Parallel maturation of hippocampal memory and CA1 task representations

Hippocampal-dependent memory is known to emerge late in ontogeny and its full development is protracted. Yet, the changes in hippocampal neuronal function that underlie this delayed and gradual maturation remain relatively unexplored. To address this gap, we recorded ensembles of CA1 neurons while charting the development of hippocampal-dependent spatial working memory (WM) in rat pups (∼2-4weeks of age). We found a sharp transition in WM development, with age of inflection varying considerably between individual animals. In parallel with the sudden emergence of WM, hippocampal spatial representations became abruptly task specific, remapping between encoding and retrieval phases of the task. Further, we show how the development of task phase remapping could partly be explained by changes in place field size during this developmental period as well as the onset of precise temporal coordination of CA1 excitatory input. Together, these results suggest that a hallmark of hippocampal memory development may be the emergence of contextually specific CA1 representations driven by the maturation of CA1 micro-circuits.

Methods).Experiments started between P17 and P24 and con�nued every day for approximately one week.Between P17 and P28 we observed a gradual improvement in the pups' average performance on the task (r = 0.97, p < 0.0001), reaching eventually performance comparable to adults (t(118) = 0.42, p = 0.68).This is in agreement with previous work that showed pups can reliably carry out this task, and other hippocampal-dependent tasks, at around 3weeks of age 8,9,19 .However, previous studies have also highlighted that the ability to carry out tasks similar to the DNMP emerges abruptly in individual animals 20 .To address this ques�on, we fited sigmoid curves to individual pups' developmental curves (Figure 1C, Online Methods).We found sigmoid fits captured the developmental curves significantly beter than linear fits (AIC sigmoid fits = -37.98(SD=6.88),linear fits = -34.26(SD=6.27),t(12) = -2.78,p = 0.0.017, Figure S1C), sugges�ng the developmental emergence of this form of hippocampal memory occurs abruptly (i.e.overnight).Further, the �ming of inflec�on points varied notably between animals (Figure 1C,D), with the earliest inflec�on point observed at P19 and the latest at P24. Importantly, this abrupt improvement in performance could not simply be explained by experience, as animals that started experiments in the post-weaning period (>P21) already performed above chance at day1 and showed a significantly shallower sigmoid fit (2-sample Kolmogorov-Smirnov test, p = 0.028, Figure 1E, Figure S1A).Further, we did not observe a difference in development curves for pups who were weaned after experiments concluded as opposed to at P21 (t(11) = 0.18, p = 0.86, Figure S1B).Is this abrupt behavioural development supported by changes in neuronal ac�vity?In the first instance, we assessed how CA1 spa�al encoding changed during this developmental period and how, or if at all, it relates to the development of the animal's ability to carry out the task.To this end, we correlated ratemaps for the two arms of the T-maze (Figure 2A(i), Online Methods).Spa�al correla�ons between ratemaps of the two arms were generally low (mean = 0.07 (SD=0.22), Figure 2A(ii)), sugges�ng place cells differen�ated between the two geometrically iden�cal arms.To ensure the low spa�al correla�ons could not merely reflect unstable place coding, we correlated ratemaps for odd and even runs on the same arms (Online Methods) and compared the distribu�on of these stability scores against the le� vs right arm remapping scores (Figure 2A(ii)).We found the remapping scores were significantly lower than the stability scores (mean = 0.30 (SD=0.15),p < 0.0001, 2-sample Kolmogorov-Smirnov test), indica�ng the low correla�ons between the le� and right arm ratemaps indeed reflect remapping.We next asked if spa�al remapping changed in tandem with the developmental improvements observed for individual animals on the DNMP task.To this end, we correlated the average session spa�al remapping scores against days to/from inflec�on (0 indicates first day a�er inflec�on).Spa�al correla�ons scores did not correlate with the development of DNMP memory (r = -0.09,p = 0.47, Figure 2A(iii)).Indeed, at the earliest ages tested CA1 cells showed reliable remapping between the two arms (r = 0.052 (SD=0.18),t (5) = -3.05,p = 0.029, 1-sample t-test P17-P18 remapping vs stability), consistent with previous research 12 , and the remapping observed did not differ from adult spa�al remapping (r = 0.18 (SD = 0.20), t(86) = -1.44,p = 0.16, Figure 2A(iii)).
As noted previously, studies in adult rodents have shown significant trajectory (le�-vs right-bound trials) and task phase (sample vs choice trials) remapping in the DNMP task.Next, we sought to chart their ontogeny.Star�ng with trajectory remapping, we found no evidence for this form of remapping in our data (Figure 2C(i)); the average correla�on between right-and le�-bound stem ratemaps did not differ from the stability distribu�on (p = 0.23, 2-sample Kolmogorov-Smirnov test, Figure 2C(ii)).Further, the degree of trajectory remapping did not show any rela�onship with the DNMP developmental curves (r = -0.001,p = 0.99, Figure 2C(iii)).On the other hand, we found the CA1 cells remapped strongly between the sample and choice trials (p = 0.0014, Figure 2D(i-ii)), and the degree of task-phase remapping correlated significantly with days to/from inflec�on point (r = -0.5, p < 0.0001), becoming comparable to adult task phase remapping in the post-inflec�on period (t(58) = 0.15, p = 0.88).To note, we also correlated the different measures of remapping to post-natal age (Figure S2A-C).These correla�ons were notably weaker than those observed for inflec�on point (trajectory remapping: r = -0.13,p = 0.25; task phase remapping: r = -0.22,p = 0.04), highligh�ng the need to account for individual varia�on when studying the neuronal basis of cogni�ve development.
