When you choose between two restaurants, two job offers, or two vacation destinations, it feels like you're consulting an internal ledger—some stable record of what you want. Decision theory has long modeled preference this way, treating utility as a stored quantity that the agent simply reads out at the moment of choice. But a growing body of evidence from cognitive neuroscience and computational modeling suggests something far more interesting: preferences are not retrieved. They are built.

The construction of preference unfolds through a series of memory queries, attentional shifts, and value computations that are deeply shaped by the architecture of the cognitive system performing them. This means that the same agent, facing the same options, can generate different preferences depending on which memories are cued first, how attention is allocated, and what cognitive resources happen to be available. Preference, in this view, is less like reading a file and more like running a program—one whose output depends critically on its execution context.

This article maps the computational architecture underlying preferential decisions. Drawing on query theory, evidence accumulation models, and neuroeconomic findings on value construction, we'll examine how memory retrieval, sequential processing, and architectural constraints conspire to produce the preferences we experience as stable and self-evident. The goal is not merely to catalog biases, but to understand the process model—the generative mechanism from which preference emerges as a computational artifact.

The classical expected utility framework treats preferences as pre-existing orderings over outcomes. An agent possesses a utility function, and choice consists of maximizing that function given available options. This is elegant mathematics, but it implies a specific cognitive claim: that somewhere in the mind, there exists a stable representation of value for each option, waiting to be consulted. The empirical evidence overwhelmingly contradicts this.

Research on constructed preferences demonstrates that valuations are assembled at the time of decision through a process of evidence sampling from memory. When you evaluate whether to take a new job, you don't access a single stored utility value for that option. Instead, you retrieve a sequence of considerations—salary, commute time, the personality of the hiring manager, that article you read about the company's culture. Each retrieved consideration contributes an increment or decrement to the emerging valuation. The preference you report is the cumulative output of this retrieval process.

Critically, memory retrieval is not exhaustive or unbiased. It is governed by associative activation, recency, emotional salience, and the specific cues present in the decision context. This means the set of considerations that gets retrieved—and therefore the preference that gets constructed—is contingent on factors that have nothing to do with the objective attributes of the options. Framing effects, context effects, and preference reversals all become natural consequences of this architecture rather than anomalies requiring ad hoc explanation.

Neuroscientific work supports this process account. Ventromedial prefrontal cortex (vmPFC) activity during valuation correlates not with a static stored signal but with the dynamic integration of attribute information over time. Studies using fMRI and eye-tracking show that value signals in vmPFC evolve as attention shifts between attributes, consistent with a constructive process rather than a simple readout. The hippocampal memory system feeds attribute-level information into this valuation circuit, and disruptions to hippocampal function alter not just memory but preference itself.

What makes this finding theoretically profound is that it dissolves the boundary between preference and cognition. If preferences are constructed through the same memory retrieval mechanisms that support episodic recall and semantic association, then preference is not a separate primitive of the decision architecture—it is an emergent property of general-purpose cognitive operations applied to a specific task. The implications for rational choice theory are significant: the stability axioms that undergird expected utility may describe an idealization that the computational architecture of the brain is not designed to satisfy.

Takeaway

Preferences are not stable entities waiting to be discovered—they are constructed in real time through memory retrieval, which means the process of deciding literally creates what you want.

Query theory, developed by Eric Johnson, Gerald Häubl, and Anat Keinan, provides a formal process model for how memory-based preference construction actually unfolds. The core insight is deceptively simple: when forming a preference, people generate internal queries—they ask themselves for reasons favoring or opposing each option. The order in which these queries are executed has a systematic and predictable effect on the resulting preference.

The mechanism relies on a well-established property of memory called output interference. When you retrieve information from memory in response to one query, the act of retrieval itself inhibits the subsequent retrieval of related but competing information. If you first query reasons to keep your current job, the retrieved considerations partially suppress reasons favoring the new opportunity. The first query enjoys a retrieval advantage, generating more considerations, which in turn tilts the constructed preference in its direction.

This framework explains the endowment effect with remarkable precision. Owners of an object tend to first query reasons to keep it; potential buyers tend to first query reasons not to acquire it. The asymmetric query order, combined with output interference, produces the valuation gap between willingness-to-accept and willingness-to-pay that has puzzled economists since Kahneman, Knetsch, and Thaler first documented it. When researchers experimentally manipulate query order—instructing participants to first generate reasons to trade rather than to keep—the endowment effect diminishes substantially or disappears entirely.

