Bastin et al. attempt to reconcile a plethora of different models concerning brain networks for the encoding, consolidation, and retrieval of episodic memory. There is much to admire, especially as a voice is given to many different ideas that have evolved over the past 20 years. The emphasis is on picking out common themes and bringing them together. A related theme is to look beyond the temporal lobe, to integrate parietal and frontal areas, as well as the medial diencephalon. These represent significant achievements, not least because they create bridges between the memory network models largely derived from classic neuropsychology, for example, the study of amnesic patients with confirmed brain pathology (which highlights the “relational representational core system”) and wider network models, largely derived from functional magnetic resonance imaging (fMRI). An almost inevitable cost of this amalgamation is that elements of the current model remain underspecified, leaving them difficult to test.
The integrative memory model is embedded within dual-process models, which distinguish recollection from familiarity. It is presumed that the principal signal for familiarity comes from perceptual fluency, the change in threshold for information that has been previously experienced. This is an odd choice for several reasons. First, as stated, some amnesics can show seemingly intact repetition priming yet catastrophic recognition memory. Second, experiments showing that perceptual fluency can contribute to recognition often require special constraints. Third, electrophysiological studies, starting from the pioneering work of Malcolm Brown, reveal that for visual information, at least, some neurons in the rhinal cortex reduce their firing following stimulus repetition. This attenuated activity is not only seen in single-cell recordings in animals but is also present in fMRI studies (Aggleton & Brown Reference Aggleton and Brown2006). The reduced activity in these neurons, thought to reflect long-term depression (Griffiths et al. Reference Griffiths, Scott, Glover, Bienemann, Ghorbel, Uney, Brown, Warburton and Bashir2008), would be sufficient to solve both familiarity and recency judgments. At the same time, it would be odd to categorise this signal reduction as perceptual fluency, as the latter would be expected to increase activity on stimulus repetition, given the fall in threshold. While it is the case that some perirhinal neurons may increase their firing gradually after hundreds of stimulus repetitions in test conditions associated with reinforcement (Holscher et al. Reference Holscher, Rolls and Xiang2003), this methodology is a far cry from cognitive tests of recognition memory. Although I agree with the authors’ statement that “lesions to the perirhinal cortex will not necessarily affect all forms of familiarity” (target article, section 4.2.1, para. 1), their model places undue emphasis on what is probably a subsidiary process (perceptual fluency).
The authors also discuss distinctions between “process” and “representational” models within the medial temporal lobe. Process models emphasise the computational properties of a structure (e.g., pattern separation by the hippocampus) whereas representational models emphasise the different kinds of information available in different brain sites (e.g., context-rich information in the hippocampus versus context-sparse information in perirhinal cortex). This distinction has been previously made, but represents a false dichotomy. Those espousing “representational” models surely do not presume that changes in representation happen by magic, they arise from the novel connections and architecture that permit different computations in different areas. At the same time, the dense, reciprocal interconnections between different medial temporal lobe sites result in the shared ownership of some representations (and processes).
Central to the integrative memory model is the “relational representation core system” (see sect. 3). The key components of this system are the hippocampus, mammillary bodies, anterior thalamic nuclei, and their interconnections – the “extended hippocampal system” (Aggleton & Brown Reference Aggleton and Brown1999). These medial diencephalic interconnections are presumed to help build the memory trace, in which item and context are bound. Recollection then emerges preferentially from reactivation of traces within this system. Surprisingly little evidence is provided by Bastin et al. for this core system, yet animal models and the analysis of patients with colloid cysts have proved most insightful. To take the latter, it has been repeatedly shown that interruption of the fornix (which provides hippocampal inputs to both the anterior thalamus and mammillary bodies, among other sites) is sufficient to cause an anterograde amnesia that preferentially impairs recollection (Vann et al. Reference Vann, Tsivilis, Denby, Quamme, Yonelinas, Aggleton, Montaldi and Mayes2009b). The resulting losses in recollection, but not familiarity, correlate closely with the extent of mammillary body atrophy (Tsivilis et al. Reference Tsivilis, Vann, Denby, Roberts, Mayes, Montaldi and Aggleton2008). Renewed interest in the mammillary body–anterior thalamic axis has provided novel insights into the memory loss in conditions such as developmental amnesia (Dzieciol et al. Reference Dzieciol, Bachevalier, Saleem, Gadian, Saunders, Chong, Banks, Mishkin and Vargha-Khadem2017), Korsakoff's syndrome (Segobin et al. Reference Segobin, Laniepce, Ritz, Lannuzel, Boudehent, Cabe, Urso, Vabret, Eustache, Beaunieux and Pitel2019), thalamic vascular damage (Carlesimo et al. Reference Carlesimo, Lombardi and Caltagirone2011), and Alzheimer's disease (Aggleton et al. Reference Aggleton, Pralus, Nelson and Hornberger2016).