Importantly, the rela�onship between task-phase remapping and DNMP developmental curves could not be explained by experience.A par�al correla�on controlling for the effect of experience had only a small effect on the inflec�on point vs task-phase remapping correla�on, and it remained robustly significant (r = -0.47,p < 0.0001).Indeed, using a general linear model to predict task-phase remapping from experience, post-natal age and days to/from inflec�on (Online Methods) showed that only days to/from inflec�on could significantly predict developmental changes in task-phase remapping (GLM: inflec�on point: t(71) = -4.56,p < 0.0001; post-natal day: t(71) = 0.94, p = 0.35; experience: t(71) = 1.2, p = 0.24).Finally, we found movement speed did not differ between the two trial phases (choice = 21.24cm/sec(SD=10.87),forced = 19.01cm/sec(SD=10.88),t(172) = 1.35, p = 0.18), although median speed increase significantly with age (r = 0.52, p < 0.0001).Importantly, task phase remapping remained a significant predictor of days to/from inflec�on a�er controlling for age-related changes in speed (r = -0.38,p = 0.0011).
To corroborate these findings and to explore where remapping occurred on the maze, we turned to a popula�on vector analysis (Online Methods).For each spa�al bin (4cm) we correlated the popula�on vectors between sample and choice ratemaps for each session, and then computed the average correla�on across all sessions recorded during the pre-and post-inflec�on period.In agreement with the remapping analysis described above, we found popula�on vector correla�ons were significantly lower during the post inflec�on period compared to the pre inflec�on period (Figure 2D).To note, dividing the post-inflec�on popula�on vector correla�ons in two (peri: inflec�on points 0-2; post: >2 days post inflec�on) showed no further changes in remapping (all spa�al bins p > 0.05, Figure S3).This suggests task-phase remapping emerges abruptly in development, mirroring the abrupt development of hippocampal memory.Black line underneath PV indicates bin where there is a significant difference between pre and post inflec�on periods.(E) Distribu�on of task phase remapping scores for pre (light purple) and post (dark purple) inflec�on periods as well as adults (grey).Circles on the x-axis show the centre of fited gaussian components.Inset: Centre of gaussian components for pre and post inflec�on periods and for adults, y-axis shows the task phase ratemap correla�on that corresponds to the centre of individual components.The size of the circle shows the propor�on of data that is captured by individual components.(F) Propor�on of cells remapping between choice and sample trials across inflec�on points.
Next, we sought to characterize whether developmental changes in remapping reflect homogeneous changes in place cell task phase coding across the CA1 popula�on.To this end, we fited gaussian distribu�ons to the PRE and POST task phase remapping distribu�ons (Online Methods).During the pre-inflec�on period, we found that the task phase remapping distribu�on was best fited with three Gaussian components (AIC = 432.35),and the three components were centred on r = -0.05(66%), r = 0.52 (30%) and r = 0.89 (3.6%) correla�on scores (Figure 2E).This suggests heterogeneous task phase encoding during the pre-inflec�on period, with some cells remapping while others did not.During the post-inflec�on period, however, the task phase correla�on distribu�on could be captured by only two components (AIC = 230.97),one centred on r = -0.13(30%) and another on r = 0.20 (70%, Figure 2E).Thus, the high correla�on component (r = 0.89) observed during the pre-inflec�on period disappeared post inflec�on, and only two sub-popula�ons -both centred on low correla�on scores -of task phase coding were apparent in the popula�on, sugges�ng task phase remapping had become nearly ubiquitous.In agreement with this, we found days to/from inflec�on could be reliably predicted by the propor�on of cells that showed task phase remapping (r = 0.42, p = 0.0001, Figure 2F).To note, fi�ng gaussian distribu�ons to adult task phase remapping distribu�ons (Figure 2E), also revealed two components (AIC = 47.75)centred on similarly low correla�on coefficients of r= -0.05 (71%) and r = 0.26 (29%) as during post-inflec�on.This underscores the observa�ons that task phase coding emerges abruptly in development.
These findings led us to ask what could explain the sudden developmental emergence of task phase remapping?We first explored whether days to/from inflec�on could be predicted by changes in place cell rate, place field size or place cell sparsity (Online Methods).Days to/from inflec�on showed no rela�onship to place cell ac�vity rate (peak rate: sample: r = 0.19, p = 0.1, choice r = 0.09, p = 0.42, Figure S4A) nor to sparsity (% of cells ac�ve: sample = r = -0.21,p = 0.06, choice: r = 0.06, p = 0.62, Figure S4B).However, we observed a significant correla�on between the size of place cell's place field and days to/from inflec�on (sample r = -0.41,p = 0.0002; choice r = -0.23,p = 0.042; Figure S4C), sugges�ng spa�al coding becomes more precise as cogni�ve development unfolds.Yet, controlling for developmental changes in place field size could not fully account for the correla�on we observed between days to/from inflec�on and task phase remapping (par�al correla�on r = -0.41,p < 0.001); sugges�ng the emergence of task phase remapping may only par�ally be explained by changes in place field size in development.
An alterna�ve hypothesis is that task phase remapping may reflect matura�on in the temporal coordina�on of input to CA1.CA1 receives primarily two glutamatergic inputs -one from layer three of the entorhinal cortex (ECIII) and another from area CA3.These two inputs are thought to reflect dis�nct hippocampal network states, suppor�ng complementary processes.ECIII input has been purported to support encoding-related processes while CA3 input may rather support memory-guided processes 21,22 .As the DNMP task requires animals to learn to alternate between encoding and memory-guided phases, perhaps the ontogene�c emergence of task-phase remapping, and thereby hippocampal memory matura�on, reflects developmental changes in CA1 input alignment to dis�nct task phases.Namely, with development CA3 input may become more dominant and specific to the choice phase -which requires an execu�on of memory-guided ac�ons -while ECIII input may be preferen�ally dominant during encoding-driven sample phases.
