It has long been hypothesized that hippocampus encodes memory traces of episodes into cell assemblies. Accumulating evidence suggests that this encoding process involves two fundamental processes, that is, pattern separation by dentate gyrus and pattern completion by the recurrent neuronal circuit of CA3. Pattern separation is considered to be essential for the discrimination of the contextual information on episodes, whereas pattern completion retrieves the memory trace of each episode, or the corresponding cell assembly, from a given cue for memory recall. Several important findings were recently made about the structure and function of both neural circuits. For example, adult neurogenesis was found to be crucial for pattern separation in dentate gyrus. However, how neurogenesis contributes to this processing is unknown. Pattern completion has been extensively studied in computational models of recurrent neural circuits. However, it was recently found that the recurrent circuit of CA3 has a long-tailed, typically lognormal, distribution of EPSP amplitudes. Such a distribution of synaptic weights was unexpected in previous models of associative memory. In this talk, I will discuss the computational implications of these novel findings for pattern separation and pattern completion by hippocampal circuits. I will demonstrate that we can model neurogenesis in pattern separation based on pattern classification theories known in machine learning. I will show how lognormal synaptic connections improve the memory storage ability of CA3 circuit and how log-STDP and short-term depression may jointly generate stable representations of memory traces with cell assemblies. The results presented were obtained primarily through collaborations with Anthony DeCostanzo and Naoki Hiratani.