Authors: Xiao Xu, Kechen Du, Dun Mao
Categories: Neuroscience, Neurophysiology, Research Article, SciAdv r-articles
Source: Science Advances
The primate hippocampus, crucial for both episodic memory and spatial navigation, remains an enigma regarding whether these functions share the same neural substrates. We investigated how identical hippocampal neurons in macaque monkeys dynamically shifted their representations between tasks. In a recognition memory task, a notable fraction of hippocampal neurons showed that rate modulation strongly correlated with recognition performance. During free navigation in an open arena, spatial view, rather than position, predominantly influenced the spatial selectivity of hippocampal neurons. Neurons selective for recognition memory displayed minimal spatial tuning, while spatially tuned neurons exhibited limited memory-related activity. These neural correlates of recognition memory and space were more pronounced in the anterior and posterior portions of the hippocampus, respectively. These opposing gradients extended further into the anterior and posterior neocortices. Overall, our findings suggest the presence of orthogonal long-axis gradients between recognition memory and spatial navigation in the hippocampal-neocortical networks of macaque monkeys.
Memory and navigation are two heavily intertwined cognitive processes. For example, memory aids our navigation in familiar surroundings, while exploring new environments helps us learn about the relationships among landmarks. The medial temporal lobe, especially the hippocampus, has been extensively implicated in both functions. The mnemonic view of the hippocampus stems from the famous patient H.M., who developed dense amnesia after hippocampal (and surrounding regions) removal (1, 2), whereas the spatial perspective originates from the discovery of hippocampal place cells in rats (3). These discoveries have fueled extensive research into the fundamental role of the hippocampus across species, at both the behavioral and neural levels of analysis.
Rodent single-cell recordings and lesion studies have contributed a substantial part to our understanding of the central role of the hippocampus in memory and navigation. Apart from hippocampal place cells, robust spatial cell types have been found in surrounding structures (4, 5), lending support for the role of the hippocampal system in navigation. Population place cell responses are specific to each spatial context (6), potentially serving as a memory code distinguishing experiences. Hippocampal lesions in rodents impair navigation performance that is based on triangulation of distal landmarks (7, 8). Moreover, the hippocampus participates in a broader network implicated in recognition memory (9, 10). These findings prompt the question of whether the memory and navigation functions of the hippocampus share a common basis.
Similar to rodents, selective damage to the monkey hippocampus disrupts spatial learning and memory in navigation tasks (11, 12). However, unlike rodents, single-cell recordings in the primate hippocampus, especially during free navigation, are scarce, and the results remain inconclusive. While the marmoset hippocampus displays place cell–like activity on linear tracks (13), such responses are rare in macaques navigating open environments (14). Instead, many cells seem selective to the attended distal location, termed spatial view (or facing location) responses (14, 15). On the other hand, hippocampal lesions in monkeys disrupt their natural preference for novel stimuli in a visual paired-comparison (VPC) task, which tests recognition memory (16, 17). This task, requiring minimal reinforcement, taps into incidentally encoded memory and is more hippocampus dependent (18). Neuronal firing and oscillatory activity in the hippocampus correlate with VPC task performance (19, 20), affirming its role in recognition memory. However, few attempts have linked behavioral performance or single-cell activity between recognition and navigation tasks, leaving the core functions of the hippocampus not directly compared.
The function of the human hippocampus appears increasingly complex. Patients with hippocampal lesions typically exhibit episodic memory deficits but often retain navigation abilities (21, 22). They often can navigate around their neighborhoods either mentally or physically. To exclude the possibility that this partial preservation of navigation arises from partial amnesia, a recent study reports almost complete navigation ability in a patient determined as having no episodic memory (21), suggesting a decorrelation between memory and navigation. Similarly, comparisons of memory and navigation functional magnetic resonance imaging (fMRI) studies reveal little overlap in activated brain regions (23). These findings therefore challenge trending models which emphasize shared neural mechanisms underlying memory and navigation (24–26). New models are emerging, highlighting differences between memory and navigation (22, 27), considering both neural and cross-species variations. This prompts questions about how identical hippocampal neurons engage in both memory and navigation and whether the hippocampus prioritizes memory in primates while favoring navigation in rodents. Macaque monkeys offer a promising model to bridge this knowledge gap, combining rodent-like electrophysiology with human-like behavioral complexities.
This study aimed to reconcile the mnemonic and spatial perspectives of hippocampal functions and explore the engagement of broader hippocampal-neocortical networks. We conducted a multitask experiment in macaque monkeys, comparing the selectivity of identical hippocampal and neocortical neurons. Using a recognition memory task followed by a free navigation task, we found that most tuned neurons were specific to either task, with only a small fraction participating in both. This dissociation trend aligned with the hippocampal long axis, where the head favored recognition and the tail favored navigation. This gradient further extended to the anterior and posterior neocortices, highlighting a broader pattern.
We trained two macaque monkeys (Macaca fascicularis) in consecutive a VPC task in front of a screen and a free navigation task in an open arena (Fig. 1A). The VPC task evaluates the animal’s judgement of prior occurrence, which is believed to rely on the role of the primate hippocampus for incidental memory without reinforcement (18). Each trial comprised a sample period with a cumulative viewing time of 10 s, a 60-s delay with a black screen, and two 5-s test periods separated by a 5-s delay. In each trial, the animals were presented with a novel set of scene images. The positions of the repeated and novel stimuli were randomized and exchanged in tests 1 and 2. The animals performed the task naturally without being rewarded. After completing 30 trials in the VPC task (~1 hour), the monkeys transitioned to the free navigation task in an open arena while we were recording from the same neurons (Fig. 1A). We tracked the monkeys’ 6-degree-of-freedom head motion using reflective markers on the head cap (Fig. 1B). We recorded binocular eye movement and the front scene using a wireless eye tracker firmly attached to the head implants. A chronic microdrive with movable single-wire bundles or tetrodes was implanted. Wireless recordings were obtained by a neural logger from the entire hippocampal long axis as well as orbitofrontal (OFC) and retrosplenial (RSC) cortices (Fig. 1C and fig. S1). Electrode locations were reconstructed by registering preoperative MRI and postoperative computed tomography (CT) images (Fig. 1C).