To address this ques�on, we analysed CA1 single-unit ac�vity and LFP markers that provide a proxy for CA3/ECIII input balance in CA1.In the first instance, we analysed theta phase preferences of place cell spikes during the sampling and choice phases.As ECIII input arrives earlier in a theta cycle rela�ve to CA3 input 23 , we reasoned that during the post-inflec�on period, sample phases should be associated with earlier theta phase spiking compared to choice phases but that no difference in phase preference should be observed in the pre-inflec�on period.Consistent with this, we found place cells tended to fire near the trough of locally recorded theta-band oscilla�ons during the pre-inflec�on period (sample mean angle = 131.5°,choice mean angle = 110.4°, Figure 3A(i)), and the phase preference between sample and choice phases of the task did not differ (95% bootstrap confidence intervals = [-0.85,0.10]).During the post-inflec�on period, however, we found phase preferences between the two task phases differed reliably (sample = 111.4°,choice = 154.5°, Figure 3A(ii)), with place cells firing at significantly earlier phases of a theta cycle during sample phases of the task rela�ve to choice task phases (95% bootstrap confidence interval = [0.01,1.42]).Addi�onally, theta phase locking -thought to derive from CA3 input -decreased significantly as the animals started to be able to perform the task reliably (r = -0.67,p < 0.001, Figure 3B).Importantly, this effect was driven by the sampling trial phases (p = 0.006, 2-sample Kolmogorov-Smirnov test, Figure 3C).Phase locking during choice trial phases did not change significantly between pre-and post-inflec�on periods (p = 0.82, 2-sample Kolmogorov-Smirnov test, Figure 3D).
Finally, to corroborate these findings we inves�gated coupling of slow and medium gamma oscilla�ons to CA1-recorded theta-band oscilla�ons in sampling and choice trials and assessed how theta-gamma coupling changed between the pre-and post-inflec�on periods.Slow gamma (~20-40Hz), recorded in CA1, is known to derive from upstream CA3 ac�vity while medium gamma oscilla�ons (~50-70Hz) are thought to originate from ECIII 24 .Thus, measuring the rela�ve coupling of the two gamma-band oscilla�ons to CA1 theta-band oscilla�ons is another approach to assess the rela�ve influence of ECIII and CA3 input over CA1 ac�vity 24,25 .During the pre-inflec�on period, we found the ra�o between slow and medium theta coupling to be comparable during sample and choice phases of the task (sample slow/medium gamma ra�o = 0.99 (SD=0.20),choice = 10.2 (SD=1.02),p = 0.48 based on bootstrapped ra�o scores, Figure 3E,F(i)).During the post-inflec�on period, however, we found a robust difference in the rela�ve theta coupling of the two gamma bands between the two task phases (sample = 0.92 (SD=0.12),choice = 1.61 (SD=0.34),p = 0.0066 based on bootstrapped ra�o scores, Figure 3E,F(ii)), with the ra�o shi�ed towards medium gamma during sample phases but towards slow gamma during choice phases (Figure 3F).To note, despite changes in movement speed between the two developmental epochs, these changes could not explain the shi� in slow-to-medium gamma balance between the sampling and choice trials post inflec�on (Figure S5, Online Methods).In sum, consistent with the theta phase locking and phase preference analyses described above, it seems that the two excitatory inputs to CA1 become beter aligned to the dis�nct task phases as the animals' hippocampal memory develops.Here, we show that hippocampal-dependent memory and CA1 task phase remapping emerge in parallel in ontogeny.As the ability to carry out a spa�al working memory matured so did CA1 neuronal representa�ons for specific task phases.These findings suggest that the development of hippocampal memory may be underpinned by the development of func�onally specific CA1 neural representa�ons.Contemporary theories of hippocampal memory development propose that one of the hallmarks of memory matura�on is that memory becomes less generic and more specific 26 .This study provides, for the first �me, insight into the changes in neuronal coding that may underlie this cri�cal cogni�ve developmental milestone.
Although the origin of this neuro-developmental milestone remains to be ascertained, we propose it may reflect the combined emergence of precise CA1 place fields and adult-like temporal organiza�on of CA1 glutamatergic input.On the one hand, as place fields become smaller, the ability to dis�nguish between related representa�ons may increase.Alterna�vely, if with development the two phases of the task become associated with dis�nct excitatory inputs, this would naturally lead to different fields emerging for the dis�nct phases.A more temp�ng hypothesis is that these phenomena are in fact inter-dependent: precise, informa�on-rich, spa�al representa�ons might depend on effec�ve integra�on of separate input streams in CA1, poten�ally relying on their temporal organiza�on.The structured interac�on of sensory-based and memory-based informa�on in CA1 could be fostered by a shi� from compe�ng ECIII and CA3 inputs early in life to their spa�o-temporal segrega�on in adults.In turn, such arrangement could provide the framework for the emergence of high-dimensional representa�ons, suppor�ng adult-like memory capabili�es.Finally, it remains to be seen what role the medial prefrontal cortex (mPFC) plays in rela�on to this process.The mPFC, and par�cularly mPFChippocampal interac�ons, are known to be crucial for accurate DNMP performance in adults and the mPFC develops significantly during this period 27 .Perhaps the ontogeny of specific CA1 representa�ons, and mature circuit func�on, reflects the matura�on of hippocampal-mPFC pathways.

