Mechanisms
Memory Systems · Non-Declarative
Classical Conditioning (Delay and Trace)
Example
A car backfires outside the window while you are working. It happens again the next week, and the week after. By the fourth time you flinch before you have identified the sound, and your heart rate rises before you have decided anything. You could describe the pattern if asked, but the flinch does not wait for the description, and it would still happen if you could not.
Classical Conditioning in Humans
Definition
Repeated pairing of a neutral stimulus with a consequence changes the response to that stimulus. In the canonical case the response and the knowledge of the pairing dissociate completely: either can be acquired while the other fails (Bechara et al., 1995). The dissociation is not uniform across procedures, and the two standard forms differ in whether the learning appears to be declarative.
- Delay conditioning. The conditioned and unconditioned stimuli overlap in time. Acquisition proceeds without awareness of the contingency and survives medial temporal lobe damage. This is the non-declarative form.
- Trace conditioning. A gap separates the two stimuli. Acquisition tracks awareness of the contingency closely and requires an intact medial temporal lobe (Clark & Squire, 1998).
Contested. Whether any form of Pavlovian conditioning proceeds without awareness is disputed. Reviewing autonomic conditioning, subliminal stimuli, eyeblink conditioning, conditioning in amnesia, evaluative conditioning, and conditioning under anesthesia, Lovibond and Shanks (2002) conclude that the evidence is largely consistent with awareness being necessary though not sufficient for conditioned performance, and that conditioning may depend on a propositional system associated with consciousness rather than a separate lower-level one. If that is right, this construct does not belong in the non-declarative block at all. They also note that many studies used inadequate measures of awareness, which is the specific objection raised against Clark and Squire.
The dispute runs through the row's own evidence. Bechara's hippocampal patient acquired the conditioned response while failing to acquire the declarative facts about the pairing, which is direct evidence against a general awareness requirement. This row treats the delay and trace forms as distinguishable and the awareness boundary as unresolved.
Neural Substrate
Amygdala for the conditioned response, hippocampus for the declarative facts about the pairing: separable systems, with damage to either sparing the other and damage to both abolishing each (Bechara et al., 1995). The hippocampal side of the dissociation carries one qualification: the hippocampal patient produced conditioned responses of normal magnitude but did not respond differentially, continuing to respond to the unpaired stimuli at magnitudes the controls had stopped producing, which the authors read as the hippocampus contributing to the discrimination between paired and unpaired stimuli rather than to the response itself (Bechara et al., 1995).
The response pathway differs by paradigm. In eyeblink conditioning, where the lesion evidence is from rabbits, the conditioned response depends on the cerebellar interpositus nucleus, whose lesion permanently abolishes it, while aspiration of the overlying lateral cerebellar cortex produces only a transient decrement followed by relearning with significant savings (Woodruff-Pak et al., 1985).
Trace and delay procedures diverge above these structures. Amnesic patients fail to acquire trace conditioning while acquiring delay conditioning normally (Clark & Squire, 1998). In a fear conditioning paradigm, the supplementary motor area, frontal operculum, middle frontal gyri, and inferior parietal lobule respond most strongly during the trace interval, with intermediate responses on delay trials and the smallest to the unpaired stimulus, while anterior cingulate and medial thalamus respond to delay and trace alike; the hippocampus is not trace-specific in the same way, showing early increases to every stimulus that fade within the first block of training, a deactivation that is largest during the trace interval, and a first-block difference tied only to how precisely participants timed the shock (Knight et al., 2004). The divergence therefore holds across both response systems, which are otherwise studied in separate literatures, with one caution: the trace intervals are not on the same scale, 500 to 1000 ms in the eyeblink work (Clark & Squire, 1998) against 10 s in the fear paradigm (Knight et al., 2004), so whether "trace" names one procedure across the two is not settled by either study.
AI at a Glance
In humans the conditioned response and the knowledge of the pairing come apart: amygdala damage leaves the facts intact and the response absent, hippocampal damage the reverse (Bechara et al., 1995). Language models are measured on conditioning directly in context and through added memory layers, and it is the paradigm the benchmark authors identify as their critical bottleneck, but the dissociation that defines the construct cannot be assessed in a system with a single output channel (Qin et al., 2026).