Fig. 1. Experimental design and behavioral performance.(A) Experimental protocol. Monkeys were engaged in a VPC task while head-fixed in front of a screen, followed by a free navigation task in an open arena. A schematic raster plot depicts the same hippocampal neurons recorded across the two tasks. An example trajectory color-coded by horizontal head direction is shown. (B) Diagram of head and eye tracking devices and chronic microdrive. Reflective markers were placed on the head cap for motion capture. Monkeys wore a backpack that carried a wireless transmitter of the eye tracker. Binocular eye and scene videos were wirelessly transmitted. Chronic, large-scale microdrive was implanted for simultaneous recordings from multiple brain regions. Untethered neural recordings were obtained by a 256-channel neural logger. (C) Preoperative MRI images and postoperative CT images were registered to obtain electrode locations. White lines show electrode tracks. Electrodes covered multiple sites from the anterior to posterior hippocampus. (D) Mean probability of viewing novel and repeated images during tests 1 and 2 as a function of time from test onsets. Shaded areas indicate SEM across all sessions (n = 93 sessions) of the two monkeys. (E) Spatial coverage of various spatial variables. Mean occupancy color maps are shown for position, spatial view, and egocentric center (F, front; B, back; L, left; R, right). Shaded areas indicate SD across all sessions (n = 84 sessions) of the two monkeys. The top right diagram illustrates egocentric center.
The monkeys showed a robust preference for viewing novel images, spending over twice as much time on them during the first test period (Fig. 1D and fig. S2). This recognition performance declined in the second test, likely due to decreased novelty. Similar to prior research (19), we used the fraction of time spent viewing novel images to quantify the animals’ recognition performance. Recognition performance negatively correlated with saccadic eye movements (fig. S2), indicating longer fixations in trials with high recognition. In the open arena, the monkeys’ behavior covered a wide spatial range, although the occupancy was biased toward salient landmarks in the environment (Fig. 1E and fig. S3). Spatial view, measuring where the monkeys visually focused (fig. S3), is crucial for primate navigation. Recording from the same neurons enabled us to examine how neuronal correlates shifted between tasks, establishing a link between hippocampal functions in recognition memory and spatial navigation, now a knowledge gap in the field.
We first assessed neural encoding in the VPC task and pinpointed hippocampal neurons exhibiting selectivity toward recognition performance. Consistent with a previous study (19), albeit several hierarchies away from the visual system, certain macaque hippocampal neurons displayed robust responses upon stimulus onset (fig. S4). Among these neurons, some showed enhanced responses to visual stimuli, while others were suppressed. As mentioned above, we quantified the animals’ memory performance by assessing the fraction of time spent viewing novel stimuli in each trial. To gauge neuronal selectivity, we used a rate modulation index (the difference in mean firing rate between viewing novel and repeated stimuli divided by their sum). In line with a previous study (19), certain neurons exhibited a robust positive correlation between the rate modulation index and memory performance (Fig. 2A), whereas others showed a negative correlation (Fig. 2B). Overall, 23.1% of hippocampal neurons (n = 210 of 909) showed a significant correlation, with the vast majority of neurons (n = 197) displaying a positive correlation with memory performance (Fig. 2C). As the stimuli swapped locations and the novelty of the image diminished in test 2, neuronal correlates with behavior dropped accordingly, observed in both positively and negatively correlated neurons (Fig. 2, D to F, and fig. S4; positive from 0.66 ± 0.01 to 0.47 ± 0.02, P = 2.9 × 10^−16^; from 1.69 to 1.22; negative from −0.60 ± 0.04 to −0.05 ± 0.09, P = 3.8 × 10^−05^; from −0.43 to −0.06; mean ± SEM, paired-sample two-sided t test). Thus, we have identified a robust, straightforward rate code in the monkey hippocampus linked to recognition memory.
Fig. 2. Neuronal selectivity in the VPC task.(A) Left: Raster plot of an example neuron during test 1 across 30 trials, overlaid on the viewing colormap. Magenta: viewing novel; viewing repeat; viewing outside the image area. Right: Rate modulation index (the difference in firing rate between viewing novel and repeat divided by their sum) as a function of memory performance (the fraction of time viewing novel images) for all trials. (B) Same as (A) but for a negatively correlated neuron. (C) Distribution of the correlation coefficients between memory performance and rate modulation index for all significantly correlated neurons (n = 210 of 909 neurons). (D) Correlation coefficients in tests 1 and 2. Solid black and red lines connect the mean correlation coefficients between tests 1 and 2. Error bars indicate SEM across all neurons in the corresponding group (gray: positively correlated neurons, P = 2.9 × 10^−16^; negatively correlated neurons, P = 3.8 × 10^−05^; paired-sample two-sided t test). (E) Average rate modulation index at 10 equally spaced memory performance bins for positively correlated neurons. Lines are linear fit to the points. Error bars, SEM across neurons. (F) Same as (E) for negatively correlated neurons.
We then examined how the macaque hippocampus encoded spatial information during free navigation. We used a free foraging task during which rich spatial cell types have been identified in rodents. Macaque hippocampal neurons showed diverse selectivity when their responses were charted against various spatial variables (Fig. 3A and fig. S5). While tuning curve–based analyses have traditionally been used to identify spatially selective neurons in rodents, recent advancements in data analysis unveiled mixed selectivity in a large fraction of tuned neurons in both the rodent and monkey hippocampus and entorhinal cortex (14, 28). Using a multimodal model–based approach (Fig. 3B and fig. S5), which considered interdependence between variables, we found that allocentric spatial view—where the monkeys looked—rather than physical position tended to dominate spatial representation in putative principal cells (with wide waveforms) [log likelihood (LLH), P < 0.05, one-sided Wilcoxon signed rank test; see Materials and Methods for details] (Fig. 3C). Conversely, in putative interneurons (with narrow waveforms), movement speed was the primarily coded information (Fig. 3C and fig. S6), consistent with previous findings (14). Position was only secondarily encoded across hippocampal neurons. We were only able to observe rodent-like place fields in only a few cells (with sharp, sparse, and isolated firing fields). These findings align with earlier studies in marmosets and macaques during free navigation in open environments (14, 29).
