Hi, We have a simple data structure like this, that we would like to expose via boost python: struct Layer { double value; }; struct Document { std::vector< Layer > layer; }; If we were writing this in pure python, then it could be written as follows: class Layer: def __init__( self ): self.value = 0.0 class Document: def __init__( self ): self.layers = [] We would like our C++ bindings to match the behaviour of python as closely as possible. But we have encountered a scenario in which it doesn't, and we wanted to get some advice on the best way forward. In pure python, the following code produces an output of '123.0': l = Layer( ) d = Document( ) d.layers.append( l ) l.value = 123.0 print d.layers[0].value When exposing C++ classes via boost python (using the vector_indexing_suite for Document.layers), it produces a result of '0.0'. The 'append' is making a copy such that 'l' and 'd.layers[0]' are two separate objects. We've noticed that one way to make this work is by using shared_ptr in our C++ data structure, but we feel we would have to compromise on our C++ API as a result. Do we have any other options? Thanks, James.