Information flows from hippocampus to auditory cortex during replay of verbal working memory items

The maintenance of items in working memory relies on a widespread network of cortical areas and hippocampus where synchronization between electrophysiological recordings reflects functional coupling. We investigated the direction of information flow between auditory cortex and hippocampus while participants heard and then mentally replayed strings of letters in working memory by activating their phonological loop. We recorded LFP from the hippocampus, reconstructed beamforming sources of scalp EEG, and - additionally in 3 participants – recorded from subdural cortical electrodes. When analyzing Granger causality, the information flow was from auditory cortex to hippocampus with a peak in the 4-8 Hz range while participants heard the letters. This flow was subsequently reversed during maintenance while participants maintained the letters in memory. The functional interaction between hippocampus and the cortex and the reversal of information flow provide a physiological basis for the encoding of memory items and their active replay during maintenance.


22
The maintenance of items in working memory relies on a widespread network 23 of cortical areas and hippocampus where synchronization between 24 electrophysiological recordings reflects functional coupling. 25 We investigated the direction of information flow between auditory cortex and 26 hippocampus while participants heard and then mentally replayed strings of letters in 27 working memory by activating their phonological loop. We recorded LFP from the 28 hippocampus, reconstructed beamforming sources of scalp EEG, and -additionally in 29 3 participants -recorded from subdural cortical electrodes. When analyzing Granger 30 causality, the information flow was from auditory cortex to hippocampus with a peak 31 in the 4-8 Hz range while participants heard the letters. This flow was subsequently 32 reversed during maintenance while participants maintained the letters in memory. 33 The functional interaction between hippocampus and the cortex and the reversal of 34 information flow provide a physiological basis for the encoding of memory items and 35 their active replay during maintenance. 36 1 Introduction 37 Working memory (WM) describes our capacity to represent sensory input for 38 prospective use [1,2]. Maintaining content in WM requires communication within a 39 widespread network of brain regions. The anatomical basis of WM was shown 40 noninvasively with EEG / MEG [3][4][5][6][7][8][9][10] and invasively with intracranial local field 41 potentials (LFP) [11][12][13][14][15][16][17][18][19][20][21] and single unit recordings [19,[21][22][23][24]. 42 In cortical brain regions, WM maintenance correlates with sustained neuronal 43 oscillations, most frequently reported in the theta-alpha range (4-12 Hz) [3][4][5][6][7][9][10][11][12][13][14][15][16][17][18][19][20] or 44 at even lower frequencies [25,26]. Also in the hippocampus, WM maintenance was 45 associated with sustained theta-alpha oscillations [15,19]. As a hallmark for WM 46 maintenance, persistent neuronal firing was reported during the absence of sensory 47 2 Results 66 2.1 Task and behavior 67 Fifteen participants (median age 29 y, range , 7 male, Table 1) 68 performed a modified Sternberg WM task (71 sessions in total, 50 trials each). In the 69 task, items were presented all at once rather than sequentially, thus separating the 70 encoding period from the maintenance period. In each trial, the participant was 71 instructed to memorize a set of 4, 6 or 8 letters presented for 2 s (encoding). The 72 number of letters was thus specific for the memory workload. The participants read 73 the letters themselves and heard them spoken at the same time. Since participants 74 had difficulties reading 8 letters within the 2 s encoding period, also hearing the 75 letters assured their good performance. After a delay (maintenance) period of 3 s, a 76 probe letter prompted the participant to retrieve their memory (retrieval) and to 77 indicate by button press ("IN" or "OUT") whether or not the probe letter was a 78 member of the letter set held in memory (Fig. 1a). During the maintenance period, 79 participants rehearsed the verbal representation of the letter strings subvocally, i.e. 80 mentally replayed the memory items. Participants had been instructed to employ this 81 strategy and they confirmed after the sessions that they had indeed employed this 82 strategy. This activation of the phonological loop [1] is a component of verbal WM as 83 it serves to produce an appropriate behavioral response [2]. 84 The mean correct response rate was 91% (both for IN and OUT trials). The rate of 85 correct responses decreased with set size from a set size of 4 (97% correct 86 responses) to set sizes of 6 (89%) and 8 (83%) (Fig. 1 b). Across the participants, 87 the memory capacity averaged 6.1 (Cowan's K, (correct IN rate + correct OUT rate -88 1)*set size), which indicates that the participants were able to maintain at least 6 89 letters in memory. The mean response time (RT) for correct trials (3045 trials) was 90 1.1 ± 0.5 s and increased with workload from set size 4 (1.1 ± 0.5 s) to 6 (1.2 ± 0.5 s) 91 and 8 (1.3 ± 0.6 s), 53 ms/item (Fig. 1 c). Correct IN/OUT decisions were made more 92 rapidly than incorrect decisions (1.1 ± 0.5 versus 1.3 ± 0.6 seconds). These data 93 show that the participants performed well in the task and that the difficulty of the trials 94 increased with the number of letters in the set. In further analysis, we focused on 95 correct trials with set size 6 and 8 letters to assure hippocampal activation and 96 hippocampo-cortical interaction as shown earlier [19]. 97