Experimental model and subject details
Thirteen Lister Hooded rat pups (30-52g at implanta�on) underwent a surgical procedure to implant a microdrive carrying eight or sixteen tetrodes of twisted 17 μm HM-L coated 90% pla�num, 10% iridium wire (California Fine Wire), targe�ng the right CA1 (ML: 2.0-2.3mm,AP: 3.0mm posterior to bregma).Electrode �ps were gold plated to reduce impedance to 150-300kOhm at 1kHz.Pups were allowed to recover from surgery housed together with litermates (and dam if pre-weaning) for 48h.The animals always had ad libitum access to water and food, and were housed on a reversed 12-h light-dark cycle.Eleven adult Lister Hooded rats (330-400g at implanta�on) underwent the same procedure as pups.Adult procedures differed from pup procedures in CA1 target coordinates (ML: 2.0-2.3mm,AP: 3.8mm posterior to bregma), one week of post-opera�ve recovery, and they were housed individually.

Electrophysiological recording
A�er the post-opera�ve recovery period, electrophysiological ac�vity was screened two to three �mes a day.All recordings were performed using an Axona recording system (Axona Ltd., St. Albans, UK) which recorded spike-threshold triggered single unit ac�vity, con�nuous LFP, and posi�on data.Each channel was amplified 5000 -15000 �mes and recorded referenced to another channel on a separate tetrode.Spikes and LFP were sampled at 48KHz.Animal posi�on was determined using an overhead infrared (IR) camera recording the loca�on of an array of IR light-emi�ng diodes (LED) mounted on the headstage.Tetrodes were gradually advanced ventrally in 62.5 -125μm steps, at least 3h apart, un�l place cells and sharp-wave ripples were detected.

