Some additional things that might be worth doing. I believe this exposes enough to allow people to build an object graph walker out of the `pickle`/`copy` protocol without having to access fragile internals, and without restricting future evolution of the internals of the protocol. See [the same -ideas thread][1] again for details on how it could be used and why. These would all be changes to the `copy` module, together with the changes to `pickle` and `copyreg` in the previous message: class Memo(dict): """Memo is a mapping that can be used to do memoization exactly the same way deepcopy does, so long as you only use ids as keys and only use these operations: y = memo.get(id(x), default) memo[id(x)] = y memo.keep_alive(x) """ def keep_alive(self, x): self.setdefault(id(self), []).append(x) def reconstruct(x, memo: Memo, reduction, *, recurse=deepcopy): """reconstruct(x, memo, reduction, recurse=recursive_walker) Constructs a new object from the reduction by calling recursive_walker on each value. The reduction should have been obtained as pickle.reduce(x) and the memo should be a Memo instance (which will be passed to each recursive_walker call). """ return _reconstruct(x, memo, *reduction, deepcopy=recurse) def copier(cls): """copier(cls) -> func Returns a function func(x, memo, recurse) that can be used to copy objects of type cls without reducing and reconstructing them, or None if there is no such function. """ if c := _deepcopy_dispatch.get(cls): return c if issubclass(cls, type): return _deepcopy_atomic Also, all of the private functions that are stored in `_deepcopy_dispatch` would rename their `deepcopy` parameter to `recurse`, and the two that don't have such a parameter would add it. [1]: https://mail.python.org/archives/list/python-ideas@python.org/thread/RTZGM7L...