Even less consideration is given in the integrative memory model for why these two medial diencephalic structures are so critical. It appears that these particular structures provide key information for memory encoding that otherwise would not reach the hippocampus (Aggleton et al. Reference Aggleton, O'Mara, Vann, Wright, Tsanov and Erichsen2010). If we just focus on the anterior thalamic nuclei, we can see that afferents potentially matching the above criteria arise from the mammillary bodies, parts of the frontal lobe (especially more dorsal areas), the reticular thalamic nucleus, and Gudden's tegmental nuclei (via the mammillary bodies). These inputs can interact with hippocampal processing via projections from the anterior thalamic nuclei to hippocampal and parahippocampal areas. A related possibility is that anterior thalamic and hippocampal efferents converge on a third site, for example, retrosplenial cortex, where their combined interaction is critical for memory. The discovery of spatial cells in the rat anterior thalamus (Jankowski et al. Reference Jankowski, Passecker, Islam, Vann, Erichsen, Aggleton and O'Mara2015) adds weight to the idea that these diencephalic processes involve individual mnemonic representations, as suggested by the integrative memory model. Key questions remain as to why there is an apparent duplication of information across medial diencephalic and temporal structures, allied to the need to test their independence and interdependence.
You have
Access
Bastin et al. attempt to reconcile a plethora of different models concerning brain networks for the encoding, consolidation, and retrieval of episodic memory. There is much to admire, especially as a voice is given to many different ideas that have evolved over the past 20 years. The emphasis is on picking out common themes and bringing them together. A related theme is to look beyond the temporal lobe, to integrate parietal and frontal areas, as well as the medial diencephalon. These represent significant achievements, not least because they create bridges between the memory network models largely derived from classic neuropsychology, for example, the study of amnesic patients with confirmed brain pathology (which highlights the “relational representational core system”) and wider network models, largely derived from functional magnetic resonance imaging (fMRI). An almost inevitable cost of this amalgamation is that elements of the current model remain underspecified, leaving them difficult to test.
The integrative memory model is embedded within dual-process models, which distinguish recollection from familiarity. It is presumed that the principal signal for familiarity comes from perceptual fluency, the change in threshold for information that has been previously experienced. This is an odd choice for several reasons. First, as stated, some amnesics can show seemingly intact repetition priming yet catastrophic recognition memory. Second, experiments showing that perceptual fluency can contribute to recognition often require special constraints. Third, electrophysiological studies, starting from the pioneering work of Malcolm Brown, reveal that for visual information, at least, some neurons in the rhinal cortex reduce their firing following stimulus repetition. This attenuated activity is not only seen in single-cell recordings in animals but is also present in fMRI studies (Aggleton & Brown Reference Aggleton and Brown2006). The reduced activity in these neurons, thought to reflect long-term depression (Griffiths et al. Reference Griffiths, Scott, Glover, Bienemann, Ghorbel, Uney, Brown, Warburton and Bashir2008), would be sufficient to solve both familiarity and recency judgments. At the same time, it would be odd to categorise this signal reduction as perceptual fluency, as the latter would be expected to increase activity on stimulus repetition, given the fall in threshold. While it is the case that some perirhinal neurons may increase their firing gradually after hundreds of stimulus repetitions in test conditions associated with reinforcement (Holscher et al. Reference Holscher, Rolls and Xiang2003), this methodology is a far cry from cognitive tests of recognition memory. Although I agree with the authors’ statement that “lesions to the perirhinal cortex will not necessarily affect all forms of familiarity” (target article, section 4.2.1, para. 1), their model places undue emphasis on what is probably a subsidiary process (perceptual fluency).
The authors also discuss distinctions between “process” and “representational” models within the medial temporal lobe. Process models emphasise the computational properties of a structure (e.g., pattern separation by the hippocampus) whereas representational models emphasise the different kinds of information available in different brain sites (e.g., context-rich information in the hippocampus versus context-sparse information in perirhinal cortex). This distinction has been previously made, but represents a false dichotomy. Those espousing “representational” models surely do not presume that changes in representation happen by magic, they arise from the novel connections and architecture that permit different computations in different areas. At the same time, the dense, reciprocal interconnections between different medial temporal lobe sites result in the shared ownership of some representations (and processes).
Central to the integrative memory model is the “relational representation core system” (see sect. 3). The key components of this system are the hippocampus, mammillary bodies, anterior thalamic nuclei, and their interconnections – the “extended hippocampal system” (Aggleton & Brown Reference Aggleton and Brown1999). These medial diencephalic interconnections are presumed to help build the memory trace, in which item and context are bound. Recollection then emerges preferentially from reactivation of traces within this system. Surprisingly little evidence is provided by Bastin et al. for this core system, yet animal models and the analysis of patients with colloid cysts have proved most insightful. To take the latter, it has been repeatedly shown that interruption of the fornix (which provides hippocampal inputs to both the anterior thalamus and mammillary bodies, among other sites) is sufficient to cause an anterograde amnesia that preferentially impairs recollection (Vann et al. Reference Vann, Tsivilis, Denby, Quamme, Yonelinas, Aggleton, Montaldi and Mayes2009b). The resulting losses in recollection, but not familiarity, correlate closely with the extent of mammillary body atrophy (Tsivilis et al. Reference Tsivilis, Vann, Denby, Roberts, Mayes, Montaldi and Aggleton2008). Renewed interest in the mammillary body–anterior thalamic axis has provided novel insights into the memory loss in conditions such as developmental amnesia (Dzieciol et al. Reference Dzieciol, Bachevalier, Saleem, Gadian, Saunders, Chong, Banks, Mishkin and Vargha-Khadem2017), Korsakoff's syndrome (Segobin et al. Reference Segobin, Laniepce, Ritz, Lannuzel, Boudehent, Cabe, Urso, Vabret, Eustache, Beaunieux and Pitel2019), thalamic vascular damage (Carlesimo et al. Reference Carlesimo, Lombardi and Caltagirone2011), and Alzheimer's disease (Aggleton et al. Reference Aggleton, Pralus, Nelson and Hornberger2016).
Even less consideration is given in the integrative memory model for why these two medial diencephalic structures are so critical. It appears that these particular structures provide key information for memory encoding that otherwise would not reach the hippocampus (Aggleton et al. Reference Aggleton, O'Mara, Vann, Wright, Tsanov and Erichsen2010). If we just focus on the anterior thalamic nuclei, we can see that afferents potentially matching the above criteria arise from the mammillary bodies, parts of the frontal lobe (especially more dorsal areas), the reticular thalamic nucleus, and Gudden's tegmental nuclei (via the mammillary bodies). These inputs can interact with hippocampal processing via projections from the anterior thalamic nuclei to hippocampal and parahippocampal areas. A related possibility is that anterior thalamic and hippocampal efferents converge on a third site, for example, retrosplenial cortex, where their combined interaction is critical for memory. The discovery of spatial cells in the rat anterior thalamus (Jankowski et al. Reference Jankowski, Passecker, Islam, Vann, Erichsen, Aggleton and O'Mara2015) adds weight to the idea that these diencephalic processes involve individual mnemonic representations, as suggested by the integrative memory model. Key questions remain as to why there is an apparent duplication of information across medial diencephalic and temporal structures, allied to the need to test their independence and interdependence.
Acknowledgments
This work was supported by the Wellcome Trust (Grant 103722/Z14/Z).