We study the question of universal provenance for regular path queries in graph databases. Universal provenance was introduced by Green, Karvounarakis, and Tannen in 2007 as the most general form of semiring provenance for positive relational algebra query results. According to this methodology, the universal provenance is to be computed during query evaluation, and any application provenance, such as e.g. the multiplicity of a tuple in the answer, can then be obtained by a homomorphic projection from the universal provenance. The methodology has subsequently been applied to SPARQL-queries and queries expressed in first order and fixed point logics, but for regular path queries only application provenance has been considered. We remedy the situation in this paper, by elucidating the classical theoretical framework and results that enable us to show that certain regular expressions form the universal provenance for regular path queries. In addition, we show experimentally that computing the universal provenance, as opposed to computing the application provenance directly, does not incur any significant overhead.