Experimental apparatus and procedures
All experiments were performed in a dark room with no natural sources of light.The experimental area of the maze was surrounded by thick opaque black curtains on all sides.The experimental arena consists of a dark brown wooden digit-8 maze (140x140cm) with textured wood running surface.A Tshaped por�on of the maze was used for the DTT task, the remaining parts of the maze were blocked off using black metal barriers.A black plas�c food well holding 0.1ml of liquid was placed at the end of each arm of the T-maze.In between trials the animals were placed in an inter-trial-interval (ITI) box (12x12cm) at the start of the stem of the Tmaze.Access to the stem was blocked off with a tall removeable barrier.Sleep recordings were conducted in a tall round opaque black plas�c box (25x48cm), filled with bedding sand, that was placed on top of the centre of the stem of the maze, directly underneath the IR camera.
Two days prior to experiments, animals were habituated for a single 10-15 minute session per day.Habitua�on was carried out on the tracks of the digit-8 maze that were unused in the T-maze, such that the physical characteris�cs of the tracks are iden�cal to the T-maze tracks, but no part of tracks overlapped between habitua�on and T-maze.During habitua�on, animals were allowed to self-ini�ate walking between two ends of a linear track with food wells filled with soy milk at each end.The experimenter refilled the food wells with milk as necessary.
The task was performed in two iden�cal sessions per day, at least 3 hours apart during which �me the animals rested in the homecage.The animals were not food nor water deprived, and the condi�oning reward was 0.1ml of soy milk formula.Each trial-pair consisted of two runs -SAMPLE and CHOICE.Each SAMPLE trial had access to one of the arms blocked off with a removable barrier, and food was placed in the open arm.The selec�on of le�/right open arms during SAMPLE trials was pseudorandomised.When the animal reached the end of the open arm and drank its food reward it placed I the ITI box for 15 seconds.The CHOICE trial started a�er the ITI door was li�ed for the animal to walk out into the stem.In this run, both arms were open.The pup got rewarded if they chose the arm opposite to the one rewarded during the SAMPLE trial.A�er each trial-pair, the animal was placed into the holding box outside the maze for 30-45s intra-trial interval.Sessions lasted between 15-45 minutes.Immediately a�er finishing each session, provided sufficient single-unit yield, the animals were placed into the sleep box and allowed to rest for an hour.