98
To investigate how cortical and hippocampal activity subserves WM 99 processing, we analyzed the LFP recorded in the hippocampus (Fig. 1 d) together 100 with ECoG from cortical strip electrodes (Fig. 2 a, Fig. 3 a, f). In the following, we 101 present power spectral density (PSD) time-frequency maps from representative 102 electrode contacts. In an occipital recording of Participant 1 (grid contact H3), strong 103 gamma activity (> 40 Hz) in the relative power spectral density (PSD) occurred while 104 the participant viewed the letters during encoding (increase >100 % with respect to 105 fixation, Fig. 2 b). Similarly, encoding elicited gamma activity in a temporal recording 106 over auditory cortex (increase >100%, grid contact C2, Fig. 2 c), similar as in [25]. 107 Gamma increased significantly only in temporal and occipital-parietal contacts 108 (permutation test with z-score > 1.96, Fig. 2 a). 109 After the letters disappeared from the screen, activity occurred in the low beta 110 range (11-14 Hz, Fig. 2 b) towards the end of the maintenance period in temporal 111 and occipital contacts (permutation test p < 0.05, Fig. 2 d). Similarly, the temporal 112 scalp EEG of Participant 2 (black rimmed disk denotes electrode site T3 in Fig 3 a) 113 showed activity during encoding and maintenance, albeit at lower frequencies (Fig 3  114 b); this pattern was found only in scalp EEG and not in ECoG, probably because the 115 strip electrode was not located over auditory cortex. In Participant 3, a similar pattern 116 occurred in the PSD of a temporo-parietal recording (most posterior strip electrode 117 contact, Fig 3 f), where the appearance of the probe letter again prompted gamma 118 activity (Fig 3 g). This site coincides with the generator of scalp EEG that was found 119 in the parietal cortex for the same task [3]. The PSD thereby confirmed the findings of 120 local synchronization of cortical activity during WM maintenance [3,8,9]. 121 In the hippocampus of all three participants, we found elevated activity in the 122 beta range (12-24 Hz) towards the end of the maintenance period (increase >100%, 123 Fig 2 e, Fig 3 c, h)

136
What was the directionality of the information flow during encoding and 137 maintenance in a trial? We used spectral Granger causality (GC) as a measure of 138 directed functional connectivity to determine the direction of the information flow 139 between auditory cortex and hippocampus in Participant 1 during the trials. To 140 improve legibility, we present GC as Granger (%) = GC*100. During encoding, the 141 information flow was from auditory cortex to hippocampus with a maximum in the 142 theta frequency range (dark blue curve in Fig. 2

h). The net information flow 143
ΔGranger (GC hippcortex -GC cortexhipp) during encoding was significant in 144 the 6-8 Hz range (blue bar in Fig. 2 h, p<0.05 permutation test against a null 145 distribution). During maintenance, the information flow in the theta frequency range 146 was reversed (dark red curve), i.e. from hippocampus to auditory cortex (dark red 147 curve in Fig. 2 h). The net information flow ΔGranger during maintenance was 148 significant in the 5-8 Hz range (red bar in Fig. 2 h, p<0.05 permutation test against a 149 null distribution). Concerning the spatial spread of the theta GC, the maximal net 150 information flow ΔGranger (GC hippcortex -GC cortexhipp) during encoding 151 occurred from auditory cortex to hippocampus (p<0.05, permutation test, Fig. 2