Classical Conditioning in AI
How Far It Has Been Measured
Measurement in AI: Indirectly Measured
AI Memory Performance: Below Human provisional
Where It Lives in the System
What Has Been Measured
- Only one study found tests AI on classical conditioning by name, a peer-reviewed 2026 benchmark covering the current conversation and add-on saved memory; it uses a text-based analogue rather than the eyeblink or fear paradigm, and no study was found that tests conditioning acquired during training, so the ability is only partly measured (Qin et al., 2026).
- Each of the 100 conditioning items pairs a cue with a bad outcome over four conversational turns (for example, requests for an "http" link each followed by an "insecure protocol" warning), adds two unrelated turns, then brings the cue back; an item counts as correct only if the first reply avoids or warns about the cue, judged against a rubric by a separate AI model (Qin et al., 2026).
- Across 17 large language models (averaged over three runs), conditioning scores ranged from 35.7% (Qwen-2.5-7B) to 69.7% for the best system (DeepSeek-R1), with GPT-5 at 64.0% and Claude-4.1-opus at 41.7%; avoidance-type items fared worst, averaging 17.6% versus 75.0% for preference-type items (Qin et al., 2026).
- The human baseline was five computer science Ph.D. students who completed the same items under the same protocol, scored with the same rubric by two other Ph.D. students (100% agreement); all five scored 100%, so the best AI fell about 30 points below the entire human range, though the human sample is tiny, expert, at ceiling, and scored by people rather than the AI judge (Qin et al., 2026).
- The test was, if anything, easier for the AI: the models could still see the whole short dialogue (about 500 tokens) in their context window, meaning the text of the current conversation, and the paper does not say whether people could reread it, so the below-human result goes against the AI's advantage and counts (Qin et al., 2026).
- Adding saved memory did not close the gap: memory-augmented agents scored 63% (MEM1), 25% (MemGPT), and 22% (MemAgent) on conditioning, and even when researchers manually stored the exact cue–outcome pairing in an external memory tool (Mem0), the best score was 76% (Qin et al., 2026).
- A 2025 preprint shows that the update rules of linear-attention transformer layers are mathematically identical to classic conditioning models (Hebbian contiguity and Rescorla–Wagner error correction) and reproduce Kamin blocking in small trained networks, but it did not test deployed language models or compare them with people (Qiao, 2025).
Where a Mechanism Can Live
04 Product
The whole configuration, and the only thing a benchmark observes directly.
01
Weights
Across deployments
02
Context Window
Within a session
03
Memory Layer
Across sessions
Which of the three a result belongs to is inferred from it, not seen in it.
Read more
Amygdala
Subcortical in the medial temporal lobe
An almond-shaped cluster of a dozen distinct nuclei in the anterior medial temporal lobe, not a single functional organ.
Grey matter in front of the hippocampus, conventionally divided into basolateral, central, medial and cortical groups whose cell types, connections and outputs differ enough that treating them as one structure obscures most of what is known. The "fear center" label is wrong on two counts: the nuclei do different things, and the responses they are implicated in are not specific to fear. Recordings and lesions support involvement in learning about biologically significant events of either valence, and whether the common factor is valence, arousal, or salience remains genuinely open.
Anterior cingulate cortex
Cortex in the cingulate cortex
The front section of the cingulate, curving around the genu of the corpus callosum, with subgenual, pregenual, and dorsal parts.
The anterior cingulate cortex wraps the front of the corpus callosum, occupying Brodmann areas 24, 32 and 33, and on some maps 25. It is connected with the amygdala, the autonomic brainstem, the ventral striatum and lateral prefrontal cortex, and in humans and great apes it contains von Economo neurons. Its function is one of the most openly contested in the field: the same dorsal tissue has been proposed as a conflict monitor, an error detector, a tracker of the expected value of control, an effort or foraging-cost signal, and part of the pain network. The proposals are not variants of one claim; they make different predictions and coexist.
Cerebellar cortex
Cerebellum in the cerebellum
The three-layered sheet covering the cerebellum, built from one repeating circuit across its whole extent.
Cerebellar cortex is folded into fine parallel folia and has three layers: an outer molecular layer of parallel fibers and interneurons, a single sheet of Purkinje cells, and an inner granule layer. Mossy fibers excite granule cells, whose axons rise and split into parallel fibers contacting many Purkinje cells weakly; climbing fibers from the inferior olive contact one Purkinje cell each, powerfully. Purkinje cells provide the only output, and it is inhibitory, onto the deep nuclei. This uniformity is why the cerebellum dominates computational modelling of the brain, and why regional differences are inferred from connections rather than structure.