Inflec�on point analysis
To determine developmental inflec�on points on the DNMP task for individual animals, we fited sigmoid curves to individual animals' daily performance data.Specifically, we fit a sigmoid curve to the data using the standard logis�c func�on (Equa�on 1) where d denotes the chance level performance which is set to 0.5, a represents the animal's highest daily performance mean, c is the steepness of the curve ranging from 0 to 20, and b the inflec�on point of the sigmoid constrained between the animal's first and last post-natal day.We used the nonlinear least squares fit op�on in Matlab R2019b (Mathworks, MA) in order to fit the model to the data.Any animals with fewer than 5 daily means (n = 3 animals) were excluded from the inflec�on point analysis to ensure a reliable fit.
In cases where the inflec�on point given by the fited model falls between two whole numbers, we always rounded up the value.Therefore, the next PD a�er the inflec�on point is determined to be day 0. For PRE vs POST analyses, we divided the data into sessions prior to the inflec�on point (PRE), and sessions from day 0 onwards (POST).For the analysis reported in Figure S4 where we further divided the POST period, PERI was defined as the data from day 0 to day 2, and POST as data form day 3 onwards.
To compare linear and sigmoid fits we computed the Akaike informa�on criterion for linear and sigmoid fits and compared them using a paired samples t-test.

Place cell analysis
All analyses were restricted to puta�ve principal cells, iden�fied by manual inspec�on of waveforms across the en�re recording session.KlustKwik was applied to spike-thresholded data to sort the data into clusters and then the clusters were manually curated in Tint (Axona Ltd.).We classified spike sorted neurons as place cells by compu�ng Skaggs Informa�on (bits per second), and compared the value we obtained against a null distribu�on generated by random permuta�ons of spike �mes.Cells that exceeded 95 th percen�le of their own shuffle were deemed place cells.Sessions containing fewer than five cells with significant Skaggs informa�on values were excluded from analysis.
To generate ratemaps, spike data was divided into the four different trial types (e.g.SAMPLE le�, CHOICE le�, etc).Spikes that occurred during sta�onary periods (<3cm/s) or when the animal was located near the start of the stem (<10cm) were excluded.Next, we linearised animals' paths, binned dwell �me and total number of spikes in 2cm bins, computed firing rates by dividing the binned spikes over binned dwell �me, and smoothed them using a Gaussian kernel (sigma= 3 bins).

Remapping
To analyse remapping between the different trial types we correlated the spa�al ratemaps for pairs of runs using the Pearson correla�on coefficient (empty bins in both ratemaps were removed).Spa�al remapping was performed only on the bins corresponding to the arms of the T-maze.The rest of remapping analyses excluded the arms and were only performed on the bins corresponding to the central stem of the T-maze, which is the only sec�on of the track common to all four run types.Correla�ons between le�-and right-bound ratemaps were refer to as trajectory remapping, while correla�ons between SAMPLE and CHOICE ratemaps we term task phase remapping.To ensure remapping scores did not just reflect unstable spa�al firing, we compared the remapping scores to correla�ons scores obtained by correla�ng the ratemaps for odd and even runs for a given trial type.To assess if a par�cular type of remapping was present in the popula�on we compared the distribu�on of mean session remapping scores to the stability scores using a 2-sample Kolmogorov-Smirnov test.
To assess if remapping changed with development we correlated average remapping scores obtained in a session with post-natal age/inflec�on point using a Pearson correla�on coefficient.To rule out the effect of experience and changes in median speed during development, we used a par�al correla�on where the rela�onship between inflec�on point and remapping was computed while controlling for the effect of experience/median speed.To compare the rela�ve influence of inflec�on point, postnatal age and experience we used a General Linear Model (GLM) with remapping scores as the response variable and inflec�on point, post-natal age and experience as the predictors.
To assess if the propor�on of cells remapping between task phases in a session correlated with inflec�on point we calculated the propor�on of cells with task phase remapping scores above 0.25 and correla�on these session propor�ons with inflec�on point.

