Physiological studies reveal a dichotomy in biological Hebbian learning: NMDA receptors are utilized for induction of long term potentiation (LTP) whereas AMPA is used for LTP expression. We propose that this dichotomy would have functional value: preventing previously stored memories from interfering with the storage of new memories. A previous hypothesis reduces this interference by temporarily reducing associative weights during learning. Complementary to this model, we propose a dual transmission algorithm in which one set of synaptic weights are used primarily for learning and another primarily for recall. This algorithm shows good performance in a simple neural network model. Biologically, the model could be mediated by a cholinergic switch from dominance of learning-insensitive NMDA receptors to dominance of learning-modifiable AMPA receptors.