Fig. 3. Neuronal selectivity during free navigation.(A) Tuning curves of four example neurons selective for spatial view, position, egocentric center, and horizontal head direction. Peak firing rates are indicated at the lower right of each colormap. Minimum firing rates are 0. (B) Top An example neuron showing tuning to spatial view, head tilt, and head height when using traditional tuning curve–based analysis. Bottom This neuron shows tuning to spatial view only when using a multimodal model–based analysis that considers interactions between variables. X axis shows different variable combinations. Red dot indicates the best model. Each gray dot represents the result from onefold during the fivefold cross-validation procedure. Black dots and error mean ± SEM across folds. Null model performance is at 0. Eight spatial variables were considered Pos, position; SV, spatial view; EC, egocentric center; HT, head tilt; AV, angular velocity; HH, head height; Spd, translational speed; HD, horizontal head direction. (C) Fraction of neurons selective to each variable for wide neurons (with wide waveforms and putative pyramidal cells) and narrow neurons (with narrow waveforms and putative interneurons) across the two monkeys, n = 84 sessions. (D) Circular graph showing interactions between variables for wide neurons. The thicker and the darker the line, the more often the connected variables were conjunctively coded in the same neurons. (E) From left to right, distribution of the preferred firing fields (red dots) for all neurons encoding spatial view, head direction (P = 0.07, Rayleigh test), and position superimposed on the average occupancy colormap. Red dot size corresponds to neuron count. Two opposite perspectives are shown for spatial view, so the entire arena is shown.
Many spatially tuned neurons exhibited mixed selectivity to multiple variables (mixed 124 neurons; single-variable 102), with the strongest interactions occurring between allocentric spatial view and head direction (Fig. 3D). The preferred spatial view “firing fields” spanned a broad spatial range without bias toward salient landmarks in the environment (Fig. 3E). Spatial view tuning covered not only the four walls but also the ceiling and floor of the arena, indicative of an overarching representation of the visual space of the entire environment. Head direction tuning did not display clear nonuniformity (P = 0.07, Rayleigh test). Position tuning tended to cluster near the arena boundaries. The above results demonstrate that the macaque hippocampus encodes a wide variety of spatial variables dominated by spatial view.
Next, we asked how these identical hippocampal neurons switch their representations across tasks. In a subset of sessions (monkey X, n = 41 sessions; monkey Q, n = 37 sessions), the monkeys shifted to the free navigation task directly after the VPC task while we recorded from the same cells. We recorded from a total of 330 identical hippocampal neurons across the two tasks (Fig. 4A). Hippocampal neurons displayed heightened activity during free navigation in the arena (Fig. 4B; 0.22, 2.97, and 19.41 Hz in VPC; 0.26, 3.55, and 26.77 Hz in navigation; 10, 50, and 90 percentiles, respectively; P = 3.0 × 10^−06^, paired, two-sided Wilcoxon signed rank test), while extracellular waveform features such as amplitude and temporal profiles remained stable across tasks (Fig. 4B and fig. S6).
Fig. 4. Neuronal selectivity between the VPC and navigation tasks.(A) Identical neurons recorded in the VPC and navigation tasks from an example session. Average waveforms and autocorrelograms of 12 identical neurons are shown. (B) Scatterplots showing firing rates (top, P = 3.0 × 10−^06^, paired two-sided Wilcoxon signed rank test) and trough-peak (TP) latency (bottom) of all identical neurons (n = 330) between the two tasks. (C) Top An example neuron showing significant correlation between rate modulation and memory performance but no spatial selectivity—a memory-only neuron. Bottom An example neuron showing no correlation between rate modulation and memory performance but with spatial selectivity—a space-only neuron. Tuning curves to four spatial variables are shown, including spatial view, head direction, position, and egocentric center. (D) Left: Contingency table showing encoding properties of all identical hippocampal neurons between the two tasks. Middle and Violin plots showing recognition correlation (P = 1.7 × 10^−06^) and spatial selectivity (P = 2.1 × 10^−12^) for all identified memory neurons and spatial neurons in the two tasks. The box plots within the violins mark the 25th, 50th, and 75th percentiles of the data. The whiskers extend to the most extreme data points that are not outliers. Data from monkey X (n = 153 neurons, 41 session pairs). (E) Same as (D) for monkey Q (n = 177 neurons, 37 session pairs; recognition P = 1.2 × 10^−12^; LLH P = 8.1 × 10^−06^; two-sided Wilcoxon rank sum test).
We identified neurons selective to either the VPC task and/or the navigation task using independent measurements mentioned in the previous two sections. Some neurons showed a significant correlation between rate modulation and memory performance without selectivity to any spatial variables (monkey X: n = 39 of 153 neurons; monkey Q: n = 33 of 177 neurons), whereas others displayed clear spatial receptive fields without selectivity to recognition memory (monkey X: n = 38 of 153 neurons; monkey Q: n = 50 of 177 neurons) (Fig. 4C). We refer to the former group as memory neurons, the latter as spatial neurons, and the remainder as memory and spatial neurons. Overall, the majority of tuned neurons were selective to either task, with approximately 10% of hippocampal neurons engaged in both recognition and navigation, which was below chance level (monkey X: n = 17 of 153 neurons, 11.1%, chance 16.3%; monkey Q: n = 17 of 177 neurons, 9.6%, chance 13.6%; alpha = 0.05) (Fig. 4, D and E). Memory neurons showed a significantly higher correlation between rate modulation and memory performance compared to spatial neurons, while the latter group contained significantly higher spatial information than the former (Fig. 4, D and E) (recognition monkey X, memory neurons 0.45, 0.61, and 0.86; spatial neurons 0.05, 0.36, and 0.74; P = 1.7 × 10^−06^; monkey Q, memory neurons 0.45, 0.59, and 0.74, spatial neurons 0.07, 0.20, and 0.40; P = 1.2 × 10^−12^. LLH monkey X, memory neurons −0.05, 0, and 0.02; spatial neurons 0.02, 0.19, and 1.36; P = 2.1 × 10^−12^; monkey Q, memory neurons −0.10, 0, and 0.04; spatial neurons 0.003, 0.03, and 0.32; P = 8.1 × 10^−06^; 10, 50, and 90 percentiles, respectively; two-sided Wilcoxon rank sum test). Thus, the majority of hippocampal neurons dynamically dissociated their involvement between recognition and navigation.
