Most computation in the neocortical local circuits is performed by the formation and segregation of cell assemblies. How neurons interact each other flexibly by their synaptic connections in not well understood. Here we investigated the local network activity in rat prefrontal cortex (PFC) during an odor-based working memory task. We recorded simultaneously ～100 units in PFC, using 64-site two-dimensional silicon probes. Approximately 80% of PFC neurons had either cue-stimulus and/or behavior selectivity. Cross-correlation analysis revealed that about 0.2% cell-pairs has robust interactions in 2～3 ms delay (i.e. monosynaptic interactions). Importantly, the spike transmission probability between the presynaptic and postsynaptic neurons varied as a function of behavior and/or the phase of the task. Statistical analysis revealed that these behavior-dependent changes of monosynaptic efficacy are often independent of the firing rates of the pair neurons. Spike transmission efficacies often depended on the spiking history of the presynaptic neurons, indicating that short-term synaptic plasticity (synaptic facilitation / depression) might contribute to the dynamic modification of functional connectivity. Furthermore, coincident firing of two or more neurons nonlinearly increased the efficacies of spike transmission. These results indicate that the efficacy of monosynaptic connections in the cortex is modified flexibly and dynamically, underlying cell assembly formations in fine time scale.