Mutualistic interactions are ubiquitous and ecologically important in nature. Though theory predicts that cheaters should sweep through populations and degrade mutual benefits, empirical evidence suggests that mutualisms are long-lived and highly variable in nature. Context-dependence and partner choice are two mechanisms that may stabilize mutualisms against the invasion of cheaters. I sampled within and among natural populations of the model legume Medicago truncatula and its symbiont Sinorhizobium meliloti, in order to investigate the ecological and evolutionary dynamics in plant and rhizobium populations that share a coevolutionary history. Using a combination of quantitative genetic and molecular techniques, I found significant genotype-by-genotype interactions for both plant and rhizobium fitness – indicating that for neither plants nor rhizobia is there one universally best partner genotype. In addition, I found evidence for plant partner choice, i.e. that plants associate more often with the most beneficial rhizobia, and that “choice” differed among plant genotypes. Finally, I asked whether plant and rhizobia populations are locally co-adapted, and found little overall evidence for this; instead, genotypes within populations varied in their response to symbiosis with local, versus distant, partner genotypes. My work suggests that genotype-dependence (GxG) for partner quality as well as selection via partner choice may be important for the persistence of mutualisms in nature. These data also provide evidence that plant-rhizobium mutualisms are dynamic and highly variable, as opposed to static interactions which result from the fixation of local genotypes for optimal mutual benefit.