The Wikipedia article on epistasis has a section on epistasis within genes, which it terms intragenic complementation:
Just as mutations in two separate genes can be non-additive if those genes interact, mutations in two codons within a gene can be non-additive. In genetics this is sometimes called intragenic complementation when one deleterious mutation can be compensated for by a second mutation within that gene.
It appears that many interesting examples of this phenomenon occur when the protein forms multimeric complexes. A multimer consisting of monomers with different mutations may function where a homogenous multimer of monomers of only one of the mutations might be inactive. See Turner et al., (1997) PNAS 94:9063-9068. The Wikipedia article mentions some examples of this under the term interallelic complementation or heterozygotic epistasis.
This reminded me of the phenomenon of protein coevolution that is currently being used to try to leverage evolutionary information to solve protein structures. When the homologous sequences of a protein are known in many different species, then a reasonable assumption is that mutations in protein residues that are in close spatial proximity will correlate, e.g. if a lysine is switched to an glutamate, then other nearby residues may mutate to accommodate the change in ambient electric charge. For more information on this effect, two very interesting papers are Hopf TA, Colwell LJ et al., (2012) Cell 149(7):1607-1621, where the sequences of a family of membrane proteins are used to predict their structure, and Skerker et al (2008) Cell 133:1043-1054, where the covariation between residues in a complex of two proteins is used to re-wire the specificity of one protein for another.