Cerebellum
Cerebellum
The densely folded structure behind the brainstem holding most of the brain's neurons in a uniform, repeating circuit.
The cerebellum sits under the occipital lobes behind the pons and medulla, attached by three pairs of peduncles. It is about a tenth of brain volume yet contains roughly four-fifths of all neurons, largely because of its granule cells. Its cortex uses the same circuit everywhere, which makes it unusually tractable: input arrives as mossy and climbing fibers, output leaves through Purkinje cells to the deep nuclei. Classically a structure of motor coordination and timing, it is now widely implicated in cognition and language on imaging and lesion evidence, though what it contributes beyond motor control is actively argued.
Hippocampus
Cortex in the hippocampal formation
A curved allocortical structure in the medial temporal lobe whose bilateral damage reliably produces dense anterograde amnesia in humans.
A paired, seahorse-shaped fold of three-layered cortex running front to back inside the temporal lobe, connected to the rest of the brain mainly through entorhinal cortex and the fornix. What it does is genuinely contested and has been for fifty years: the human lesion literature reads it as necessary for forming new declarative memories, the rodent recording literature as a map of space, and a third position treats both as instances of binding arbitrary elements into relations. The accounts differ on which species and which task they explain best, and the question of which is right remains open.
Inferior parietal lobule
Cortex in the parietal lobe
The parietal cortex below the intraparietal sulcus, comprising the supramarginal and angular gyri.
The inferior parietal lobule lies beneath the intraparietal sulcus and above the lateral fissure, made up of the supramarginal gyrus in front and the angular gyrus behind, with the temporoparietal junction at its lower border. It is disproportionately expanded in humans relative to other primates and is among the last cortical regions to myelinate. Left-sided lesions contribute to apraxia and to phonological and semantic deficits; right-sided lesions are the commonest cause of hemispatial neglect. Its parcellation is unsettled: cytoarchitectonic studies report five to seven areas where the gross anatomy shows two gyri.
Medial temporal lobe
Cortex in the temporal lobe
The inner fold of the temporal lobe (hippocampus, the cortices wrapped around it, and the amygdala), removed together in the classic amnesia cases.
A regional grouping rather than a single structure: the hippocampal formation, the entorhinal, perirhinal and parahippocampal cortices that feed it, and, on some definitions, the amygdala. It became the canonical "memory region" because the surgeries and infarcts that produce dense anterograde amnesia (H.M. most famously) take out most of it at once. Whether the parts constitute one memory system or several with different jobs is an open argument, and lesions in patients often cross the boundaries that would settle it.
Middle frontal gyrus
Cortex in the frontal lobe
The gyrus between the superior and inferior frontal sulci, occupying most of the lateral prefrontal surface.
The middle frontal gyrus lies between the superior and inferior frontal sulci and extends forward to the frontal pole. Much of what is called dorsolateral prefrontal cortex falls on it, along with the frontal eye field near its junction with the precentral sulcus. It is reported across a wide range of imaging tasks, activated during working memory, attentional control and rule-guided behavior. That ubiquity is a caution rather than a finding: the gyrus is large, its sulcal pattern varies considerably between individuals, and group-average coordinates on it can correspond to different cytoarchitectonic areas in different brains.
Operculum
Cortex in the cerebral hemisphere
The lips of cortex from the frontal, parietal and temporal lobes that fold over and conceal the insula.
The operculum (Latin for lid) is the cortex that overhangs the insula and closes the lateral fissure over it, contributed by the frontal, parietal and temporal lobes and named accordingly. It forms during development as the surrounding cortex expands faster than the insular region and grows across it. The parietal operculum carries secondary somatosensory cortex; the frontal operculum includes pars opercularis and cortex associated with speech production and taste. Failure of this folding, as in some malformations of cortical development, leaves the insula exposed and is visible on imaging.
Supplementary motor area
Cortex in the motor cortex
The medial part of Brodmann area 6, on the inner surface of the hemisphere just in front of the leg area of motor cortex.
The supplementary motor area lies on the medial wall of the superior frontal gyrus, anterior to the primary motor leg representation and above the cingulate sulcus. It projects directly to the spinal cord as well as to primary motor cortex, and stimulation evokes complex postures rather than single twitches. It is engaged by internally generated and sequential movements, by bimanual coordination, and by timing. The readiness potential preceding voluntary action arises largely here, which is why the region features in arguments about volition that go well beyond what the recordings themselves establish.
