Since the early 1900s it has been suspected that long-term memory is somehow encoded in the strengths of synaptic connections between nerve cells. It is known that at least in specific cases such strengths can remain unchanged for at least hours or more, but can immediately change if connected nerve cells have various patterns of simultaneous excitation. The changes that occur appear to be associated changes in ionic channels in cell membranes and sometimes with the addition of new synapses between cells.
Observations suggest that in humans there are several different types of memory, with somewhat different characteristics. (Examples include memory for facts and for motor skills.) Usually there is a short-term or so-called working component, lasting perhaps 30 seconds, and typically holding perhaps seven items, and a long-term component that can apparently last a lifetime. Specific parts of the brain (such as the hippocampus) appear necessary for the long-term component to form. In at least some cases there is evidence for specialized areas that handle particular types of memories. When new data is first presented, many parts of the brain are often active in processing it. But once the data has somehow been learned, only parts directly associated with handling it usually appear to be active.
Memories often seem at some level to be built up incrementally, as reflected in smooth learning curves for motor skills. It is not clear whether this is due to actual incremental changes in nerve cells or just to the filling in of progressively more cases that differ in detail.
Experiments on human learning suggest that a particular memory typically involves an association between components from several sensory systems, as well as emotional state.
When several incomplete examples of data are presented, there appears to be some commonality in the character of generalizations that we make. One mathematically convenient but probably unrealistic model studied in recent years in the context of computational learning theory involves building up minimal Boolean formulas consistent with the examples seen.