Chapter 8 · One Resource, Shared by Three

If the narrowness of working memory was a matter of how many you can hold at once, there is one more limit apart from it: how many operations you can run at once on top of what you hold. Holding five things and sitting still, and binding those five together while at the same time bringing up something new, place different burdens on working memory. The second limit is this amount of computation. And the heart of this limit is that working memory's three operations do not each run on their own but share one household.

A Single Budget, Shared by Three

Formation, maintenance, and retrieval all draw from the same limited resource. What that resource is, memory research has long worked on; it is less the storage slots lined up in working memory than the attention that commands those slots—the control that allocates what to focus on. Working memory's storage clearly has grains that divide by kind, like words and images, but what the three operations contend over all at once is not those storage slots but the one attention that commands them. Holding a representation, binding a new composite representation, and pulling something from long-term memory all draw on this attention.

From the fact that there is one resource and three share it, one thing follows at once: if one operation takes much of the resource, the others' shares shrink by that much, and as one side improves the other worsens. This trade-off shows clearly first in a familiar scene outside working memory. Light chatter while driving a familiar road both roll along without strain. But the moment the road suddenly tangles or the conversation turns to a vexing negotiation, one eats the other. Pour attention onto the tricky road and your words stall; absorb yourself in the negotiation and you miss the turn. This trade-off—tilt attention to one side and it comes alive while the other dies—is the sign that the two are sharing one resource. There are only two things in your head, the road and the conversation, so there is room enough, yet you cannot run both at full force at once. If Chapter 7's narrowness was the limit of how many you hold at once, the second limit revealed here is how much you run on top of them.

Within working memory the same trade-off plays out among the three operations. When one binding grows complex and the inference runs long, the material you were holding meanwhile dims; when you must keep pulling from long-term memory, that retrieval takes the share meant for binding. Because formation, maintenance, and retrieval split one resource at every moment, when one grows heavy another collapses.

Retrieval Is Not Free

The point most easily doubted in the claim that the three operations contend over one resource is retrieval. Recalling something is just pulling it from long-term memory, so it looks free, as if it does not use working memory's resource separately. But there is an experiment showing it is not so. If, while retrieving something, you are at the same time made to do another task that divides attention, the accuracy of what is recalled itself barely drops, yet performance on that side task clearly worsens. If retrieval used no attention, the side task should be fine—but it is not. This is direct evidence that recalling draws on the resource of attention. Retrieval does not fall free from long-term memory; it is an operation that takes its own share from the same household as the other operations.

Why does this matter? Because when, in reading, the thing to bind has been pushed out of working memory and must be retrieved separately from long-term memory, that retrieval takes the very resource that should have gone to binding. If the preceding and following sentences point to the same object, that object still remains in working memory and binding finishes cheaply; but if there is no cue shared between them, the object must be hauled separately from long-term memory, and that hauling eats resource. If the relation to connect is itself absent from the text and must be made by inference, still more resource is spent. This is why the difficulty of a text is set not by the length of words or sentences but by how often such high-cost operations are forced.

The Depth Failure, Revisited

The depth failure seen in the previous chapter—having read all the letters yet with no understanding standing—can be reread through this resource competition. We said "what was built earlier is pushed out"; through the eye of resource, it is that retrieval and formation take so much resource that not enough is left to return to holding. With no resource to hold with, the context dims before reaching the topmost binding. Narrowness and competition are two faces of one failure.

Where to Channel the Resource

Here the direction of learning design takes shape: to let the limited resource reach the operations that actually make learning. The three operations stand differently with respect to learning. Formation builds a new composite representation and leaves in long-term memory a structure that was not there; and among retrievals, the kind recalled with effort where retrieval strength has fallen grows storage strength most, as seen earlier. These two make learning. Maintenance is different: it only holds material until it is bound, and merely keeping the same thing afloat in working memory leaves nothing of itself. Retrieval too has a kind that leaves no learning. The incidental retrieval forced when something to bind is pushed out while reading and must be hauled back is mere overhead that robs the resource meant for formation. So a good learning condition is one that lets resource flow to formation and hard retrieval and clears away the share that leaks without aiding learning. The leaking share is unnecessary maintenance that floats the same thing over and over, and the incidental retrieval that a sloppy text keeps forcing.

Actual learning methods are levers that turn this allocation in a particular direction. Some relieve the burden of incidental retrieval and search to drive resource into formation; others, conversely, deliberately spend more resource on something like discrimination and so change what remains. By which operation you let the one household of resource flow—how you turn that handle—the common learning techniques divide. In the next chapter we look at those levers one by one.