Hippocampal networks link temporally contiguous memories through spiking sequences that encode sensory cues (external world) and time between them (internal representations), constructing ‘memory-maps’ of related experiences. It was recently shown that such sequences combine particularly stable representations of sensory cues with highly dynamic, learning- dependent temporal-codes. This combination of stability and flexibility allows linking fixed elements of the external world across variable time intervals. But how does feedback inhibition by parvalbumin- (PV) and somatostatin-expressing (SOM) interneurons (INs) control the shaping and stability of these multi-modal representations? First, we employed two-photon calcium imaging in vivo in the CA1 area of head-fixed mice, to track the activity of non-specific GABAergic INs across days, while mice learned and performed an olfactory delayed non-match-to-sample task (DNMS) requiring working memory. We found that a subset of INs yielded significant odor- and time-fields, collectively forming odor-specific sequences. These ‘odor-INs’ and ‘time-INs’ had lower odor-selectivity and noisier activation than pyramidal cells (PY), yielding poor cue-decoding performance. Moreover, odor-IN had more stable fields than time-IN, but overall, much more unstable ones than PY fields across days. Importantly, unlike PYs, the number of time-cell INs did not increase during learning of the task. Secondly, through ultra-fast, voltage imaging in vivo on transgenic mice, we recorded single action potentials and subthreshold membrane dynamics from multiple PV and SOM INs during DNMS performance. We followed the same cells across multiple days, before and after DNMS training. We confirmed that both PV and SOM INs yielded odor-fields with reduced odor-specificity. Moreover, increased spiking during odor-presentation was accompanied by an increase in intracellular theta and beta oscillations in subthreshold traces, which was stronger for PV than SOM INs. Therefore, PV and SOM IN ensembles provide timed but not cue-specific inhibition. They thus increase the signal-to-noise-ratio of PY sequences by silencing the non-cue-specific PY population.