This paper reviews the skilled memory theory and relevant findings (see summary of Ericsson & Staszewski, 1989, for a summary). Briefly, a positive effect of experience on memory for domain-specific information has been demonstrated in a wide range of domains. For example, Egan and Schwartz demonstrated the skilled memory effect using diagrams of circuits. Akin showed that architects recall building plans in terms of ordered patterns. Shneiderman showed that expert computer programmers had superior memory for lines of code from a meaningful, but not from a nonsense, FORTRAN computer program. Why are these findings important? After all, it only makes sense that experts would have better memory for domain-specific material than novices.
As discussed in Ericsson and Staszewski (1989), the skilled memory theory consists of three basic principles: 1) information is encoded into LTM in terms of prior knowledge, 2) a separate LTM knowledge structure (a meta-structure) is used to keep track of the order in which the information was presented, and 3) encoding and retrieval speedup with practice.
A number of important themes were developed in research on skilled memory theory.
First, exceptional memory (e.g., for dinner orders) was explained without violating assumptions about the capacity of STM. For example, verbal reports from the mnemonist SF indicated that he held no more than about 7 numbers in memory at any given time. Specifically, he processed semantically groups of 3-4 numbers—that is, he encoded them as a running time—while temporarily storing the following numbers. Second, and related, research on skilled memory emphasized that exceptional skill is acquired through practice.
Second, research on skilled memory provided a mechanism to account for experts’ rapid processing of information. The argument was that information is rapidly encoded and retrieved through use of hierarchical knowledge structures called retrieval structures. According to the skilled memory theory, it takes extensive practice to refine this knowledge structure. As Chase and Ericsson explain, "we assume that it takes practice, extensive practice, to use [a retrieval structure] . . . and that practice involves learning to generate a set of distinctive features to generate a set of distinctive features to differentiate one location from another" (p. 27). From this view, skilled memory is an essential aspect of expertise, and skilled memory is predicated on organized knowledge structures, which develop through practice.
Skilled Memory Theory
Three questions are addressed: 1) What is the structure of LTM?; 2) What are the storage and retrieval mechanisms that operate on this structure; and 3) What is the role of retrieval structures in skilled memory?
1. Hierarchical Organization
A hierarchical structure of LTM is assumed. LTM can be represented as a network of nodes. For example, DD’s knowledge of running times is represented as a hierarchy in which the most abstract level of organization is running time. The next level of organization includes labels such as "1 Mile" and "2 Mile." The "1 Mile" label subsumes finer level nodes such as "Near New Barrier," which in turn branches into "Coe & Ovett (346-349)." Finally, "Coe & Ovett" branches into two terminal nodes: "New World Recod (347)" and "John Walker (349)." An assumption of this model is that search to a terminal node results in a link between that node and the contents of STM. In sum, the contents of STM result in LTM activation. When the activation spreads to a terminal node, a link between the contents of STM and that knowledge is established.
2. Retrieval Structures
Retrieval structures are abstract, hierarchical knowledge structures used to keep track of the order in which information is presented. Retrieval structures are also hierarchically organized, and, again, a link is established between the contents of short-term memory and a terminal node when that node is activated, that is, when the node and the contents of short-term memory are activated at the same time. Chase and Ericsson conceptualize a retrieval structure as "a featural description of a location that is generated during encoding . . . " (p. 27). The important point is that these descriptions are linked to, encoded with, the digits during encoding. The descriptions then serve as retrieval cues for recall.
Finally, attended information (e.g., digits) is automatically bound to the context: the day, the time, the length of the list, etc.
Short-Term Memory and Attention
Short-term memory is defined as the currently active subset of LTM nodes, or knowledge structures. Attention is equated with the process whereby the contents of STM are linked with LTM knowledge structures. Attention is limited and capacity; this limitation corresponds to a limitation on the number of LTM knowledge structures that can be activated.
The model posits a featural representation system in which the contents of STM and activated semantic features are bound ("chunked together") by virtue of simultaneous activation. An episodic event (e.g., a group of numbers) is associated with semantic information. The contents of STM are linked with 1) a mnemonic code, 2) a retrieval structure, and 3) the context. Location information is included in the memory trace and provides location information to be used during retrieval. As Chase and Ericsson explain, "What the semantic code does is narrow the search in long-term memory for the memory trace" (p. 30).
Chase and Ericsson argue that retrieval begins through activation of the current context and first location in the retrieval structure. Recall that context information is linked with encoded information. In addition, the first location in the retrieval structure can be intentionally activated given that a retrieval structure is a permanent, LTM structure. The spreading of activation that ensues begins the retrieval process.
When two groups of information occupy the same semantic category, updating must occur to differentiate them. For example, 4054 and 4062 occupy the same semantic category: good college times. Chase and Ericsson argue that when this happens, "a new hierarchical memory trace is formed" (p. 34).
Chase and Ericsson claim that forgetting occurs because of loss of order information—that is, confusion among similar locations in the retrieval structure—and not because of a weakening of semantic links in the knowledge structure. For example, in after-session recall, SF and DD showed the best recall for digits they had the most difficult time with during after-trial recall. "These data clearly suggest that the buildup of proactive interference over trials is due to a loss of connections between the location in the retrieval structure and the memory trace, because the memory trace is clearly accessible through the semantic code" (p. 38).
Chase and Ericsson’s theory has implications for the construct of working memory. (These implications were eventually elaborated in Ericsson and Kintsch’s theory of long-term working memory.) In brief, they asserted that working memory should be reconceptualized to include retrieval mechanisms that expand the functional capacity of working memory. They state: "Retrieval structures provide direct access to knowledge structures, and they provide relatively fast access . . . , thus avoiding the difficulties normally associated with long-term memory retrieval (such searches take a lot of time and cause interference by activating competing knowledge structures)" (p. 41). Retrieval structures are viewed as a kind of intermediate-term memory. They provide a context for STM and WM processes to work within. For example, "The working memory of good readers is expanded . . . because they have developed better structures for organizing and retrieving information of various types relevant to the comprehension process from semantic memory during the reading process" (p. 42).
Studies on expert mental calculators, waiters, and skill in reading provide evidence for skilled memory theory. Studies of mental calculators and memory for dinner orders are discussed above. Briefly, the reported research on memory for sentences showed that people have better memory for words arranged into meaningful sentences than for the same words presented in a random order. What mediates this advantage? Words are not necessarily encoded as individual units. For example, Ericsson and Karat found that subjects had better memory for some long sentences than for much shorter sentences. Furthermore, errors tend not to affect the meaning of the sentence. Finally, when presented a cue word, subjects recalled nearly all of the words from sentences, but very few words from word scrambles. As Chase and Ericsson explain, "This clearly suggests to us that a single cue word was able to access an integrated relationship rather than just a single chunk or unit" (p. 54).
Summary, Questions, and Comments
The goal of this article is to show how people can use knowledge to achieve superior performance. That is, "The major theoretical point we wanted to make here is that one important component of skilled performance is the rapid access of a sizable set of knowledge structures that have been stored in directly retrievable locations in long-term memory" (p. 55). According to skilled memory theory, retrieval structures increase the functional capacity of working memory by enabling fast access to knowledge structures in LTM. A revised model of working memory is provided accordingly. This model includes three basic components: 1) short term memory, 2) intermediate term memory consisting of domain-specific knowledge structures, and 3) context.