Let's start from scratch in thinking about what memory is for, and consequently, how it works. Suppose that memory and conceptualization work in the service of perception and action. In this case, conceptualization is the encoding of patterns of possible physical interaction with a three-dimensional world. These patterns are constrained by the structure of the environment, the structure of our bodies, and memory. Thus, how we perceive and conceive of the environment is determined by the types of bodies we have. Such a memory would not have associations. Instead, how concepts become related (and what it means to be related) is determined by how separate patterns of actions can be combined given the constraints of our bodies. I call this combination “mesh.” To avoid hallucination, conceptualization would normally be driven by the environment, and patterns of action from memory would play a supporting, but automatic, role. A significant human skill is learning to suppress the overriding contribution of the environment to conceptualization, thereby allowing memory to guide conceptualization. The effort used in suppressing input from the environment pays off by allowing prediction, recollective memory, and language comprehension. I review theoretical work in cognitive science and empirical work in memory and language comprehension that suggest that it may be possible to investigate connections between topics as disparate as infantile amnesia and mental-model theory.

A functional theory of memory has already been developed as part of a general functional theory of cognition. The traditional conception of memory as “reproductive” touches on only a minor function. The primary function of memory is in constructing values for goal-directedness of everyday thought and action. This functional approach to memory rests on a solid empirical foundation.

Glenberg's theory is rich and provocative, in our view, but we find fault with the premise that all memory representations are embodied. We cite instances in which that premise mispredicts empirical results or underestimates human capabilities, and we suggest that the motivation for the embodiment idea – to avoid the symbol-grounding problem – should not, ultimately, constrain psychological theorizing.

Glenberg's rethinking of memory theory seems limited in its ability to handle abstract symbolic thought, the selective character of cognition, and the self. Glenberg's framework can be elaborated by linking it with theoretical efforts concerned with cognitive development (Piaget) and ecological perception (Gibson). These elaborations point to the role of memory in specifying the self as an active agent.

We are sympathetic to most of what Glenberg says in his target article, but we consider it common wisdom rather than something radically new. Others have argued persuasively against the idea of abstraction in cognition, for example. On the other hand, Hebbian connectionism cannot get along without the idea of association, at least at the neural level.

(1) Non-projectable properties as opposed to the clamping of projectable properties play a primary role in triggering and guiding human action. (2) Embodiment in language-mediated memories should be qualified: (a) Language imposes a radical discretization on body constraints (second-order embodiment). (b) Metaphors rely on second-order embodiment. (c) Language users sometimes suspend embodiment.

This commentary connects some of Glenberg's ideas to similar ideas from artificial intelligence. Second, it briefly discusses hidden assumptions relating to meaning, representations, and projectable properties. Finally, questions about mechanisms, mental imagery, and conceptualization in animals are posed.

Corresponding to Glenberg's distinction between the automatic and effortful modes of memory, I propose a distinction between cued and detached mental representations. A cued representation stands for something that is present in the external situation of the representing organism, while a detached representation stands for objects or events that are not present in the current situation. This distinction is important for understanding the role of memory in different cognitive functions like planning and pretense.

Researchers in the field of discourse processing have investigated how mental models are constructed when adults comprehend stories. They have explored the process of encoding various classes of inferences “on-line” when these mental microworlds are constructed during comprehension. This commentary addresses the extent to which these inferences and mental microworlds are “embodied.”

Glenberg argues for embodied representations relevant to action. In contrast, we propose a grouping of representations, not necessarily all being directly embodied. Without assuming the existence of representations that are not directly embodied, one cannot account for the use of knowledge abstracted from direct experience.

Glenberg's conception of “meaning from and for action” is too narrow. For example, it provides no satisfactory account of the “logic of Elfland,” a metaphor used by Chesterton to refer to meaning acquired by being told something.

All that we call spirit and art and ecstasy only means that for one awful instant we remember that we forget.

G. K. Chesterton (in Gardner 1994, p. 101)

Glenberg provides a new and exciting view that is especially useful for capturing some functional aspects of memory. However, memory and its functions are too multifarious to be handled by any one conceptualization. We suggest that Glenberg's proposal be restricted to its own “focus of convenience.” In addition, its value will ultimately depend on its success in generating detailed and testable theories.

Glenberg focuses on conceptualizations that change from moment to moment, yet he dismisses the concept of working memory (sect. 4.3), which offers an account of temporary storage and on-line cognition. This commentary questions whether Glenberg's account adequately caters for observations of consistent data patterns in temporary storage of verbal and visuospatial information in healthy adults and in brain-damaged patients with deficits in temporary retention.

To model potential interactions, memory must not only mesh prior patterns of action, as Glenberg proposes, but also their internal consequences. This is necessary both to discriminate sensorimotor information by its relevance and to explain how go als about the world develop. In the absence of internal feedback, Glenberg is forced to reintroduce a grounding problem into his otherwise sound model by presupposing interactive goals.

Can memory be cast as a system that meshes events to actions? This commentary considers the concepts of mesh versus association, arguing that thus far the distinction is inadequate. However, the goal of shifting to an action-based view of memory has merit, most notably in emphasizing memory as a skill and in focusing on processes as opposed to structures.

Glenberg tries to explain how and why memories have semantic content. The theory succeeds in specifying the relations between two major classes of memory phenomena – explicit and implicit memory – but it may fail in its assignment of relative importance to these phenomena and in its account of meaning. The theory is syntactic and extensional, instead of semantic and intensional.

There are three major weaknesses with Glenberg's theory. The first is that his theory makes assumptions about internal representations that cannot be adequately tested. The second is that he tries to accommodate data from three disparate domains: mental models, linguistics, and memory. The third is that he makes light of advances in cognitive neuroscience.

The functional theory of memory set out in Glenberg's target article accords with recent proposals in the developmental literature with respect to event memory, conceptualization, and language acquisition from an embodied, experiential view. The theory, however, needs to be supplemented with a recognition of the sociocultural contribution to these cognitive processes and emerging structures.

What would Glenberg's attractive ideas look like when computationally fleshed out? I suggest that the most helpful next step in formalizing them is neither a connectionist nor a symbolic implementation (either is possible), but rather an implementation- general analysis of the task in terms of the informational content required.

The ability of Glenberg's model to explain the development of complex symbolic abilities is questioned. Specifically, it is proposed that the concepts of clamping and suppression fall short of providing an explanation for higher symbolic processes such as autobiographical memory and language comprehension. A related concept, “holding in mind” (Olson 1993), is proposed as an alternative.