The primate hippocampus is topographically organized along its long axis in both anatomical and functional aspects (30, 31). Thus, we tested whether hippocampal engagement in recognition and navigation tasks was spatially organized. Our electrodes were distributed along the entire long axis, enabling us to categorize neurons into three groups based on their recording anterior (A), intermediate (I), and posterior (P) hippocampus (Fig. 5, A and B).
Fig. 5. Spatial organization of memory and space selectivity along the hippocampal long axis.(A) Left: Registered MRI and CT images showing electrodes as white tracks. The hippocampus was divided into anterior (A), intermediate (I), and posterior (P) portions along the long axis. Right: Stacked bar plots showing the fractions of all identical neurons with only memory selectivity, only space selectivity, or both effects along the long axis (P = 0.03, 6 df, χ^2^ = 14.1). Data from monkey X. (B) Same as (A) for monkey Q (P = 0.03, 6 df, χ^2^ = 9.3).
Upon examining identical neurons recorded in both tasks, an interesting trend emerged when plotting the fraction of neurons selective to recognition and/or navigation from the anterior to posterior hippocampus. Neurons selective to either task exhibited a divergent Memory-only neurons were more prevalent in the anterior-intermediate portions, while spatial coding progressively intensified toward the posterior region (Fig. 5, A and B) (memory-only neuron monkey X, A, I, and P: 0.35, 0.30, and 0.16; monkey Q, A, I, and P: 0.23, 0.21, and 0; space-only neuron monkey X, A, I, and P: 0.16, 0.13, and 0.36; monkey Q, A, I, and P: 0.18, 0.27, and 0.65; monkey X: P = 0.03, 6 df, χ^2^ = 14.1; monkey Q: P = 0.03, 6 df, χ^2^ = 9.3). Hence, it appears that the monkey hippocampus demonstrates a differential spatial organization along its long axis concerning its role in recognition memory and spatial navigation.
In addition, the primate hippocampus is also topographically connected to various neocortical regions (32). The anterior hippocampus projects to frontal cortices such as the OFC, medial prefrontal, and anterior cingulate cortices (33), while the posterior hippocampus primarily connects with posterior cortices like the RSC and posterior parietal cortices (34). Therefore, we asked whether such a gradient in neural coding for recognition and navigation extends to the anterior and posterior neocortices. Recognition coding was the strongest in the OFC, whereas overall spatial coding was most prominent in the RSC (Fig. 6) (memory-only neuron monkey X, HPC 0.25 and RSC 0.08; monkey Q, OFC 0.22, HPC 0.19, and RSC 0.05; space-only neuron monkey X, HPC 0.25 and RSC 0.41; monkey Q, OFC 0.29, HPC 0.28, and RSC 0.51; monkey X: P = 0.002, 3 df, χ^2^ = 15.1; monkey Q: P = 1.8 × 10^−04^, 6 df, χ^2^ = 26.5). Moreover, different regions showed preferences for distinct aspects of spatial The hippocampus and OFC leaned toward encoding allocentric spatial view, whereas the RSC showed a preference for encoding egocentric information such as movement speed (fig. S7). Consistent with the dichotomy between memory and spatial coding, nonmemory neurons displayed stronger spatial selectivity across regions (fig. S7).
Fig. 6. Spatial organization of neuronal selectivity between the two tasks in the hippocampal-neocortical networks.(A) Registered MRI and CT images showing electrodes as white tracks. Neurons were recorded from the RSC and hippocampus in monkey X. Neurons were recorded from the OFC, hippocampus, and RSC in monkey Q. (B) Fractions of all identical neurons in each region showing selectivity to both memory and space, memory only, and space only between the two tasks (P = 0.002, 3 df, χ^2^ = 15.1). Data from monkey X. (C) Same as (B) for monkey Q (P = 1.8 × 10^−04^, 6 df, χ^2^ = 26.5). (D) Diagram showing spatial organization of neural correlates with recognition and navigation along the OFC-HPC-RSC pathway.
In summary, these findings reveal opposing gradients between recognition and navigation along the extended anterior-posterior hippocampal-neocortical axis (Fig. 6D).
Memory and navigation are central functions often associated with the hippocampus, yet debates persist regarding its core function—whether it primarily serves memory, navigation, or both through a shared algorithm. Some argue for the primacy of space in mapping daily experience, suggesting that the role of the hippocampus in memory stems from its role in processing spatial information (35). Conversely, others propose that the hippocampus performs a general computation, encoding relationships among stimuli, thereby contributing to both spatial and nonspatial tasks (36, 37). Some human researchers, on the other hand, posit that memory and navigation represent distinct processes, with the primate hippocampus potentially evolving to play a more notable role in declarative memory (38). Related to this is the observation that posterior neocortices in humans showed stronger activation than the hippocampus during navigation tasks (23).
The current study aimed to reconcile the above debates and bridge insights from rodent and human studies using macaque monkeys as an animal model. Through a multitask experiment while recording from identical hippocampal neurons, we found that hippocampal coding for item memory (recognition) and space is topographically biased along the anterior-posterior axis, and such a gradient further extends to the anterior and posterior neocortices. However, while our study intended to compare the neural substrates of recognition memory and spatial navigation, the observed distinctions between item memory and space may have certain limitations when extrapolating to a general differential response between memory and navigation. Future studies should confirm or refute these differences using complementary tasks, such as a space recognition task using the same visual stimuli used during the navigation task, or a navigation task tapping into spatial memory. Furthermore, discrepancies in the coding properties of spatial neurons and the proportions of wide and narrow neurons between the two monkeys were noted. Uneven sampling from different subregions likely contributed to these discrepancies, although precise identification of recorded sites necessitates further histological examination.
