When Ivan Pavlov documented that a conditioned response fades with repeated unreinforced presentations, he also recorded something inconvenient: given time, the response returned. A century later, that observation has become one of the most consequential findings in affective neuroscience. Extinction does not erase the original fear memory. It writes a second one, and the two compete for control of behaviour.

This distinction is not semantic. If extinction were unlearning, a successful course of exposure therapy would leave a patient functionally equivalent to someone who had never acquired the fear. Instead, rodent lesion studies, human neuroimaging, and decades of behavioural work in the tradition of Mark Bouton converge on a different conclusion. The original association between the conditioned stimulus (CS) and the unconditioned stimulus (US) survives largely intact. What changes is the addition of an inhibitory memory, encoded in prefrontal–amygdala circuitry, that signals this cue is safe, here, now.

That framing changes how we should read relapse after treatment. A return of fear is not necessarily evidence that therapy failed. It is the predictable consequence of a memory architecture in which the safety trace is fragile, bound to context, and easily outcompeted under the wrong conditions. Understanding why this happens, and what can be done about it, remains one of the most productive translational problems in clinical neuroscience. It sits precisely where basic learning theory meets the consulting room.

The behavioural evidence came first. Four phenomena show that an extinguished response can reappear without any new CS–US pairing. In spontaneous recovery, fear returns with the passage of time. In renewal, it returns after a change of context. In reinstatement, it returns after unsignalled exposure to the aversive outcome. In savings, reacquisition proceeds faster than the original learning did. Each effect implies a memory that persisted beneath suppressed behaviour.

The pharmacological evidence points the same way. In 1992, Falls, Miserendino and Davis showed that infusing an NMDA receptor antagonist into the rat amygdala blocked extinction. That is the signature of active plasticity, not passive decay. Erasure through disuse would not obviously require the molecular machinery of synaptic learning. Acquiring a new association does. Extinction recruits much the same glutamatergic and intracellular signalling cascades as acquisition, consistent with the brain encoding something new rather than deleting something old.

Gregory Quirk and colleagues then localised the inhibitory trace. In rats, lesions of the infralimbic cortex left within-session extinction largely intact but impaired recall of extinction the following day. Milad and Quirk (2002) found that infralimbic neurons responded to the tone only after extinction. Pairing infralimbic microstimulation with the tone also mimicked extinction. The proposed mechanism runs through infralimbic projections to GABAergic intercalated cells. These cells gate output from the central amygdala, suppressing fear expression without dismantling the original association stored upstream.

Human data map onto this circuit. Phelps and colleagues (2004) reported ventromedial prefrontal cortex (vmPFC) engagement during extinction recall. Milad et al. (2007) found that vmPFC cortical thickness correlated with how well extinction was retained. At the cellular level, Herry et al. (2008) identified distinct populations of fear neurons and extinction neurons in the basolateral amygdala. Switches between the two populations tracked behavioural state and depended on hippocampal input. Competing memories, in other words, are physically instantiated in separable ensembles.

The clinical implication is significant. Patients with post-traumatic stress disorder often acquire and extinguish conditioned fear within a session at near-normal rates. Their deficit emerges at recall. Milad et al. (2009) found that impaired extinction retention in PTSD was accompanied by reduced vmPFC and hippocampal activation alongside heightened amygdala and dorsal anterior cingulate responses. The core problem may lie less in learning safety than in retrieving it. That shifts the therapeutic target from the extinction session itself to the conditions under which its memory is later accessed.

Takeaway

Extinction leaves two memories, not one. Recovery from fear is less a matter of deletion than of which trace wins at the moment of retrieval.

Bouton's central insight was an asymmetry. Excitatory conditioning generalises across contexts relatively well, while extinction is tightly bound to the circumstances in which it occurred. After extinction, a cue carries two meanings: it predicted danger, and it later predicted nothing. Organisms resolve this ambiguity by consulting context. From an adaptive standpoint the bias makes sense. An animal that discarded a predator association after a few safe encounters in one location would pay heavily when it moved elsewhere.

Renewal is the clearest demonstration. In the canonical ABA design, fear is acquired in context A, extinguished in context B, and tested back in A, where it returns robustly. ABC and AAB designs also produce renewal, though typically weaker. This shows that simply leaving the extinction context is sufficient, not only returning to where fear was learned. Clinically, the therapist's office is context B. A patient's home, workplace and daily routes function as A or C.

Context is defined broadly, and that is where the problem deepens. Interoceptive states qualify, including drug states, hormonal milieu, arousal and mood. Bouton, Kenney and Rosengard (1990) showed that extinction conducted under benzodiazepines was state-dependent, with fear renewing once the drug was withdrawn. Spontaneous recovery can be understood the same way: elapsed time is itself a context shift. On this account, recovery over time is a form of temporal renewal rather than a separate mechanism.

