The fascinating question of the relation of information and coding theory to the memories stored in the brain is our research scope. We speculate there is a similar code used to represent different memories, rather than unique code for different memories. The uniform cortex structure supports our speculation. Recently we suggested holographic coding that can fit Pribram's holographic memory theory. Using the holographic coding metaphor, the memory should be retrieved by a reference beam as in a hologram. We explore the possibility that the brain learns its directory (possibly in the temporal lobe), during memory consolidation. This directory is a neural network that is used for sending signals to the cortex to recall memories. The network learns to distinguish between objects during saving, in order to signal the correct recall. Haar features (HF) are 0/1 matrices used for face recognition. We use HF to learn to differentiate between objects. Namely when objects are saved, our system learns what is the best set of HF to distinguish between them using a genetic algorithm. The sets of HF are tested for the best clustering set without knowing their semantics (unsupervised learning). Later semantics is learned by interaction with the environment. The best sets continue to the next generation. We chose unsupervised learning due to the idea that it is possible to distinguish objects without knowing their identity.