Thalamus
Subcortical in the diencephalon
Paired egg-shaped grey masses flanking the third ventricle, through which nearly all sensory and cortico-cortical traffic is relayed.
Each thalamus is a walnut-sized mass of at least thirty named nuclei, divided by a Y-shaped sheet of white matter (the internal medullary lamina) and wrapped in the reticular nucleus. Every sensory modality except olfaction relays here before reaching cortex, and cortical areas talk to one another through thalamic nuclei as well as directly. Whether the thalamus is a relay that passes signals on or an active participant that gates and sharpens them is a live argument; higher-order nuclei such as the pulvinar and mediodorsal receive more input from cortex than from the periphery, which is hard to square with a pure relay account.
Bechara et al. (1995)
Double Dissociation of Conditioning and Declarative Knowledge Relative to the Amygdala and Hippocampus in Humans
In
Science, 269(5227), 1115-1118
Abstract
A patient with selective bilateral damage to the amygdala did not acquire conditioned autonomic responses to visual or auditory stimuli but did acquire the declarative facts about which visual or auditory stimuli were paired with the unconditioned stimulus. By contrast, a patient with selective bilateral damage to the hippocampus failed to acquire the facts but did acquire the conditioning. Finally, a patient with bilateral damage to both amygdala and hippocampal formation acquired neither the conditioning nor the facts. These findings demonstrate a double dissociation of conditioning and declarative knowledge relative to the human amygdala and hippocampus.
Find it
Clark & Squire (1998)
Classical Conditioning and Brain Systems: The Role of Awareness
In
Science, 280(5360), 77-81
Abstract
Classical conditioning of the eye-blink response, perhaps the best studied example of associative learning in vertebrates, is relatively automatic and reflexive, and with the standard procedure (simple delay conditioning), it is intact in animals with hippocampal lesions. In delay conditioning, a tone \[the conditioned stimulus (CS)\] is presented just before an air puff to the eye \[the unconditioned stimulus (US)\]. The US is then presented, and the two stimuli coterminate. In trace conditioning, a variant of the standard paradigm, a short interval (500 to 1000 ms) is interposed between the offset of the CS and the onset of the US. Animals with hippocampal lesions fail to acquire trace...
Find it
Knight et al. (2004)
Neural Substrates Mediating Human Delay and Trace Fear Conditioning
In
The Journal of Neuroscience, 24(1), 218-228
Abstract
Previous functional magnetic resonance imaging (fMRI) studies with human subjects have explored the neural substrates involved in forming associations in Pavlovian fear conditioning. Most of these studies used delay procedures, in which the conditioned stimulus (CS) and unconditioned stimulus (UCS) coterminate. Less is known about brain regions that support trace conditioning, a procedure in which an interval of time (trace interval) elapses between CS termination and UCS onset. Previous work suggests significant overlap in the neural circuitry supporting delay and trace fear conditioning, although trace conditioning requires recruitment of additional brain regions. In the present...
Find it
Qiao (2025)
Attention as Conditioning: What Classical Learning Theory Predicts About Linear Transformers
In
arXiv
Abstract
Attention is widely understood as an associative memory, but that description alone does not predict how the memory will behave. Predictive theories do exist, but in the literature on animal learning. We show that the state updates of the major linear-attention families are term-for-term identical with named models from a century of animal learning theory: linear attention implements Hebbian contiguity, DeltaNet implements Rescorla--Wagner error correction, and decay variants such as RetNet implement contiguity with a stimulus trace. This dictionary turns conditioning phenomena into testable statements about the in-context behavior of linear transformers, while distinguishing algebraic...
Find it
Qin et al. (2026)
ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models
In
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 28232-28261
Abstract
Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning...
Find it
Woodruff-Pak et al. (1985)
Trace conditioning: Abolished by cerebellar nuclear lesions but not lateral cerebellar cortex aspirations
In
Brain Research, 348(2), 249-260
Abstract
The trace conditioning paradigm in which the conditioned stimulus-unconditioned stimulus (CS-US) interval is longer and the US and CS do not overlap requires retention of a ‘trace’ of the CS to associate it with the US. This task is difficult, requiring for learning about 5 times the number of trials to criterion as the standard delay task. The present study was undertaken to determine if: (1) the cerebellar interpositus nucleus is essential in trace as well as in delay conditioning; and (2) the lateral cerebellar cortex is involved when CS-US association over longer time intervals is required. Sixteen adult male New Zealand white rabbits had recording and in some cases lesion electrodes...
Find it