Bastin et al.’s integrative memory model of recognition memory conceptually separates the subjective experience, which is created in an attribution system, from the memory contents, which are stored in and retrieved from core memory modules. Basing recognition memory on generic perceptual-mnemonic systems seems more appropriate to us than postulating two dedicated memory systems based on differences in phenomenology (Brown & Aggleton Reference Brown and Aggleton2001). Since phenomenology is private to the individual, it cannot confer an evolutionary benefit and, therefore, evolution cannot select for memory systems based on subjective experiences (Cheng et al. Reference Cheng, Werning and Suddendorf2016; Suddendorf & Corballis Reference Suddendorf and Corballis1997). In the integrative memory model, the memory modules have qualitatively different functions: The entity representation core system processes single items and the relational representation core system processes the relationships between the items. The integrative memory model postulates that retrieval from the former is always associated with familiarity, but familiarity can also arise from the latter, when certain types of stimuli, for example, images of scenes, are used. By contrast, recollection arises only based on the relational representation core system. However, memory retrieval from a particular system by itself is not sufficient to account for the phenomenology, according to the integrative memory model. Instead, subjective experiences of familiarity and recollection are generated by a separate attribution system that evaluates the retrieved memory.

Although we embrace the diversity of memory systems (Werning & Cheng Reference Werning, Cheng, Bernecker and Michaelian2017), we argue that memory modules might differ in ways other than those considered by Bastin et al., and that other differences are potentially more consistent with the available evidence on recognition memory. Memory modules can (1) have qualitatively different functional properties, (2) receive different inputs, and (3) have different quantitative properties. The integrative memory model considers the first two cases. Here, we present the confounds of their assumptions and discuss the third possibility.

The integrative memory model predicts that the perirhinal and entorhinal cortices are part of the entity representation core system, which performs entity pattern separation, that is, distinguishing between similar stimuli based on conjunctive representations. The authors suggest that experiments manipulating the similarity between targets and lures can test the role of the perirhinal/entorhinal damage on recognition performance. Such studies exist. When lures were highly similar to targets, recognition performance of aged individuals with mild cognitive impairment (MCI) and Alzheimer's disease (AD) is indeed impaired, compared to age-matched controls (Westerberg et al. Reference Westerberg, Paller, Weintraub, Mesulam, Holdstock, Mayes and Reber2006). However, the deficits might arise from comorbid hippocampal damage in the early stages of AD and in MCI (Du et al. Reference Du, Schuff, Amend, Laakso, Hsu, Jagust, Yaffe, Kramer, Reed, Norman, Chui and Weiner2001) rather than from perirhinal damage, or from hippocampal impairment due to aging (Raz et al. Reference Raz, Lindenberger, Rodrigue, Kennedy, Head, Williamson, Dahle, Gerstorf and Acker2005), resulting in difficulty in distinguishing similar items (Stark et al. Reference Stark, Yassa, Lacy and Stark2013).

Moreover, evidence suggests that it is the hippocampus that is important for distinguishing highly correlated items. In the Westerberg et al. (Reference Westerberg, Paller, Weintraub, Mesulam, Holdstock, Mayes and Reber2006) paradigm, patients with selective hippocampal lesions rejected highly related lures less frequently than healthy controls (Bayley et al. Reference Bayley, Wixted, Hopkins and Squire2008; Holdstock et al. Reference Holdstock, Mayes, Roberts, Cezayirli, Isaac, O'Reilly and Norman2002), whereas recognition performance with unrelated lures is often preserved. In agreement with these findings, theoretical work concludes that representational overlap in cortex is higher than in the hippocampus (Greve et al. Reference Greve, Donaldson and van Rossum2010; Norman & O'Reilly Reference Norman and O'Reilly2003). These experimental and theoretical results seem to oppose the predictions of the integrative memory model.

The second possibility is that memory modules differ in their inputs. According to the dual stream model (Mishkin et al. Reference Mishkin, Ungerleider and Macko1983), perirhinal cortex processes object information (“what” stream), while parahippocampal cortex receives spatial inputs (“where” stream). Information from both streams converges in the hippocampus (Beer et al. Reference Beer, Vavra, Atucha, Rentzing, Heinze and Sauvage2018). Because, both perirhinal cortex and hippocampus receive object information and almost all recognition memory experiments employ visual stimuli in the same physical location, difference in inputs cannot account for possible differences in phenomenology in recognition memory task. In contrast to Bastin et al., we regard images of scenes as “what” information, which is quite different from information about the animal's current location (“where” information) (Azizi et al. Reference Azizi, Schieferstein and Cheng2014; Neher et al. Reference Neher, Azizi and Cheng2017).

Finally, memory modules can differ in their quantitative properties. The phenomenology of familiarity and recollection, in principle, could be generated within a single type of memory module, for example, in a memory retrieval process with attractor dynamics (Greve et al. Reference Greve, Donaldson and van Rossum2010). Specifically, after the presentation of a retrieval cue, the state of the memory network is updated until it converges to an attractor state. The success of retrieval depends on the attractor landscape. If the attractor state is veridical, it contains indices to neocortical representations providing additional details in the spirit of the hippocampal indexing theory (Fang et al. Reference Fang, Demic and Cheng2018a; Reference Fang, Rüther, Bellebaum, Wiskott and Cheng2018b; Teyler & DiScenna Reference Teyler and DiScenna1986; Teyler & Rudy Reference Teyler and Rudy2007). If these keys lead to the retrieval of meaningful information, the retrieved details are assigned higher weights and lead to high-confidence recollective experiences. However, if the attractor state is spurious, then either no details are retrieved or the retrieved information seems improbable. So, a familiarity response is generated with a strength depending on the depth of the attractor state. Therefore, high- and low-confidence responses can rely on familiarity and recollection (Ingram et al. Reference Ingram, Mickes and Wixted2012) depending on the attractor depth, the amount of recollective details, and the consistency of the details. This suggestion is akin to the one in the integrative memory model that the attribution system assesses the amount of mnemonic information and leads to recognition phenomenology based on the relevance and strength of retrieved details.

In conclusion, we suggest that the perirhinal cortex gives rise to familiarity more often, because the attractors are shallower due to weaker plasticity, and the network is more prone to generating spurious attractors due to higher noise or less robust representations. By contrast, the hippocampus has stronger plasticity and is less prone to generate spurious attractors, consistent with its specialization for one-shot encoding of episodic memories (Cheng Reference Cheng2013; Cheng & Werning Reference Cheng and Werning2016).