Introduction

Since the 1940s many theorists have suggested that information might be stored in the cognitive system in a different way over short time periods than for longer time periods. A number of possible mechanisms have been proposed – special or general-purpose buffers, the continuing but temporary activation of the structures that have just processed an input, and the formation of temporary associations or temporary changes in association strength. Theoretical arguments for such mechanisms have been produced from a range of scientific fields from physiological psychology (e.g., Hebb, 1949), information-processing psychology (Broadbent, 1958), and symbol-processing artificial intelligence (Newell & Simon, 1972) through to connectionist neuroscience (e.g., Crick, 1984; Hinton & Plaut, 1987). Empirical support has come from a narrower set of approaches. Most of the relevant findings have come from experimental psychology (e.g., Waugh & Norman, 1965; Atkinson & Shiffrin, 1968; Glanzer, 1972; Baddeley & Hitch, 1974), but these have been subject to many criticisms (see, e.g., Craik & Lockhart, 1972; Crowder, 1982).

From a neuropsychological perspective, if a short-term memory trace were being carried by certain of the aforementioned mechanisms, a selective deficit to that mechanism would be realized in a selective impairment in behaviour. This is clearly true for the most favoured possibility – damage to a specific short-term memory buffer - given that the buffer was not involved in all cognitive operations. It would also apply for a specific impairment of one of two association-weight-changing mechanisms. For instance, Hinton and Plaut (1987) have simulated a connectionist network using two types of weights – “slow” weights (which change slowly and carry the long-term learning) and “fast” weights (which change rapidly but decay fairly quickly to zero).