Reinstatement fits the framework too. Unsignalled presentations of the aversive outcome after extinction restore fear to the cue. Bouton and Bolles (1979) showed that in rodents this depends substantially on where those presentations occur. The surrounding context acquires aversive value, which then raises responding to the cue tested there. The clinical analogue is uncomfortable but familiar. A stressful life event, an unexpected panic attack, or a minor accident can reinstate fears that seemed resolved months earlier.

The hippocampus appears to perform the gating. Corcoran and Maren (2001) found that inactivating the dorsal hippocampus abolished renewal in rats. In humans, Kalisch et al. (2006) and Milad et al. (2007) reported hippocampal–vmPFC coupling during successful extinction recall in the extinction context. The safety memory is retrieved on condition of a contextual match. Anything that weakens that match, or stress that degrades prefrontal function through catecholamine and glucocorticoid surges, tilts the competition back toward the original fear trace.

Takeaway

Extinction answers the question of whether a cue is dangerous with a qualified answer: not here, not now. The brain takes that qualifier literally.

If the clinical problem is a fragile, context-bound inhibitory memory, two broad levers follow. One is to strengthen consolidation of the extinction trace. The other is to broaden the conditions under which it can be retrieved. Pharmacological approaches have mostly targeted the first lever and behavioural approaches the second. A third, more radical strategy aims to modify the original fear memory itself, challenging the dual-memory account at its foundation.

D-cycloserine, a partial agonist at the NMDA receptor glycine site, is the most studied enhancer. Walker and Davis (2002) showed that it facilitated extinction in rats, and Ressler et al. (2004) reported augmented exposure outcomes in acrophobia. Enthusiasm was later tempered. Mataix-Cols et al. (2017), in an individual-participant meta-analysis, found only small advantages. One mechanistic caveat matters: D-cycloserine appears to enhance consolidation of whatever was learned. After a session in which fear failed to decline, it may strengthen the wrong lesson.

Other candidates show similar promise with similar caveats. Glucocorticoid administration enhanced exposure outcomes in some phobia trials from de Quervain's group, likely by acting on consolidation and on retrieval of fear. Noradrenergic manipulation with yohimbine has produced mixed results. Sleep after training supports extinction consolidation, and aerobic exercise may help via BDNF signalling. Individual differences are substantial. Soliman et al. (2010) found impaired extinction in carriers of the BDNF Val66Met variant, a reminder that responsiveness is partly genetically constrained.

Behaviourally, Michelle Craske's inhibitory learning approach reorganises exposure around retrieval rather than habituation. Within-session fear reduction turns out to predict long-term outcome poorly. Expectancy violation, the mismatch between what the patient predicts and what happens, appears more important. Practical techniques follow from this. Clinicians can conduct exposure in multiple contexts, which reduced renewal in Vansteenwegen et al. (2007), and combine feared cues to deepen extinction, following Rescorla (2006). They can also remove safety behaviours, introduce variability, and provide retrieval cues that patients carry beyond the clinic.

The most provocative strategy exploits reconsolidation. Monfils et al. (2009) in rats, and Schiller et al. (2010) in humans, briefly reactivated a fear memory and then conducted extinction within the subsequent lability window. Fear appeared resistant to spontaneous recovery and reinstatement. Kindt et al. (2009) reported comparable effects with propranolol given at reactivation. Replication has been inconsistent, however, and boundary conditions such as memory age, strength and prediction error remain poorly specified. If the effect proves reliable, it would mean updating the original trace rather than merely competing with it.

Takeaway

The aim of exposure may not be to lower fear during the session, but to build a safety memory robust enough to be retrieved anywhere, at any time.

The dual-memory account reframes extinction as a retrieval problem as much as a learning problem. The fear trace persists. The safety trace competes. Context, time, internal state and stress determine which one is expressed. Relapse after exposure therapy is therefore less an anomaly than a default outcome of this architecture.

Several questions remain open. How exactly does the hippocampus compute contextual match, and can that computation be biased pharmacologically toward retrieving the safety trace? Under what boundary conditions does reconsolidation-based updating reliably modify the original association? And can individual markers, from vmPFC structure to BDNF genotype, predict who will need augmented approaches rather than standard exposure?

There is also a deeper theoretical issue. If memories are rarely erased, only outcompeted, then recovery from fear disorders resembles a continuing negotiation more than a cure. The research frontier lies in learning how to tip that negotiation durably, and in making safety as portable as fear already is.