Developing Intelligence

Reworking Working Memory

Memory, defined as “any lasting effect of experience,” is an overly broad term. Those with damage to the hippocampus lose their long-term memory but retain the ability to maintain conversations (at least for short periods of time). But new perspectives on the nature of short-term or “working” memory suggests that such a neat division between memory systems is overly simplistic.

In the current issue of Psychological Review, Unsworth & Engle argue that working memory is best understood as an “override” mechanism serving two functions: 1) the active maintenance of novel or important information in the face of interference (either internal or external), and 2) discrimination and retrieval of appropriate information by cue- or context-specified search of memory.

Unsworth & Engle suggest that working memory fulfills these functions through a two-part architecture: 1) a “primary” memory system based on active maintenance of information with a small but dynamic capacity limit (maxing out at around 4 items) and the ability to protect these maintained items from proactive interference, and 2) a “secondary” memory system which stores items that have been displaced from primary memory, along with the context in which they were processed, which can later serve as a retrieval cue. Unsworth & Engle suggest that information is kept in primary memory only as long as attention is actively directed at those items; once attention is removed, the items are displaced into secondary memory.

Capacity limits in working memory therefore result from individual differences in the ability to actively maintain information in primary memory, as well as the ability to recall “supra-span” items from secondary memory. Unsworth & Engle suggest that individuals with a lower working memory capacity (from now on: WMC) may have noisier representations in the primary memory system, which could lead to decreased capacity, decreased fidelity or decreased stability in primary memory as well as less effective retrieval from secondary memory due to “noisier” context or cue specification in the service of retrieval. Interestingly, Unsworth & Engle assume that WMC differences do not correlate with differences in terms of cue monitoring or decision processes (although this is not a necessary assumption of their theory).

This proposal contrasts with more traditional models of working and short-term memory (such as Baddeley’s influential model), which are themselves sufficient for explaining a wealth of data. However, Unsworth & Engle present a large amount of new data that is elegantly explained by their model, and might be more difficult to accomodate with more standard models of working memory. Below, I outline each of the major pieces of evidence used to motivate Unsworth & Engle’s “primary & secondary” model of working memory.

1) WMC correlates with tasks requiring the over-ride of prepotent responses because primary memory can be used to strongly maintain the goal or action plan, and bias behavior appropriately. Accordingly, individual WMC differences predict reaction times on antisaccade tasks, where subjects must look directly away from a flashed stimulus, presumably by maintaining the goal “look away” to override the more automatic response of “look towards.” Since “look towards” is the automatic behavioral response to suddenly-presented stimuli, individual differences in “prosaccade” tasks should not (and in fact, do not) correlate with WMC.

2) When prosaccade and antisaccade tasks are intermixed, as in a task-switching paradigm, individual differences in WMC correlate with errors on both types of trials: goal maintenance is now required for both trial types, given that they are intermixed.

3) Similarly, WMC relates to performance in the Stroop task, where subjects must strongly represent the goal of “name the ink color” and override the more automatic “read the word” response. Individual differences in WMC more strongly relate to error rates when the need for goal maintenance is not reinforced by every trial, but rather when a large proportion of congruent stimuli like “BLUE“. In this case, those with low WMC seem to lapse in their maintenance of the “ink color naming” goal, while individuals with a higher WMC are less likely to do so. (I’ve previously reviewed these results here).

4) Tests of working memory are actually contaminated by proactive interference from previous trials, a phenomenon usually considered only in the context of long-term memory. This supports the idea that a retrieval component contributes to measures of WMC. Accordingly, those with lower WMC show exaggerated signs of proactive interference, which Unsworth & Engle suggest are related to poorer context or cue specification in the service of retrieval from secondary memory (but could alternatively be due to weaker context or item maintenance in the first place, when those items were in primary memory. If this were the case, one would expect proactive interference among those with low WMC to be abolished by giving them more time to rehearse the target items during complex span tasks).

5) WMC correlates with performance in “directed forgetting” paradigms, which is most easily explained as a difference arising from differences in context maintenance or cue specification; previous models of WMC do not address these findings directly. The basic result is that those with higher WMC are generally better at remembering but are paradoxically worse at remembering if they’re told not to remember or if they’re encouraged to change their mental context (e.g., by imagining they are invisible; note this is one of the few known examples where greater WMC is actually a disadvantage). This again implicates WMC in cue and context specification.

6) WMC is related to performance on recollection but not recognition tasks, and free recall performance loads as strongly on latent WMC variables as complex span tasks, again supporting the idea that controlled search of long-term or secondary memory is integrated in measures of WMC. Likewise, WMC is related to recall latencies (mean RTs as well as mean inter-item RTs) on tasks where proactive interference is likely.

7) Measurements of primary and secondary memory suggest that while they are moderately correlated, they appear to load on independent factors and contribute unique variance to measures of WMC. For example, measures of primary memory significantly correlate with complex span when controlling for secondary memory, as do measures of secondary memory when controlling for primary memory. Primary and secondary memory do not correlate when controlling for complex span.

8) WMC correlates with many measures of higher-level cognition, but seems to do so because of the cognitive conlifct involved in many of those tasks. For example, vocabulary, general science knowledge and knowledge retrieval do not show strong correlations with WMC.

The authors have assumed that WMC differentiates individuals partly in maintenance, and partly their ability to use cues to interrogate long term memory, but this second point could also include individual differences in monitoring ability. It will be important for future work to tease the influences of these two variables apart (both of which are thought by some to rely on the inferior frontal gyrus).

Related Posts:
ACC: Monitoring Conflict or Response Frequency?
Symbols in Cognition: How Do Labels Help
Dissociations in the Development of Long Term Memory
Source Monitoring and Childhood Amnesia
Neural Cascades in Prefrontal Cortex
Visualizing Working Memory

Comments

  1. #1 Joanna
    February 27, 2007

    Interesting.

    Memory is so fascinating, and so much of “memory” as a concept is oversimplified – just as much in scholarly journals as in popularized media reports.

    For example: multiple memory systems, systems consolidation (hippocampal independent memory in the cortex), LTP/LTD not equating with memory, the molecular mechanisms involved and needed for learning and memory, etc.

  2. #2 tbell
    February 28, 2007

    just read this paper, very stimulating thanks for the link. The paper’s model does not require it, but the possibility that inhibitory processes may facilitate the active maintenance of information in the face of interference is potentially relevant. (I come out of a research lab with a strong interest in inhibitory control of memory)

  3. #3 Chris Chatham
    February 28, 2007

    That’s ironic – I come out of a research lab with a strong interest in explaining inhibition in terms of other mechanisms. :) A difference of perspective I guess.