Popula�on Vector Correla�on Analysis
To examine the distribu�on of task phase remapping remapping across the track, we correlated the popula�on vectors for choice and sample trials.Specifically, for each spa�al bin (4cm) we computed the Pearson correla�on coefficient for z-scored ac�vity of all cells ac�ve in a session.The first 10cm at the start of the stem were removed to exclude areas of the maze associated with immobility.The average popula�on vector was then computed for PRE and POST inflec�on periods and 95% confidence intervals (CIs) computed to assess which spa�al bins differed significantly between the two periods.The average PRE and POST popula�on vector correla�on was then smoothed with a guassian kernel (sigma = 8cm).To compute the CIs we bootstrapped the session popula�on vector correla�ons 10,000 �mes, repea�ng the analysis separately for PRE and POST inflec�on periods, for each itera�on of the bootstrap we computed the mean popula�on vector correla�on.From the bootstrapped data we obtained the 2.5 th and 97.th percen�le for each inflec�on period, if the CIs for the two periods did not overlap we deemed the comparison significant.The same procedure was used to compared the popula�on vector correla�ons between the period immediately a�er inflec�on (inflec�on points 0-2, PERI) and the subsequent days.

Fi�ng gaussian components
To fit gaussian components to the distribu�on of task phase remapping scores during the PRE and POST inflec�on periods we used the Matlab func�on fitgmdist, we fited 1 and 4 components and used the Akaike Informa�on Criterion (AIC) to determine the model with the best fit.

Place field analyses
Field size and peak rate were assessed by first using the regionprops func�on in Matlab on rate thresholded ratemaps (bins > 50% of the peak rate of the ratemap).Small fields detected with this method (<20cm long) were removed.Then the area of the field was used as a measure of field size and the highest rate within the field a measure of peak rate.If a cell had mul�ple fields the average size and peak was computed.To measure sparsity we computed the propor�on of all cells recorded there were had a significant skaggs informa�on value in all four run types.
To assess how these place cell features changed with development we correlated them against postnatal age/inflec�on point using the Pearson correla�on coefficient.To control for the effect of field size we used a par�al correla�on, where inflec�on point and task phase remapping were the predictor and response variables, respec�vely, and field size the covariate.

Theta phase analyses
To analyse theta phase preference and phase locking to theta-band oscilla�ons during SAMPLE and CHOICE trials we first iden�fied the electrode in the CA1 region with the highest power in the theta band (5-12Hz) using LFP data downsampled to 1.2kHz.We performed a wavelet transform to extract the instantaneous phase of the channel's signal in the theta band.We then used the extracted phase to iden�fy the theta phase of each spike.We filtered out low theta periods where the power of theta was below the mean.Further, we excluded sta�onary period (<3cm/sec) and only analysed the phase for cells who s�ll had at least 10 spikes within their place field for given run a�er this filtering.
To compute the phase preference of each cells during SAMPLE and CHOICE runs we calculated the circular mean for each trial type.To assess phase locking we computed the Resultant Vector Length for each cell's phases.To compare phase preference between SAMPLE and CHOICE runs for the two inflec�on periods we first boostrapped the preferred phase distribu�ons for each run type 10,000 �mes.For each itera�on of the bootstrap we computed the circular mean of the SAMPLE and CHOICE data.Subsequently, we computed the circular distance between SAMPLE and CHOICE bootstrapped mean phase distribu�ons, and computed confidence intervals based on the returned circular distance distribu�ons.If the CI did not include 0 we concluded that the different between the SAMPLE and CHOICE phases different significantly.To note, analyses were done separately for PRE and POST inflec�on periods.For phase locking data, we first assessed if phase locking changed with inflec�on point.To this end, we used a Pearson correla�on coefficient between the session mean phase locking scores (for all trial types) and inflec�on points.We then divided the phase locking data by run type (SAMPLE and CHOICE) and compared the distribu�on of phase locking scores for each trial type during PRE and POST inflec�on periods, using a 2-sample Kolmogorov-Smirnov test.

