Schelling (1969, 1971a, b, 1978) observed that macro-level patterns do not necessarily reflect micro-level intentions, desires or goals. In his classic model on neighborhood segregation which initiated a large and influential literature, individuals with no desire to be segregated from those who belong to other social groups nevertheless wind up clustering with their own type. Most extensions of Schelling's model have replicated this result. There is an important mismatch, however, between theory and observation, which has received relatively little attention. Whereas Schelling-inspired models typically predict large degrees of segregation starting from virtually any initial condition, the empirical literature documents considerable heterogeneity in measured levels of segregation. This paper introduces a mechanism that can produce significantly higher levels of integration and, therefore, brings predicted distributions of segregation more in line with real-world observation. As in the classic Schelling model, agents in a simulated world want to stay or move to a new location depending on the proportion of neighbors they judge to be acceptable. In contrast to the classic model, agents' classifications of their neighbors as acceptable or not depend lexicographically on recognition first and group type (e. g., ethnic stereotyping) second. The FACE-recognition model nests classic Schelling: When agents have no recognition memory, judgments about the acceptability of a prospective neighbor rely solely on his or her group type (as in the Schelling model). A very small amount of recognition memory, however, eventually leads to different classifications that, in turn, produce dramatic macro-level effects resulting in significantly higher levels of integration. A novel implication of the FACE-recognition model concerns the large potential impact of policy interventions that generate modest numbers of face-to-face encounters with members of other social groups.