7Short-term memory is associated with persistent neural activity without sustained input, arising from the 8 interactions between neurons with short time constants 1,2 . A variety of neural circuit motifs could account 9 for measured neural activity 3-7 . A mechanistic understanding of the neural circuits supporting short-term 10 memory requires probing network connectivity between functionally characterized neurons 8 . We 11 performed targeted photostimulation of small (< 10) groups of neurons, while imaging the response of 12 hundreds of other neurons 9,10 , in anterior-lateral motor cortex (ALM) of mice performing a delayed 13 response task 11 . Mice were instructed with brief auditory stimuli to make directional movements (lick left 14 or lick right), but only after a three second delay epoch. ALM contains neurons with delay epoch activity 15 that is selective for left or right choices. Targeted photostimulation of groups of neurons during the delay 16 epoch allowed us to observe the functional organization of population activity and recurrent interactions 17 underlying short-term memory. These experiments revealed strong coupling between neurons sharing 18 similar selectivity. Brief photostimulation of functionally related neurons produced changes in activity in 19 sparse subpopulations of nearby neurons that persisted for several seconds following stimulus offset, far 20 outlasting the duration of the perturbation. Photostimulation produced behavioral biases that were 21 predictable based on the selectivity of the perturbed neuronal population. These results suggest that ALM 22 contains multiple intercalated modules, consisting of recurrently coupled neurons, that can 23 independently maintain persistent activity. discrete attractors 24 , failed to capture key features of the responses to photostimulation: the sparse and 105 persistent changes in activity (Fig. 4a, Extended Data Fig. 10). 106Imposing sparse, local connectivity on such models did not produce sparse, persistent activity (Fig. 4b). 107We tested a 'modular' network, with strong within-module connectivity and weak between-module 108 connectivity (Fig. 4c), in which each module can independently maintain persistent activity. Such a 109 network as a whole produced sparse persistent activity consistent with experiments. 110 111 | |/70 , i j i j w e --=