Acute infection with murine cytomegalovirus (mCMV) is controlled by CD8+ T cells and develops into a state of latent infection, referred to as latency, which is defined by lifelong maintenance of viral genomes but absence of infectious virus in latently infected cell types. Latency is associated with an increase in numbers of viral epitope-specific CD8+ T cells over time, a phenomenon known as “memory inflation” (MI). The “inflationary” subset of CD8+ T cells has been phenotyped as KLRG1+CD62L- effector-memory T cells (iTEM). It is agreed upon that proliferation of iTEM requires repeated episodes of antigen presentation, which implies that antigen-encoding viral genes must be transcribed during latency. Evidence for this has been provided previously for the genes encoding the MI-driving antigenic peptides IE1-YPHFMPTNL and m164-AGPPRYSRI of mCMV in the H-2d haplotype. There exist two competing hypotheses for explaining MI-driving viral transcription. The “reactivation hypothesis” proposes frequent events of productive virus reactivation from latency. Reactivation involves a coordinated gene expression cascade from immediate-early (IE) to early (E) and late phase (L) transcripts, eventually leading to assembly and release of infectious virus. In contrast, the “stochastic transcription hypothesis” proposes that viral genes become transiently de-silenced in latent viral genomes in a stochastic fashion, not following the canonical IE-E-L temporal cascade of reactivation. The reactivation hypothesis, however, is incompatible with the finding that productive virus reactivation is exceedingly rare in immunocompetent mice and observed only under conditions of compromised immunity. In addition, the reactivation hypothesis fails to explain why immune evasion genes, which are regularly expressed during reactivation in the same cells in which epitope-encoding genes are expressed, do not prevent antigen presentation and thus MI. Here we show that IE, E, and L genes are transcribed during latency, though stochastically, not following the IE-E-L temporal cascade. Importantly, transcripts that encode MI-driving antigenic peptides rarely coincide with those that encode immune evasion proteins. As immune evasion can operate only in cis, that is, in a cell that simultaneously expresses antigenic peptides, the stochastic transcription hypothesis explains why immune evasion is not operative in latently infected cells and, therefore, does not interfere with MI.