Caitlin Mace: Vehicle Indeterminacy in Memory Neuroscience

On Vehicle Indeterminacy in Memory Neuroscience

Caitlin Mace

University of Pittsburgh

Neural representations are broadly understood to be individuable structures or processes in the brain that have a representational role in cognition or behavior. This representational role can be elucidated in a number of ways, such as by the causal role the structure or process plays (Egan 2020; Cao 2022), how the structure or process is used (Baker et al. forthcoming), or the information it carries (Pohl et al. 2026). But the structure or processes themselves, understood as representational vehicles, are undertheorized. These vehicles are often loosely defined as neurons, spikes, electrical or chemical processes, or other patterns of neural activity in the brain. Once one attempts to get a theoretical grip on what precisely are the vehicles of neural representations, it becomes apparent that there are many ill-defined candidates.

To show this, I draw on neuroscientific investigations of memory engrams, which are neural representations of memory that result from learning-induced changes, retain information about some fact, feature of the world, or past experience, and are used to retrieve that information for memory recall. Commitment to memory engrams leaves open whether the engram is a neural structure, mechanism, or activities (Robins 2018). As such, there are many candidates. Models of memory engrams can be grouped into three classes: molecular models, synaptic models, and supracellular models (Colaço and Najenson 2024; Mace, O’Sullivan, and Wilson, forthcoming). Each class postulates disparate vehicle candidates to explain memory. While there have been attempts to reconcile some of these models or integrate them, it remains unclear which kind of structure—molecules, synapses, or neural ensembles—performs a representational role in any case. I argue in the paper that there is radical vehicle indeterminacy in neuroscience and raise two distinct vehicle indeterminacy issues.

The first, vehicle indefinity, is the problem of identifying meaningful properties by which to individuate vehicles. This indeterminacy results from blurry and arbitrarily posed boundaries on vehicles that are typically set by the specificity of the intervening or measuring apparatus. Consider each kind of memory engram model. The limits of neural ensembles are not obvious, and the addition or removal of neurons may be interpreted not as better approximations of the vehicle but as new vehicles altogether. A similar concern exists for synaptic processes underlying memory. While neuroscientists reasonably focus on understanding which synaptic processes are important for memory encoding, storage, or recall, there has been little investigation of which synapses are vehicles for memory. Some studies use labeling techniques to tag active synapses during remembering, but such techniques do not show whether the activity was increased or reduced (Eom, Kim, and Ho Hyun 2025). Finally, molecular models posit various vehicle candidates, from certain types of RNA to epigenetic processes. Insofar as these candidates are postulated to perform a representational role, rather than, say, provide instructions for creating engrams, there is no understanding of which nucleic acid sequences or genes are the vehicle. Here, again, neuroscience aims at understanding the processes rather than defining the bound of the vehicle. But without knowing such bounds, neuroscientific practices undermine vehicle postulation insofar as the properties by which vehicles are individuated and identified are not meaningful vehicle properties.

To see this, consider that evidence for vehicle candidates requires re-identifying tokens of that vehicle type. Re-identifying vehicles over time, in different brains, and in different labs shows that the entity in question reliably serves as a vehicle (Cao 2022). Moreover, re-identifying vehicles with different tools reduces the effects of errors, noise, or artifacts produced by any particular tool. This kind of robust analysis can tell us what the entity is, which is prerequisite for determining if the entity in question is the kind of thing that can perform a representational role. But identifying a vehicle with different detection methods across various contexts requires knowing what precisely should be re-identified, which is underdetermined in the case of vehicle indefinity.

The second kind of vehicle indeterminacy, vehicle disjunctivism, is the problem of determining which of many candidate vehicles is the real vehicle. This indeterminacy results from competing accounts of vehicle status and coarse-grained interventions. Part of the problem in determining which candidate vehicle has a representational role in the system is that interventions in the brain involve modifications to several variables. The brain is not the kind of system in which one can hold other components and processes stable to determine what is sufficient for causing representation-driven behavior like memory. While causal specificity is important for identifying and individuating engrams (see Najenson 2021, 2025), figuring out the vehicle for a specific memory content requires tools and theorizing that can target more and more specific vehicle candidates. Researchers aim to strengthen causal inferences by converging evidence across research programs involving various experiments and tools. But comparisons of populations, synapses, and molecules across contexts can only be inexact without resolving vehicle indefinity.

