Background: The distinct immunologic functions performed by CD4 T cells of central memory (CM) and effector memory (EM) phenotypes require distinct gene expression programs that could differentially affect HIV replication. We tested the difference between CM and EM cells as HIV hosts in vivo by quantifying and sequencing HIV genomes from these cellular subsets in a group of HIV-infected donors with spontaneous control of plasma HIV-1 RNA levels to <1,000 copies/mL in the absence of cART. Methodology: CM and EM cells were sorted from PBMC of 6 HIV controllers by FACS. RNA genomes were extracted from plasma virions and reversetranscribed. Env gene fragments were quantified and cloned from cell-associated DNA or virion cDNA using a novel, high-throughput single-genome amplification method termed fluorescence-assisted clonal amplification (FCA). Clonal sequences were checked for hypermutation using the Hypermut algorithm. Coreceptor specificities of clonal sequences were predicted using the Geno2pheno algorithm. Maximum-likelihood phylogenetic trees were reconstructed from alignments of non-hypermutant clonal sequences using the general time-reversible model of substitution with gamma distribution. Compartmentalization was assessed by Slatkin-Maddison testing. Results: Levels of HIV genomic DNA were indistinguishable in CM and EM cells from HIV controllers. By contrast, env DNA sequences from these two subsets diverged markedly, with genetic compartmentalization between CM and EM in all 6 controllers studied. No clonal sequence from any controller was predicted to be X4-tropic, and few sequences were hypermutant (CM median 3.9% of all sequences per donor, range 0-22.7%; EM median 2.15%, range 0-9.5%). Sequences from CM cells were more closely related to sequences from plasma virions than were sequences from EM cells, with CM/plasma compartmentalization in 3/6 controllers and EM/plasma compartmentalization in 6/6. Average genetic distances between groups of clonal sequences were smaller between CM and plasma than between EM and plasma. Phylogenetic trees from controllers but not from two untreated subjects with viral loads >20,000 copies/mL showed large clusters of matching sequences, often separated by large genetic distances. Several large clusters represented HIV clones with truncations in gp120, suggesting non-viable proviruses amplified by T cell clonal expansion. Conclusions: The population of productively infected CD4 T cells in HIV controllers is very small. This is particularly true in the EM subset, where most cells harboring HIV DNA appear to have expanded from a few non-productively infected precursors. The key focus of ongoing HIV replication in controllers may be a tissue site containing CM-like cells.