The primary reason is that the SystemC Proof-of-Concept (PoC) implementation uses some global data structures, which get constructed during elaboration, which are not fully cleaned up when sc_stop() is called. These data structures are only freed by the OS when the application quits. There have been some efforts improve the PoC implementation in this area, but these are currently stalled. Unfortunately, partially due to bad code examples, bad habits, and laziness, many SystemC models also not cleanly deallocate their dynamically allocated resources, as they still use manual resource allocation using the new operator, but never release that memory using delete as they don't implement a destructor out of laziness. Since the general availability of smart pointers in C++'11, especially std::unique_ptr<T>, managing cleanly dynamic memory management has become much more convenient and there is no justification anymore to write such dirty code. However, there is a lot of legacy code out there making a bad example and hardware engineers often lack a formal education in C++ instead learning the language basically on the job. This will only resolve slowly over time.
@David Black's suggestion to launch your multiple simulations from a separate process implemented using your favourite scripting language or even in C++, is the easiest workaround for the above problems.