The present study explored the origins of word learning in early infancy. Using event‐related potentials (ERP) we monitored the brain activity of 3‐month‐old infants when they were repeatedly exposed to several initially novel words paired consistently with each the same initially novel objects or inconsistently with different objects. Our results provide strong evidence that these young infants extract statistic regularities in the distribution of the co‐occurrences of objects and words extremely quickly. The data suggest that this ability is based on the rapid formation of associations between the neural representations of objects and words, but that the new associations are not retained in long‐term memory until the next day. The type of brain response moreover indicates that, unlike in older infants, in 3‐month‐olds a semantic processing stage is not involved. Their ability to combine words with meaningful information is caused by a primary learning mechanism that enables the formation of proto‐words and acts as a precursor for the acquisition of genuine words. Three‐month‐old infants learned pairings of eight novel objects and eight novel words within only four consistent repetitions, but they did not retain them in long‐term memory until the next day. In the memory test infants relearned the combinations, even though objects and words were paired consistently in only half of their overall pairings. Instead of the N400 component that evidences the learning of word meanings in the ERP of older infants, a late left‐hemispheric negativity indicated the learning of object‐word pairings in 3‐month‐olds. This result represents the first neurophysiological distinction between ‘higher’ lexical‐semantic learning and associative ‘proto‐word’ learning.