Query theory also accounts for temporal discounting asymmetries, default effects, and certain framing phenomena. In each case, the situational context biases which query is executed first, and output interference does the rest. The theory makes quantitative predictions: the number of considerations generated per query mediates the preference shift, and this mediation can be measured and modeled. It is not merely a verbal redescription of bias—it is a mechanistic account with testable, falsifiable implications at the level of cognitive process.

From a computational perspective, query theory recasts preference formation as a sequential sampling problem where the sampling distribution is endogenous—it depends on what has already been sampled. This links it naturally to evidence accumulation frameworks like the drift-diffusion model, but with an important twist: the drift rate itself changes as a function of retrieval history. Preference is not determined by a fixed signal-to-noise ratio; it is determined by the trajectory of an inherently path-dependent process. The order of thought is not epiphenomenal to choice—it is constitutive of it.

Takeaway

The order in which you mentally query reasons for and against an option doesn't just reflect your preference—it actively determines it, because earlier queries suppress competing evidence through output interference.

If preferences are constructed through sequential memory retrieval, then the architectural constraints of the cognitive system—working memory capacity, attentional bandwidth, retrieval dynamics—are not incidental to choice. They are constitutive parameters of the preference-generation mechanism. Understanding these constraints is essential for understanding why preferences take the specific form they do.

Working memory imposes a hard bottleneck on the number of considerations that can be simultaneously active and integrated. George Miller's classic estimate of 7±2 items has been revised downward by modern research to roughly 3-4 chunks in the focus of attention. This means that at any given moment during preference construction, the decision-maker is working with a severely limited subset of the total evidence available in long-term memory. The preference that emerges is not a function of everything the agent knows—it is a function of what the agent can hold in mind at the critical moments of integration.

Attention further constrains the process. Eye-tracking studies of multi-attribute choice show that people do not uniformly sample all attributes of all options. Instead, attention is allocated in a pattern shaped by salience, spatial position, and the emerging preference itself—a phenomenon called the gaze cascade effect, where people increasingly fixate on the option they are about to choose. This creates a positive feedback loop: the option that captures more attention accumulates more evidence, which in turn attracts further attention. The attentional architecture doesn't just filter the input to the decision—it actively amplifies certain signals at the expense of others.

Retrieval from long-term memory introduces its own biases. Associative activation means that contextual cues disproportionately surface memories that are semantically linked to the cue, regardless of their diagnostic value for the decision at hand. Emotional arousal at the time of encoding enhances retrieval probability, so emotionally charged experiences receive outsized weight in the construction process. And the temporal dynamics of retrieval—the fact that earlier-retrieved items have longer to influence the accumulating preference signal—mean that retrieval fluency is confounded with evidential weight.

These constraints collectively define what we might call the computational envelope of preferential choice. Within this envelope, the system does something remarkably sophisticated: it integrates heterogeneous evidence from memory into a coherent value signal under time pressure with limited resources. But the shape of the envelope guarantees that the output will be systematically different from what an unconstrained optimizer would produce. This is not a failure of rationality—it is the signature of a system that evolved to make adequate decisions quickly, not optimal decisions eventually. The architecture is the theory.

Takeaway

The limitations of working memory, attention, and retrieval are not bugs in the decision system—they are fundamental parameters that define the space of preferences the system can construct, making architecture itself a theory of choice.

The architecture of preferential choice reveals that what we call a preference is the output of a constructive computational process, not the readout of a stored value. Memory retrieval, query sequencing, and the hard limits of working memory and attention jointly determine the preference that emerges at any given moment. The elegant stability assumed by classical utility theory is an approximation—useful for some purposes, but misleading as a description of mechanism.

This process-level understanding has practical consequences. If query order shapes preference, then whoever controls the framing controls which queries fire first. If attentional constraints amplify early signals, then the design of choice environments is not neutral—it is architecturally determinative.

The deepest implication may be philosophical. If preferences are constructed rather than revealed, then the question "What do you really want?" may not have a determinate answer independent of the computational context in which it is asked. The architecture doesn't just constrain choice—it constitutes it.