Dear all, I was trying to use the Boost Serialization library again, but I ran into a problem which cost me a full day. The Boost Serialization library is a very handy library, but often when I use it I run into problems which cost me days. This time I try to serialize just 'track_never' objects. First of all the compiler firewall which is built in the library does not work when using the nvp macro: void Foo() { boost::archive::xml_oarchive oa(...) SomeStruct s; oa << s; // compile time error oa << BOOST_SERIALIZATION_NVP(s); //ok } But this wasn't the real problem. I was looping to serialize non tracking objects: void Bla() { boost::archive::xml_oarchive oa(...) for (int i = 0; i < 1000; ++i) { const SomeStruct s; oa << s; } } However the library did track them by pointer due to a previous request to serialize a pointer. It decides this on a call to 'basic_serializer::serialized_as_pointer' deep in the library. struct SomeStructHolder { SomeStructHolder() : m_p(NULL){} template <class Archive> void serialize(Archive& ar, const unsigned int /*version*/) { ar & BOOST_SERIALIZATION_NVP(m_p); } const SomeStruct* m_p; }; somewhere else: void Bla() { boost::archive::xml_oarchive oa(...) const SomeStructHolder holder; oa << holder; } So this is again an unexpected property of the library. Maybe the ambition for automatic object tracking is too high, and specify it explicitly per class / per archiving operation would be more work for the user, but also clearer. And still the original problem was too find out if Boost Serialization could also be used on a per demand loading of (value-)objects: in my problem the final storage can be potential bigger than computer memory, so I was wondering if objects could be loaded on demand (instead of all in once). wkr, me