There is a near consensus that the hippocampus is vital for memory, particularly episodic memory, which allows individuals to vividly relive past events. The seminal case of patient H.M. and decades of subsequent research have underscored the importance of the hippocampus to memory in both patients with amnesia and nonhuman primates. Recognition memory, a facet of episodic memory, is generally believed to contain two recollection and familiarity. Some have argued that the hippocampus is only critical for the recollective component of recognition, while others have found that the hippocampus supports both familiarity- and recollection-based recognition in humans (10). Studies using nonhuman primates have also produced controversial results using different recognition tests (16, 18). One line of research used VPC tasks which tested the monkey’s natural preference for novel stimuli. They found a robust link between hippocampal damage and performance decline. Another line of research used matching tasks which tested the monkey’s explicit choices matched or unmatched to samples. Unexpectedly, matching tasks often produced minor deficits or complete sparing following hippocampal lesions. It remains controversial whether the VPC tasks test the recollective component of recognition memory. However, it is likely that VPC measures incidental recognition memory, which appears to be particularly hippocampus dependent. Given its consistent reliance on the hippocampus, we opted to adapt the VPC task for the present study. Nonetheless, it remains to be seen whether these findings can be extrapolated to other recognition memory tests, or when recognition and navigation are combined into a single task.
Neurophysiological studies of the primate hippocampus in recognition memory are relatively scarce compared to behavioral studies. Using a VPC-like task with sequentially presented stimuli, Jutras and Buffalo (19) found firing rate modulations that were correlated with recognition performance in many hippocampal neurons (n = 30 of 131 neurons). Human studies have similarly documented hippocampal neuron firing rate changes following a single exposure to a visual stimulus (39, 40). Our results align with these previous findings by revealing firing rate modulations between viewing novel and repeated images. Since we used a novel set of images in each trial, the recognition-linked neural signals unlikely reflect specific features of a certain picture but rather a general, abstract signal related to recognition memory.
Besides the mnemonic perspective, another viewpoint regards the hippocampus as the hotspot for navigation in physical space, by performing (or inheriting) path integration and triangulation among environmental landmarks. Unlike rodents, behavioral studies using primates to establish the causal role of the hippocampus in spatial navigation are rare, with findings often appearing contradictory. Hippocampal lesions disrupted monkey’s ability to establish or use allocentric, but not egocentric, representations to guide behavior (11, 12). In humans, hippocampal damage often caused minor deficits in spatial navigation (41, 42), whereas RSC damage yielded more pronounced effects (43). Previous controversies may have arisen from the failure to distinguish between allocentric and egocentric components of navigation. This interpretation is consistent with current findings that the macaque hippocampus contains stronger allocentric spatial coding whereas the RSC cares more about egocentric components.
The view about the hippocampus in navigation is also largely fueled by the groundbreaking discoveries of hippocampal place cells and entorhinal grid cells in rodents. In the present study, the monkeys were highly familiar with the arena environment. The free foraging task adapted here is more like a beacon navigation task in which the monkeys navigated to clearly visible targets (food pellets). Similar tasks have been extensively used to reveal rich cell types in rodents. However, important distinctions need to be made when extrapolating spatial representations found in rodents to primates (44). Primates and rodents explore environments in fundamentally different Rodents typically move about to gather information in their proximity to guide future behavior; primates can visually inspect distal space and then decide whether to move or not. Therefore, it is likely that diurnal and nocturnal species exploit vision and other senses to different degrees during navigation. Diurnal primates can inspect their environment by frequent rapid head gaze shifts even when remaining stationary (29). This has an impact on spatial representations. The way the rodent hippocampal system represents physical location may be similar to how the primate hippocampal system represents visual space. Converging studies have started to show that the monkey hippocampus and entorhinal cortex encode visual space and heading more than self-position, both in restrained and freely moving monkeys (14, 15, 45, 46). We show here that, in line with previous studies, spatial view coding dominates spatial selectivity in the hippocampus and OFC.
Spatial view is an allocentric component in that the animals can view the same location from different perspectives, that is, with different self-position and head orientation. It remains to be tested what spatial view exactly it could be a memory code for the viewed space or the relationship between the viewed space and an object (that is, a vectorial representation). One recent study demonstrated that view-centered, large-scale background information may be one important input to spatial view representation (47). It would also be interesting to investigate what aspects of the spatial codes in the OFC and RSC are inherited from the hippocampus, and how they are integrated with local information processing. Furthermore, it is intriguing to see how egocentric first-person perspectives are transformed along the ventral and dorsal pathways into allocentric view representations in the hippocampus (48).
The above-discussed mnemonic and spatial views of the hippocampus prevail in our current understanding of its functions. These two lines of research have often been carried out in parallel, leaving a gap in our understanding of how these two seemingly different functions are linked, or whether they actually reflect the same computation (22, 24, 27, 49, 50). It also remains an important question whether there is a cross-species difference. On the one hand, there is no doubt that memory and navigation are connected cognitive processes, and relevant models have emphasized their common neural mechanisms (24). On the other hand, it has become increasingly clear that there are major differences between them. For example, patients with amnesia with bilateral hippocampal lesions showed intact path integration performance (51, 52), distinct from that in rodents. One case study reported that a patient with dense amnesia showed intact navigation (21). A meta-analysis using a large sample of fMRI studies revealed little to no overlap between brain regions activated by memory and navigation tasks (23). These findings suggest that memory and navigation depend on dissociable brain substrates, at least partially in primates. Our results are consistent with these human studies by further showing that neural correlates with recognition memory and spatial navigation tend to be allocated at different hippocampal portions, and such a dissociation further extends to anatomically corresponding neocortices.
A potentially fruitful way forward would be to compare the responses of identical hippocampal neurons across multiple tasks that are motivated by theoretical premises (53), as we did in the current study. Such a strategy has been underestimated in previous research, impeding our adjudication among the various functions the hippocampus may be implicated in. It would further our understanding of the flexibility and generalizability of the inner workings of the hippocampus.