Vehicle disjunctivism moreover shows why coarse-graining vehicle candidates is an inadequate strategy for re-identification. Coarse-graining involves, for example, defining ensembles by the percentage of active neurons in a particular brain region, using a volume range to determine similarity in dendritic spine growth, identifying molecular processes by their functional role, or smoothing spike trains. While coarse-graining lends to re-identifiability, one cannot rule out other vehicle candidates. Consider the individuation of ensembles by the percentage of active neurons. Here, researchers are abstracting away details about which neurons are activated, which synaptic connections are formed or modified, what molecular mechanisms are active, and other properties that are alternative candidates for vehicle properties. Thus, other vehicle candidates remain plausible competitors. The coarse-graining strategy provides a way around the vehicle indefinity problem but still faces the problem of vehicle disjunctivism since indeterminacy persists about candidate vehicles.

Importantly, these indeterminacies are epistemic indeterminacies in being about shortcomings in our knowledge of vehicles and which properties of patterns are important for attributing vehicle status to those patterns. While there are some metaphysical consequences of epistemic indeterminacy, such as whether we should be realists about the vehicles that have been postulated by neuroscientists, these consequences are only considered insofar as a goal neuroscientists might have is to discover vehicles as they naturally exist in the brain. Meeting this goal requires resolving indeterminacy issues without coarse-grained, arbitrary, or merely stipulated bounds. And while resolving epistemic indeterminacy is just normal, everyday science, I try to show in my paper that the vehicle indeterminacy issues faced are radical in that the very methods that resolve epistemic indeterminacy in science—such as robust analysis across methods and contexts—break down in the case of vehicles. The discussion of vehicle indeterminacy here is one step to diagnosing the failure to resolve epistemic indeterminacies in ways that converge across methods and contexts. And finally, I suggest that these problems provide good reason to embrace pragmatism about neural representational vehicles.

Works Cited

1. Baker, Ben, Richard Lange, Andrew Richmond, Nikolaus Kriegeskorte, Rosa Cao, Xaq Pitkow, and Odelia Schwartz. [forthcoming]. “Use and usability: concepts of representation in philosophy, neuroscience, cognitive science, and computer science.” Neurons, Brains, Data Analysis, and Theory
2. Cao, Rosa. 2022. “Putting Representations to Use.” Synthese 200 (151): 1–24.
3. Colaço, David, and Jonathan Najenson. 2024. “Where memory resides: Is there a rivalry between molecular and synaptic models of memory?” Philosophy of Science 91: 1382–1392.
4. Egan, Frances. 2020. “A Deflationary Account of Mental Representation.” In What Are Mental Representations?, ed. Joulia Smortchkova, Krysztof Dołęga, and Tobias Schlicht, 26–53. Oxford: Oxford University Press.
5. Eom, Kisang, Donguk Kim, and Jung Ho Hyun. 2025. “Engram and behavior: How memory is stored in the brain.” Neurobiology of Learning and Memory 219 (108047): 1–16.
6. Mace, Caitlin, Fionn O’Sullivan, and Scott Wilson. Forthcoming. “Reductionism in Engram Neuroscience.” European Journal of Neuroscience 60. https://doi.org/10.1111/ejn.70497
7. Najenson, Jonathan. 2021. “What have we learned about the engram?” Synthese 199: 9581–9601.
8. Najenson, Jonathan. 2025. “Engrams and Causal Specificity.” Philosophical Psychology: 1–27. https://doi.org/10.1080/09515089.2025.2475173
9. Pohl, Stephan, Edgar Walker, David Barack, Jennifer Lee, Rachel Denison, Ned Block, Florent Meyniel, and Wei Ji Ji Ma. 2026. “Clarifying the conceptual dimensions of representation in neuroscience.” Nature Reviews Neuroscience https://doi.org/10.1038/s41583-026-01030-8
10. Robins, Sarah. 2018. “Memory and Optogenetic Intervention: Separating the Engram from the Ecphory.” Philosophy of Science 85: 1078–1089.