Theta-Gamma Coupling
To measure coupling of slow and medium gamma oscilla�ons to theta-band oscilla�ons during SAMPLE and CHOICE task phases, we filtered the LFP data so to include only samples where theta power was above the mean and divided the data into SAMPLE and CHOICE periods.We computed the phase-amplitude coupling between theta phase and amplitude of oscillatory components between 15 and 200 Hz: we first extracted the phase of the oscillatory component in the theta band (5-12 Hz).Then we binned theta cycles into 26 phase bins and for each phase bin we computed the average power of the oscillatory components at higher frequencies (>15Hz) obtained from a wavelet decomposi�on of the full signal.Thus, we obtained a phase-amplitude 2D-matrix of coupling strengths where each value correspond to a par�cular combina�on of theta phase interval and degree of amplitude modula�on of a faster oscilla�on.These couplings were then z-scored across each frequency band and average coupling computed for PRE and POST inflec�on periods and SAMPLE and CHOICE phases.To assess the ra�o between slow and medium gamma coupling to theta, we iden�fied the highest coupling observed in the phase-amplitude analysis in the two gamma bands (slow gamma: 18-30Hz, medium gamma: 40-70Hz), and divided the highest coupling observed in the slow gamma band by the highest coupling observed in medium band.Here a value above 1 indicates strongest coupling to slow gamma compared to medium gamma.To assess if slow-tomedium gamma coupling to theta-band oscilla�ons differed between SAMPLE and CHOICE trials during the two inflec�on periods, we bootstrapped the session cross-frequency coupling spectrograms 10,000 �mes, and for each itera�on of the bootstrap we computed the difference in slow-to-medium gamma coupling.We then computed 95% confidence intervals for the difference score distribu�on.
To control for the effect of speed during different inflec�on periods, we repeated the analysis above for different speed bands (low: 3-20cm/s, mid: 20-40cm/s, high: >40cm/s), we then assessed if the difference in slow-to-medium gamma ra�os during each inflec�on period for individual speed bands differed significantly from the original data.To this end, we computed the difference between bootstrapped slow-to-medium gamma ra�os during individual inflec�on periods as described above.

Figure 1 .
Figure 1.Hippocampal-dependent memory develops abruptly.(A) Schema�c of the DNMP task.(B) Mean percentage of correct alterna�ons of pups across age.Shaded area shows SEM.(C) Fited

Figure 2 .
Figure 2. Task phase remapping ontogeny predicts matura�on of hippocampal memory.(A) (i) Spa�al remapping was computed by correla�ng place cell ratemaps for le� and right arms.(ii) Distribu�on of average session remapping scores (blue) and stability scores (grey).(iii) Mean session remapping scores as a func�on of days to/from inflec�on point.Horizontal grey line shows mean stability score along with standard devia�on (dashed line).(B,C) Same as A but for trajectory and task phase remapping, respec�vely.(D) Popula�on vector correla�on between sample and choice ratemaps for pre and post inflec�on periods.Shaded area shows 95% confidence intervals from bootstrapping data.Black line underneath PV indicates bin where there is a significant difference between pre and post

Figure 3 .
Figure 3. Hippocampal memory matura�on is associated with task phase-based coordina�on of CA1 input.(A) (i) Circular histogram of CA1 cell phase preferences for sample and choice trials during the pre inflec�on period.Coloured diagonal lines show the circular mean for each trial type.(ii) Same as (i) but for the post inflec�on period.(B) Mean session phase locking (Raleigh vector length) as a func�on of days to/from inflec�on.(C) Frequency distribu�on of phase locking scores for the pre and post inflec�on periods for sample trials.(D) Same as (C) but for choice trials.(E) (i) Cross frequency coupling to theta-band oscilla�ons during sample (top) and choice (botom) trials, pre inflec�on, x-axis shows phase in a theta cycle.Note, three theta cycles are shown for clarity.(ii) same as (i) but for the post inflec�on period.(F) Average sample and choice slow/medium gamma coupling ra�o for pre and post inflec�on periods, error bars show 1SD.