a, The size of n=94 independent ensembles of simultaneously recorded neurons varies between 1-6. b, Fraction of neural ensembles with significant previous stimulus decoding accuracy (z > 1.96, see Methods) computed for all ensembles (dashed line) and only for those ensembles with strongest previous stimulus code averaged across the whole delay (see Methods). The incidence of stimulus decoding was significant in delay and reactivation, but not at ITI (two-sided binomial test at p=0.05, with n=94 and n=27 ensembles, for ‘all ensembles’ and ‘highest delay code’, respectively). Error bars are bootstrapped ±s.e.m. c, across-ensemble Pearson correlation between delay decoding accuracy (averaged in the entire delay) and decoding accuracy at different time points (two-sided p-values: 6.5e-30, 0.87, 0.035, n=94 ensembles). The ensembles with strongest delay code also had stronger decoding during reactivation, demonstrating the neural association between delay representations and reactivations despite absent code in the ITI. Error bars denote ±s.e.m. computed with a bootstrap procedure. d, Individual ensemble values from c, orange (Pearson correlation, two-sided p=0.035, n=94 ensembles).