A prominent anatomical feature of the primate hippocampus is that the anterior portion has more densely packed neurons, and its intrinsic connections are different from the body of the hippocampus (31) (the anterior/posterior hippocampus in primates corresponds to the ventral/dorsal hippocampus in rodents). This difference may underlie functional differentiations in the primate hippocampus along its long axis. The anterior and posterior hippocampus in humans maps large- and fine-scale spatial information, respectively (54). In memory domain, specific details and abstracted concepts are organized from the posterior to anterior hippocampus, respectively (55). Our results resonate with previous findings in humans by showing that the abstract neural correlates with recognition are more prevalent in the anterior, while specific spatial information is more strongly associated with the posterior hippocampus. This underscores the importance of appropriate recording location based on the research question at hand or adopting an unbiased approach encompassing the entire longitudinal axis. Failure to do so may lead to inaccurate or potentially misleading conclusions arising from uneven sampling.
The primate hippocampus is also topographically connected with various neocortical regions. For example, the posterior hippocampus is strongly connected with the RSC, while the anterior hippocampus sends moderate projections to the OFC. We found in the present study that neural correlates with recognition memory further increased from the anterior hippocampus to the OFC, and spatial code strengthened further from the posterior hippocampus to the RSC. These results are consistent with a recently proposed network model trying to explain the similarities and differences between episodic memory and navigation—the posterior medial/anterior temporal (PMAT) model (56). The PMAT model suggests that the anterior hippocampus, perirhinal cortex, and OFC contribute to item processing; whereas the posterior hippocampus, RSC, and parahippocampal cortex play a role in spatial contextual processing. In this model, the hippocampus is no longer considered as the top hierarchy in information processing but rather is a critical interface between the PM and AT systems, which can bind item and context to form memories. It would be interesting to examine how rapid item-context binding is accomplished within the hippocampal circuit, and how this signal can link distributed neocortical targets using feedback projections.
Two long-tailed macaques (M. fascicularis, female, 6 years old, 3 kg and male, 6 years old, 7 kg) were used in this study. All experimental procedures and surgeries were approved by the Biomedical Research Ethics Committee of the Institute of Neuroscience, Chinese Academy of Sciences and were in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals (CEBSIT-2021034R02).
We obtained CT (Samsung) and MRI (3T, Siemens) images for each monkey to segment the regions of interest and to design form-fitting head implants and microdrive. The CT and MRI volumes were manually registered to the stereotaxic coordinates in three-dimensional (3D) slicer software (57) (https://slicer.org). The hippocampus, RSC, and OFC were segmented from the T1-weighted MRI images, with reference to standard atlas (58, 59). Major blood vessels were segmented from the 3D time-of-flight MRI images, and the skull model was segmented from the CT images. On the basis of these segments, we customized each monkey’s head plate, heat ring, microdrive system, and head cap using SolidWorks (https://solidworks.com) and blender (https://blender.org/). The microdrive system consisted of four groups of electrodes allocated along the long axis of the right hippocampus, two groups of electrodes in the OFC, and two groups in the RSC. Each group consisted of 32 channels and was driven by one shuttle screw (350 μm per revolution). In monkey X, we used brush arrays (MicroProbes). Each array was made of 32 12.5-μm platinum-iridium wires that were fitted into a polymicro capillary tubing, which was further loaded into a stainless-steel guide tube. The wires extended beyond the polymicro tubing for 2 to 3 mm. In monkey Q, we used tetrode arrays. Each array consisted of eight tetrodes that were made of 17-μm nickel-chromium wires (California Fine Wire Co.). Each tetrode was loaded into a 100-μm polymicro capillary tubing (100 μm inner diameter, 170 μm outer diameter) (Molex). No guide tubes were used for the tetrode arrays. All electrodes were gold plated to reduce the impedance to approximately 250 kilohm. All implants were autoclaved or gas-sterilized (Ethylene Oxide) before surgery. CT scans were performed at least every 2 months to reconstruct current electrode locations.
Three sequential surgeries were performed on each animal under anesthesia, involving implanting the head plate, the head ring, and the microdrive system. In the first surgery, we implanted a form-fitting titanium plate fixed on the skull by 15 bone screws to offer a base for the following implants. The skin was sutured back to completely seal over the head plate. This procedure created an ideal environment for tight sealing between the head plate and bone screws and the skull (60). In the second surgery (at least 2 months after the first surgery), we implanted the head ring made of polyetheretherketone (PEEK), which was attached to the head plate by five M4 stainless-steel screws. The head ring was used for head fixation in the monkey chair and to house the microdrive system and later the recording system. We avoided using dental acrylic in the first two surgeries, therefore reducing the chance of granulation tissue growth and infections. In the third surgery, we implanted the microdrive system. The form-fitting microdrive chamber was attached to the skull and three anchor screws using C&B Metabond (Parkell) and dental acrylic. The microdrive body was then inserted into the chamber and hermetically sealed with the chamber using sterile silicone grease. For monkey X, the entire microdrive body and the electrodes were implanted as a single unit. The guide tubes rested at about 3 mm from the target regions. The electrodes were moved out of the guide tubes by shuttle screws once the animals recovered from surgery. For monkey Q, each tetrode was manually lowered to a depth of about 3 mm from the target regions. The tetrodes were then glued to the shuttles for future fine advancement. In the first implant of monkey Q, we were unable to record any neurons in the anterior hippocampus. We managed to replace the microdrive and record from the anterior hippocampus in the second implant.
