum(natives, ())
+    all_paths = sum(native_paths, [])
+    all_np = [choose_backend(n).numpy(n) for n in all_natives]
+    np.savez(file, tree=tree, specs=specs, paths=paths, **{p: n for p, n in zip(all_paths, all_np)})
 
 
 def load_tree(file: str):
     data = np.load(file, allow_pickle=True)
-    np_natives = [data[k] for k in data if k not in ['tree', 'specs']]
-    specs = data['specs'].tolist()
-    tensors = assemble_tensors(np_natives, specs)
+    all_np = {k: data[k] for k in data if k not in ['tree', 'specs', 'paths']}
+    specs = [unserialize_spec(spec) for spec in data['specs'].tolist()]
+    tensors = assemble_tensors(list(all_np.values()), specs)
     tree = data['tree'].tolist()  # this may require outside classes via pickle
-    obj = assemble_tree(tree, tensors, attr_type=all_attributes)
-    return obj
+    stored_paths = data['paths'].tolist()
+    new_paths = attr_paths_from_container(tree, all_attributes, 'root')
+    if tuple(stored_paths) != tuple(new_paths):
+        lookup = {path: t for path, t in zip(stored_paths, tensors)}
+        tensors = [lookup[p] for p in new_paths]
+    return assemble_tree(tree, tensors, attr_type=all_attributes)
+
+
+def serialize_spec(spec: dict):
+    from ._sparse import SparseCoordinateTensor, CompactSparseTensor, CompressedSparseMatrix
+    type_names = {NativeTensor: 'dense', TensorStack: 'stack', CompressedSparseMatrix: 'compressed', SparseCoordinateTensor: 'coo', CompactSparseTensor: 'compact'}
+    result = {}
+    for k, v in spec.items():
+        if k == 'type':
+            result[k] = type_names[v]
+        elif isinstance(v, dict):
+            result[k] = serialize_spec(v)
+        else:
+            assert not isinstance(v, type)
+            result[k] = v
+    return result
+
+
+def unserialize_spec(spec: dict):
+    from ._sparse import SparseCoordinateTensor, CompactSparseTensor, CompressedSparseMatrix
+    type_names = {NativeTensor: 'dense', TensorStack: 'stack', CompressedSparseMatrix: 'compressed', SparseCoordinateTensor: 'coo', CompactSparseTensor: 'compact'}
+    lookup = {v: k for k, v in type_names.items()}
+    result = {}
+    for k, v in spec.items():
+        if k == 'type':
+            result[k] = lookup[v]
+        elif isinstance(v, dict):
+            result[k] = unserialize_spec(v)
+        else:
+            result[k] = v
+    return result