We propose a characterization of provability in BI's Pointer Logic (PL) that is based on semantic structures called resource graphs. This logic has been defined for reasoning about mutable data structures and results about models and verification have been already provided. Here, we define resource graphs that capture PL models by considering heaps as resources and by using a labelling process. We study provability in PL from a new calculus that builds such graphs from which proofs or countermodels can be generated. Properties of soundness and completeness are proved and the countermodel generation is studied.