The monkeys were placed on a food and water delayed schedule. They received juice rewards during the eye calibration procedure. They received food pellet rewards during the navigation task in the arena. The monkeys were trained to remain head-fixed for the whole period of the VPC task. We monitored the movement of the left eye using an infrared monocular tracker (ISCAN, Inc) during the VPC task. No formal training was required for the VPC task as it tested the monkey’s natural preference. In the free navigation task, we used a motion capture system (Vicon) combined with a portable wireless binocular eye tracker (ISCAN) to monitor the monkey’s 6-degree-of-freedom head motion and eye movements. The monkeys were trained to enter the arena from the chair and return voluntarily. They were habituated to the arena environment until they could explore the arena for randomly scattered food pellets throughout a ~1-hour session. They were trained to wear a jacket in which the backpack carried the battery (3.7 V) and transmitter of the eye tracker. They were accustomed to the eye tracker that was firmly fixed to the head ring implant. Eye images were reflected off two hot mirrors from which they could see through. Thus, the monkeys’ sight was not obstructed in any way. The eye tracker also contained a scene camera that captured the animal’s front view. Binocular eye (60 Hz) and scene (30 Hz) images were wirelessly transmitted to the receiver outside the arena. Three reflective markers were placed on the monkey’s head cap to capture the translational and rotational movements of the head. The 3D position (x, y, and z) of each marker was recorded at 100 Hz at submillimeter resolution. Start signals from the eye tracker and the motion capture system were recorded in the electrophysiological recording system to synchronize the data. The whole weight of the apparatus the monkeys carried during navigation is ~475 g. This did not restrict their head or body movements in any way.
We used a 256-channel neural logger to record broadband electrophysiological signals at 30 kHz (SpikeGadgets). The logger was connected to the microdrive via jumper cables. The logger was powered by a 3.7-V battery and had a 256-GB micro-SD card inserted to store the broadband raw signals. To avoid clock drift over time, the electrophysiological system regularly transmitted radio frequency signals to the logger for synchronization. Start signals of other behavioral apparatuses were recorded in the electrophysiological system, enabling us to merge the neural signals in the SD card with the behavioral data recorded on the local computers. We typically advanced the electrodes by ~40 to 180 μm each time, and recordings were performed the following day to allow the electrodes to fully stabilize.
The monkeys were head-fixed in a dark room, 48.5 cm from the monitor (Samsung, 1.016 m, 30 Hz). Eye position was tracked by an infrared eye tracker (ISCAN). The VPC task was controlled using MonkeyLogic 2.2.32 (61) (https://monkeylogic.nimh.nih.gov/). Each trial comprised three sample, delay, and two tests. The sample phase started with the presentation of a novel stimulus subtending 15° at the center of the screen. After the monkeys had accumulated a total viewing time of the image for 10 s, the image disappeared, and a dark screen was presented for 60 s (delay period). In the first test, two images were displayed simultaneously on the the image shown previously during the sample period (repeat) and a novel one that the monkey had not encountered before. The images were symmetrically arranged on both sides of the center of the screen, with a separation of 15°, in randomly generated positions and displayed for 5 s. Following a 5-s interval, the second test was conducted with the same two images as the first test, but their positions were exchanged. All stimuli in the experiment were obtained from the Free Gallery (www.unsplash.com). The monkeys were required to complete three blocks, each containing 10 trials—thus, a total of 30 trials per session. No reward was delivered during the task. The monkeys were rewarded with juice during eye calibration.
Before transferring the monkey to the free navigation task, we affixed the wireless binocular eye tracker onto the monkey’s head ring to conduct binocular calibration and a random fixation task using Monkeylogic. During the random fixation task, a yellow dot randomly appeared on the screen at different positions in each trial. The monkey was required to maintain fixation at the target spot for 200 ms to receive juice reward. The data from this task were used to calculate eye-in-head positions (in degrees) and the positions where the monkeys were looking in the arena. After the calibration, the monkey was released from the monkey chair to the arena to perform the free navigation task. During the free navigation task, monkeys searched for small food pellets (190 mg) randomly scattered throughout the session in an open arena (3.5 m by 3.5 m by 2.2 m) for about 1 hour. The arena was built using a framework of metal mesh that allowed monkeys to climb and was enclosed by green curtains on two sides. The monkey could see the decorated elements on the wall through the remaining two sides. At the top of the arena, there were two Arduino-controlled food dispensers positioned at opposite diagonal corners. These devices released food pellets at random intervals from 20 to 30 s. A motion capture system (Vicon) consisting of 12 infrared cameras was mounted in the arena’s interior. These cameras captured the real-time positions of three reflective markers attached to the monkey’s head cap, allowing us to extract spatial behavioral variables of the monkeys.
We used a semiautomatic procedure for the preprocessing of electrophysiological data. We used established pipelines for spike detection and spike sorting followed by manual curation (Kilosort 2.0, SpykingCircus, and Phy) (62, 63). Raw data were first band-passed (300 to 6000 Hz) for spike detection. Automatic spike sorting was performed using Kilosort 2.0 (https://github.com/MouseLand/Kilosort) and SpykingCircus (https://spyking-circus.readthedocs.io/en/latest/), followed by manual curation in Phy (https://github.com/cortex-lab/phy). Manual curation was based on spike waveforms, waveform features, auto- and cross-correlograms. Putative nonsomatic spikes (tri-phasic, positive spikes) were excluded from further analyses as we were not sure about the origins (local or from elsewhere) of these signals. Only well-isolated single units were included in further analyses. Neurons recorded on different days were considered unique. Neurons were classified as wide and narrow based on their spike waveform trough-peak (TP) latency (threshold at 0.5 ms). Neurons with wide waveforms (TP ≥ 0.5 ms) were considered putative pyramidal cells. Neurons with narrow waveforms (TP < 0.5 ms) were considered putative interneurons.
For each neuron, we calculated its mean firing rates during the 600-ms windows before (pre-stimulus) and after stimulus (post-stimulus) onset in the sample phase of each trial. A neuron was considered visually responsive if pre-stim and post-stim responses were statistically different using paired t test (P < 0.05, one-sided paired sample t test). Visually responsive neurons exhibited either suppressed or enhanced responses by stimulus presentation.
We calculated the fraction of time spent viewing the novel image over the total viewing time (of both images) during tests as an index to evaluate recognition performance. We used the same method as described before (19) to identify neurons correlated with recognition performance. For all trials in a session (in test one), the firing rate modulation index of individual neurons was calculated as the difference between the mean firing rates when viewing novel and repeat images divided by their sum. This index was then correlated with the animal’s recognition performance (viewing novel fraction). We excluded trials in which the animals exclusively viewed either the novel or repeat images. Neurons with a significant correlation coefficient (Pearson correlation, P < 0.05) were identified as selective to recognition memory.
