A taxonomy of long-term memory systems together with specific brain structures involved in each system (adapted from ref. 66).

The serial reaction time task as presented on a computer (22). Four dashes appear continuously at the bottom of the screen to denote four possible locations of an asterisk (A, B, C, or D). During training, the asterisk appears sequentially, moving from one to another of the four locations. Subjects respond to each appearance of the asterisk as rapidly as possible by pressing a key directly beneath the cue. Five hundred milliseconds after each response, the asterisk appears at a new location. Unbeknownst to the subject, a sequence of 10 locations (e.g., DBCACBDCBA) repeats every 10 trials throughout 400 training trials—i.e., there are 40 repetitions of a 10-trial sequence. Learning is demonstrated by gradually improving reaction times when the asterisk appears in the repeating sequence of locations, as compared with reaction times when a random sequence of locations is presented.

Prototype learning (27). Examples of the 40 study items and 84 test items used to study prototype learning. The study items are all distortions of a prototype (average) dot pattern that is never presented. For training, the 40 study patterns are presented for 5 sec each, and the subject points to the dot closest to the center of the pattern (to guarantee attention). Five minutes later, subjects are instructed that the dot patterns they have just seen all belong to a single category of patterns in the same sense that, if a series of different dogs had been presented, they would all belong to the category “dog.” Then for each of 84 new dot patterns subjects judge (yes/no) whether or not it belongs to the same category as the training patterns. The test items consist of four repetitions of the prototype, 20 new “low” distortions of the prototype, 20 new “high” distortions of the prototype, and 40 random dot patterns.

Schematic view of the medial temporal lobe memory system. The entorhinal cortex is a major source of projections to the hippocampal region (which includes the dentate gyrus, the cell fields of the hippocampus, and the subicular complex). Nearly two-thirds of the cortical input to entorhinal cortex originates in the adjacent perirhinal and parahippocampal cortices, which in turn receive projections from unimodal and polymodal areas in the frontal, temporal, and parietal lobes. The entorhinal cortex also receives other direct inputs from orbital frontal cortex, insular cortex, and superior temporal gyrus. All these projections are reciprocal.

Probabilistic classification learning: the weather prediction task (2). Subjects decide on each trial which of two weather outcomes (rain or sunshine) will occur based on a set of one, two, or three cues (out of four possible cues) that appear on a computer screen. Each cue is independently associated to a weather outcome with a fixed probability, and the two outcomes occur equally often. There are four possible cue–outcome association strengths: throughout training, a cue is associated either 75%, 57%, 43%, or 25% (approximately) with sunshine. During each trial, one, two, or three of the four possible cues are presented. Subjects respond by pressing a key to predict the weather outcome, and feedback is given immediately after each choice (correct or incorrect).

Artificial grammar learning (24). Letter strings are generated from a finite state rule system. Grammatical letter strings can be formed by traversing the diagram from the in arrow to the out arrows, adding a letter at each transition from one node to the next. In a typical experiment, 23 grammatical items are used for training, and a different 23 items are used for testing. An additional 23 nongrammatical test items are also used as foils for testing. These are generated by introducing an error in each of 23 different grammatical items. Subjects first study 23 grammatical letter strings one at a time. Five minutes later, they are informed for the first time that the letter strings they have just seen were formed by a set of rules. They are told that their task is to classify new letter strings according to whether they appear to conform to these rules. The 46 test items are then displayed one at a time, and subjects judge the item to be correct or incorrect (grammatical or nongrammatical).

Each panel shows the results for four control subjects tested six times each (open bars) and the results for the profoundly amnesic patient EP (average of six tests; filled bars). (A) Classification of 84 novel dot patterns after studying 40 different training patterns (see Fig. 5 caption). Control subjects and EP performed similarly, endorsing the test items as a function of how closely they resembled the prototype of the training category (27). (B) Exactly the same task as in A but now with instructions to recognize the dot patterns that had been presented before (i.e., subjects made yes/no judgments). Actually, none of the 40 training patterns appeared on the test. Instead, the 84 test patterns varied in their resemblance to the training patterns as in A. (C) Classification of 84 novel dot patterns after studying a single dot pattern presented 40 times in succession. The training pattern was a prototype dot pattern, and the test patterns consisted of four repetitions of the prototype, 40 low distortions of the prototype, 20 high distortions of the prototype, and 40 random dot patterns. The instructions were as in A. (D) Exactly the same task was presented as in C, but now with instructions to recognize the dot patterns that had been presented before, as in B. Actually, only one dot pattern had been presented during training. The test items consisted of four repetitions of this same pattern and 80 other patterns that varied in their resemblance to the training pattern. A four-way ANOVA (EP versus controls, classification versus recognition instructions, 40 different study items versus one study item, and four types of test item) revealed significant effects of group, type of study item, and type of study item (F_s > 17.0, P_s < 0.002), but the effect of instructions fell short of significance [F(1,103) = 4.7, P = 0.06]. EP performed entirely normally at classification after seeing 40 different training patterns (A), but he performed significantly worse and at chance when he had to recognize a single pattern presented 40 times in succession (D). When asked to recognize 40 stimuli that had been presented once each (B), both EP and control subjects tended to use a classification strategy. When asked to classify after seeing only one pattern 40 times (C), normal subjects tended to rely on declarative memory, but EP could not perform the task. Subjects were more influenced by the kinds of material they studied than by the instructions given at test. Classification learning can proceed nondeclaratively when there is some variability in the training stimuli (A and B).

Mean z scores based on data from four measures of memory (47) for ten normal monkeys (N), five monkeys with surgical aspiration lesions of the parahippocampal cortex (the PH lesion; cortical areas TH and TF; ref. 52), four monkeys with surgical stereotaxic damage limited to the hippocampal region (H; dentate gyrus, the cell fields of the hippocampus, and the subicular complex), and five monkeys with surgical aspiration lesions of perirhinal cortex (PR; cortical areas 35 and 36; ref. 53). The PR group was impaired relative to the N, PH, and H groups. The PH group was not impaired (51). Error bars indicate standard errors of the mean.