This dissertation is devoted to the development of quantum memories for light. Quantum memory is an important part of future long-distance quantum fiber networks and quantum processing. Quantum memory is required to be efficient, multimode, noise free, scalable, and should be able to provide long storage times for practical applications in quantum communications and beyond. Here I concentrate on solving particular problems of different quantum protocols and find ways for extending the performance of memories and adding new capabilities. I theoretically show that an array of whispering-gallery resonators is capable of being an efficient and noise-free optical memory with an adjustable storage time. The potential for on-chip realization at room temperature makes the scheme attractive for easy implementation. The effect of Raman scattering in echo memory was evaluated experimentally and theoretically. The noise performance of gradient echo memory in Λ configuration proves, that the developed theory is in a good agreement with an experiment. I proposed a mechanism for extending the bandwidth of impedance-matched memories via a white-light cavity effect. The introduced additional dispersion compensates a bandwidth decrease induced by the cavity and hence increases the spectral zone of impedance matching. Theoretically the scheme allows to increase the bandwidth of high efficient storage (>90%) several times without adding extra noise. Finally, I have proposed an architecture of quantum random-access memory for time-bin photons. The architecture consists from a memory unit and a strongly coupled three-level atom. Both of them are placed in their own cavities, which are coupled to each other. The protocol allows to achieve quantum addressing of quantum information stored in the memory with only a single control unit. This is useful for numerous tasks in quantum machine learning.