Spatial behavioral variables were extracted from the raw marker data that were preprocessed to fill small gaps (<1 s). All data were resampled to 50 Hz. Combining eye and head tracking data, we extracted eight spatial variables, including position (the animal’s horizontal self-location), spatial view (where the animal looks), egocentric center (the animal’s distance and direction relative to the arena center), head tilt (pitch and roll angles), head angular velocity (yaw and pitch rotational velocity), head height (vertical position of the head), translational speed, and horizontal head direction. The variables were binned to fit in the linear-nonlinear Poisson (LNP) model. Detailed descriptions of the variable binning are given
Position (Pos), the monkey’s location in the arena’s horizontal plane (XY plane). A 2D variable; X and Y: 0 to 350 cm, 20 by 20 bins;
Spatial view (SV) is the intersection of the monkey’s fixation point and the interior surfaces of the arena. X, Y, and Z: 0 to 350 cm, 0 to 350 cm, and 0 to 220 cm, 15 by 15 by 9 bins;
Egocentric center (EC) was first defined in a polar coordinate system, with the radial component corresponding to the distance from the head to the arena center and the polar angle defined between the arena center-to-head vector and the azimuth head direction. This polar coordinate was then converted to Cartesian coordinate. 2D variable; X and Y: −240 to 240 cm, 20 by 20 bins;
Head tilt (HT) referred to the head gesture which only considered pitch and roll angles in this study. 2D variable; −30° to 30°(up-down), −60° to 60°(left-right), 18 by 9 bins;
Angular velocity (AV), the velocity components about the head’s yaw, pitch, and roll axes. We only considered the yaw and pitch angular velocities since the head rotated mostly in these two dimensions. 2D variable; yaw, −120° to 120°/s, 16 by 16 bins;
Head height (HH), the head position in the vertical z axis. Position and head height determined the 3D position of the monkey’s head. 1D variable, 10 to 100 cm, 15 bins;
Translational speed (Spd), linear movement speed of the head. 1D variable, 0 to 150 cm/s, 15 bins;
Head direction (HD) was defined in the horizontal plane. 1D circular variable, −180° to 180°, 15 bins.
We calculated spatial view in the following First, we performed a random fixation task using MonkeyLogic in a room before the navigation task, in which head-fixed monkeys gazed at some randomly presented points shown on the screen. We constructed a transformation matrix between the eye movement data recorded in MonkeyLogic and the physical eye-in-head angle for each eye. Second, during the free navigation task, we used MonkeyLogic to record eye movements through the wireless binocular eye tracker. We transferred the eye movement data to the eye-in-head angle using the transformation matrix calculated during calibration. Combined with 3D head position and orientation captured by the motion tracking system, we obtained a view vector for each eye. Third, to calculate the spatial view, we only considered the left eye’s view vector and calculated the intersection of this view vector and the surface of the arena. In general, we found a good match between the two eyes.
Traditional methods of analyzing spatial tuning do not consider the interactions between behavioral variables. In this study, as described elsewhere (14, 28), we used a multimodal framework, the LNP model, to analyze which spatial variables neurons encoded. The spike trains were divided into time bins of 0.02 s, which matched the binning of the behavioral variables. To assess the significance of the model fitting and reduce the risk of overfitting, we applied a fivefold cross-validation procedure. To minimize fitting bias, the entire session was divided into three chunks. Within each chunk, the data were further divided into five subchunks. During the cross-validation process, for each fold, the ith [where i ∈ (1, 2, 3, 4, and 5), representing the five folds] subchunk in each chunk was combined and used as the test set, which accounted for 20% of the data. The remaining subchunks were used as the training set, which accounted for 80% of the data. The recorded spike train r was convolved using a Gaussian kernel with an SD of 0.06 s (equivalent to three time bins). The optimal parameter was determined using the MATLAB fminunc function. We used LLH as a metric to measure the performance of the model, namely, its goodness of fit. The performance of the model was compared to a null model, namely, the mean firing rate model, to determine significance. This was done using a one-sided Wilcoxon signed rank test with a significance level (alpha) of 0.05. We used an optimized forward search approach to identify the best model. First, one-variable models were fitted. The best, significant one-variable model (if there is any) was selected for further fitting by adding another variable one at a time. Then, the best, significant two-variable model (if any) was selected for further fitting. This procedure continued until adding more variables did not further improve model performance. The fraction of neurons encoding each variable was based on the final best model. That is, if a neuron encoded more than one variable, then the neuron counts at each encoded variable were all increased by one.
The degree to which variables were encoded conjunctively was defined as the similarity of the encoding between variables, that is, the size of the intersection divided by the size of the union—Jaccard index. The circular graph (Fig. 3D) (https://github.com/paul-kassebaum-mathworks/circularGraph) included only variable pairs with a Jaccard index larger than 0.1.
The free navigation task and the VPC task were interleaved by the eye calibration procedure of the binocular wireless eye tracker. We did not perform continuous recording across the two tasks and the intermediate procedure. Spike sorting was performed separately between the two tasks. The identical neurons were identified if they showed identical average waveforms on the same channel between the two tasks.
To assess whether the proportion of neurons showing both recognition and navigation selectivity was at chance level, we randomly drew the corresponding numbers of memory neurons and space neurons from the population and calculated the overlap. We repeated this process 1000 times and obtained a distribution. We then compared the actual overlap with the 95th percentile of the distribution. We found that the actual overlap was below chance for both monkeys.
We thank Y. Huang and A. Liu for help with data preprocessing, Y. Zhang for help with animal training, and the monkey facility for veterinary care.
Funding: This work was supported by National Science and Technology Innovation 2030 Major Program 2022ZD0205000 (D.M.), National Natural Science Foundation of China 32371076 (D.M.), and The Lingang Lab LG202105-01-08 (D.M.).
**Author ** Conceptualization: D.M. Methodology: X.X. and D.M. Investigation: X.X. and K.D. Duration: X.X., K.D., and D.M. Formal X.X., K.D., and D.M. Visualization: X.X., K.D., and D.M. Supervision: D.M. Writing—original X.X. and D.M. Writing—review and X.X. and DM.
**Competing ** The authors declare that they have no competing interests.
**Data and materials ** All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials.