The human brain forms episodic memories by integrating information about 'what', 'where', and 'when' an event occurred. While rodent and primate studies have illuminated aspects of these representations in the medial temporal lobe (MTL), the precise neuronal mechanisms in humans remain unclear. This study aimed to address this gap by directly recording single-neuron activity in the human MTL during an associative memory task involving item-location pairings. The researchers hypothesized that specialized neuron populations would encode the 'what' (item) and 'where' (location) components of the memories, with their firing rates reflecting successful encoding. Understanding these mechanisms is crucial for comprehending how our perceptions transform into lasting memories, particularly within the context of episodic memory formation which depends on linking various aspects of an experience into a cohesive whole. The MTL's role in this process is widely acknowledged, but the specifics of how this integration takes place at the neuronal level remains a central question within the field of memory research. This study aimed to directly investigate this, providing much-needed data from human participants, which bridges the gap between animal models and the complex human memory system.

Existing literature highlights various spatial representations in the MTL, including hippocampal place cells and entorhinal grid cells. Studies in rodents and primates also demonstrate neurons sensitive to temporal aspects of events. While different models exist regarding the organization of information within the hippocampus (ranging from domain-specific to event-sequence-based), strong parallels between place and time cells suggest a less domain-specific organization. Research using fMRI and MEG in humans has provided support for grid-like representations in the entorhinal cortex during virtual navigation. However, human entorhinal neurons' involvement in scene and spatial information processing remains debated. Conversely, parahippocampal activity has been clearly linked to spatial navigation in 3D tasks, and evidence supports the existence of allocentric and egocentric spatial representations in the human parahippocampal cortex (PHC) and hippocampus. The concept of 'concept cells' – neurons selectively responding to specific concepts – has also emerged, with these cells potentially acting as building blocks of memory. However, previous studies on selective MTL activity in memory tasks yielded inconsistent results, often showing population-level effects but lacking consistent single-neuron findings. This study sought to resolve these inconsistencies by focusing on subpopulations of visually selective neurons and their involvement in memory encoding.

The study utilized an associative memory paradigm with a 3x3 grid, presenting images at different locations. Participants were required to recall item-location associations. Data were recorded from 361 signed multi-units in 13 neurosurgical patients with implanted depth electrodes in the amygdala, hippocampus, entorhinal cortex (EC), and prefrontal cortex (PFC). The difficulty of the task was adaptively adjusted during the experiment, maintaining performance around 50% accuracy. This involved adjusting presentation duration and set size (number of images to memorize) in real-time. Trials were classified as either subsequently remembered or forgotten. A screening session preceded the main experiment to identify response-eliciting images for each participant. The researchers analyzed neuronal activity during encoding and retrieval trials using linear mixed-effects models and various statistical tests, such as binwise rank-sum tests, binomial tests, and cluster permutation tests, to determine neuron responsiveness, compare remembered and forgotten trials, and test for subsequent memory effects. Responses to items and locations were assessed separately. A 15-second counting task was inserted between encoding and retrieval phases to prevent rehearsal. The total number of neurons recorded was 3681 (168 single units and 1863 multi-units) from 51 electrodes, across hippocampus, amygdala, EC and PHC. The signal was recorded at 32 kHz using a Neuralynx ATLAS system and spikes were extracted and sorted using CombinatoR™. Statistical analyses included linear mixed-effects models to examine the relationship between memory performance and experimental parameters (set size, trial duration, reaction time) while controlling for individual differences. Further analyses investigated the effects of adaptation, memory interference, and verified the consistency of preferred stimuli during retrieval.

The study revealed a significant subsequent memory effect (SME) in distinct neuron populations. Item-selective neurons in the amygdala, hippocampus, and EC showed higher firing rates during encoding for subsequently remembered items compared to forgotten ones. This effect was observed in a time window after the initial peak activity. Importantly, this effect was not seen in the parahippocampal cortex (PHC). Conversely, a significant proportion of neurons in the PHC were location-selective, exhibiting higher firing rates during encoding for locations of subsequently remembered items. This SME in location-selective neurons occurred at a later time window than the item-selective SME. Notably, the PHC was the only brain region with a significantly larger proportion of location-selective neurons compared to other regions. There was no significant reactivation of neurons during the delay period (15-second counting task). Control analyses showed that excluding neurons classified as both item and location neurons did not substantially alter the main findings. The study also confirmed that preferred stimuli remained consistent during retrieval trials, reinforcing the encoding specificity of the identified neurons. The PHC showed a high proportion (80%) of location-selective neurons, significantly more than the amygdala or hippocampus, and their activity was strongly associated with location during both encoding and retrieval.

The findings support the hypothesis that distinct neuronal populations in the MTL contribute to the 'what' and 'where' aspects of episodic memory. Item-selective neurons in the amygdala, hippocampus, and EC appear to provide the 'what' information, consistent with the concept of concept cells serving as building blocks of memory. Parahippocampal location neurons seem to encode the 'where' information. This is supported by the higher proportion of location-selective neurons found specifically in the PHC, along with their consistent response to locations during both encoding and retrieval. The hippocampal indexing theory offers a possible framework to interpret these findings. Here, item-selective neurons act as pointers to neocortical representations, while parahippocampal location neurons may directly represent spatial information. The later time window of the SME for location responses compared to item responses could reflect the hierarchical processing of information, with 'what' being processed before 'where'. The lack of significant reactivation during the delay period suggests that rehearsal was effectively prevented by the counting task. The data suggest that distinct neuronal populations in the MTL, rather than a single population showing generalized activity, are critically involved in encoding different features of an experience. Furthermore, the findings support the differential roles of hippocampus and PHC in episodic memory encoding, emphasizing their specialization in 'what' and 'where' information respectively.

This study provides compelling evidence for the specialized roles of distinct neuronal populations in the human MTL for encoding the 'what' and 'where' aspects of episodic memories. Concept cells in the hippocampus, amygdala, and EC appear crucial for encoding item information, while location-selective neurons in the PHC are responsible for encoding spatial context. The findings support the hippocampal indexing theory and highlight the importance of considering hierarchical processing of different information components during memory formation. Future research should investigate the temporal aspect of memory encoding and explore the interactions between these distinct neuronal populations and neocortical networks involved in long-term memory storage.

The study sample was limited to 13 neurosurgical patients, which could affect the generalizability of the findings. The task design, although carefully controlled, might not fully capture all aspects of natural memory formation. Also, the focus on visually presented stimuli and the specific associative memory task limits the extent to which the results can be generalized to other sensory modalities or memory tasks.