The transient period of memory instability that can be triggered when memories are retrieved under certain conditions offers an opportunity to modify the maladaptive memories at the heart of substance use disorders (SUDs). However, very well-learned memories (such as those in excessive drinking and alcohol use disorders) are resistant to destabilisation when retrieved or may not destabilise at all. Memory retrieval and intervention procedures that reliably destabilise and update maladaptive motivational memories may help to improve the long-term treatment of SUDs. In 59 hazardous drinkers, we tested a novel retrieval procedure for destabilising well-learned cue-drinking memory networks that maximises prediction error (PE) via guided expectancy violation during retrieval of these memories. This was compared with a retrieval procedure without PE and no-retrieval controls. We subsequently counterconditioned alcohol cues with disgusting tastes and images in all groups and assessed responding to alcohol stimuli 1 week later. Counterconditioning following PE retrieval produced generalised reductions in oculomotor attentional bias, explicit valuation and outcome expectancies in response to alcohol cues 1 week after intervention, evidence of updating of distributed motivational drinking memory networks. These findings demonstrate that well-learned cue-drinking memories can be destabilised and that learning history need not constrain memory destabilisation if PE is maximised at retrieval. Broad rewriting of diverse aspects of maladaptive memory by counterconditioning is achievable following this procedure. The procedure described may provide a platform for the development of novel memory-modifying interventions for SUDs.