i). 152
During maintenance, the theta ΔGranger was significant from hippocampus to both 153 auditory cortex and occipital cortex (permutation test p<0.05, Fig. 2j). Interestingly, in 154 Participant 1, the distribution of high ΔGranger coincides with the distribution of high 155 PLV: both show a spatial maximum to grid contacts over auditory cortex and both 156 appear in the theta frequency range. 157 We next tested the statistical significance of the spatial spread of contacts with 158 high ΔGranger (4-8 Hz) during maintenance ([-2 0] s). To provide a sound statistical 159 basis, we tested the spatial distribution of GC on the grid contacts against a null 160 distribution. The activation on grid contacts was reshaped into a grid vector. The 161 spatial collinearity of two grid vectors was captured by their scalar product. We next 162 performed 200 iterations of random trial permutations. For each iteration we selected 163 two subsets of trials and we calculated the scalar product between the two vectors 164 corresponding to these subsets. We then tested the statistical significance of the 165 scalar product (Fig. 2 k). The true distribution (red) is clearly distinct from the null 166 distribution (gray, blue bar marks the 95th percentile). The analogous procedure was 167 performed for PSD (Fig. 2 a, d), PLV (Fig. 2 g) and GC during encoding ( Fig. 2 i), 168 which gave equally significant results in all cases. 169 As a further illustration of the ΔGranger time-course, the time-frequency plot 170 ( Fig. 2 l) shows the difference between GC spectra (GC hipp cortex -GC cortex 171 hipp) at each time point, where blue indicates net flow from auditory cortex to 172 hippocampus and red indicates net flow from hippocampus to auditory cortex. 173 Similarly in Participant 2, the time course of GC followed the same pattern 174 between auditory cortex (anterior strip electrode contact in Fig. 3 a) and 175 hippocampus ( Fig. 3 d,e). Among the three participants that had both LFP and 176 temporo-parietal ECoG recordings, Participant 3 had an electrode contact over visual 177 cortex; the sensory localization was indexed by the strong gamma activity in the most 178 posterior contact of the strip electrode ( Fig. 3 g). The time-course of information flow 179 between visual cortex and hippocampus ( Fig. 3 i,j) followed the same pattern as 180 described for the auditory cortex above. Thus, letters were encoded with information 181 flow from sensory cortex to hippocampus; conversely, the information flow from 182 hippocampus to sensory cortex indicated the replay of letters during maintenance. 183

184
We used beamforming [30] to reconstruct the EEG sources during encoding 185 and maintenance for each of the 15 participants (Table 1). We tested whether the 186 sources during fixation differed from sources during encoding and during 187 maintenance (non-parametric cluster based permutation t-test [31,32]). In each 188 participant, the proportion of significant sources in the left hemisphere exceeded 80% 189 of all significant sources. Across all participants, the spatial activity pattern during 190 both encoding and maintenance showed the highest significance in frontal and 191 temporal areas of the left hemisphere (Fig S1). 192

194
The synchronization between hippocampal LFP and EEG sources (N = 15 195 participants) confirmed the directed functional coupling found in the three participants 196 with ECoG. We first calculated the GC between hippocampus and the EEG 197 beamforming sources in the auditory cortex. We found that the mean GC spectra 198 resembled the GC spectrum for ECoG in the theta frequency range ([4 8] Hz, Fig.  199 4a). During encoding, the net information flow was from auditory cortex to 200 hippocampus (light blue curve -dark blue curve, blue bar, group cluster-based 201 permutation test). During maintenance, the net information flow was reversed (dark 202 red curve -light red curve, red bar, group cluster-based permutation test), i.e. from 203 hippocampus to auditory cortex. Thus, both for ECoG and EEG sources, GC showed 204 the same bidirectional effect in theta between auditory cortex and hippocampus. 205 To explore the spatial distribution, we computed GC also for other areas of 206 cortex. We averaged the net information flow (ΔGranger) in the theta range across 207 the participants and projected it onto the inflated brain surface (Fig 4b, c). During 208 encoding, the mean information flow was strongest from auditory cortex to 209 hippocampus (ΔGranger = -4.9 %, p = 0.0009, Kruskal-Wallis test, Fig 4b). For all 210 other areas, the mean ΔGranger was also from cortex to hippocampus but the effect 211 was weaker (mean ΔGranger = [-3 0]%, Dunn's test, Bonferroni corrected). During 212 maintenance (Fig 4c) the information flow was reversed. While all areas had 213 information flow from hippocampus to cortex (ΔGranger = [0 2]%, Dunn's test, 214 Bonferroni corrected), the strongest flow appeared from hippocampus to auditory 215 cortex (ΔGranger = 3.4%, p = 0.001, Kruskal-Wallis test). 216

217
The reversal of ΔGranger appeared in all 15 participants individually (Fig 4d). 218 We averaged ΔGranger for each participant in the [4-8] Hz theta frequency range. 219 The ΔGranger between hippocampus and auditory cortex, was negative during 220 encoding and was positive during maintenance in the (p = 4.1e-10, paired 221 permutation test). The directionality and its reversal was missing for all other areas, 222 e.g. lateral prefrontal cortex (p = 0.16, paired permutation test, Fig 4e). Of note, all 223 analyses up to here were performed on correct trials only. 224 Finally, we established a link between the participants' performance and 225 ΔGranger. For incorrect trials, the net information flow ΔGranger from auditory cortex 226 to hippocampus did not show the same directionality in all participants and did not 227 reverse in direction (p = 0.37, paired permutation test, Fig. 4f). Since participants 228 performed well (median performance 91%), we balanced the numbers of correct and 229 incorrect trials. We calculated the GC in a subset of correct trials (median of 200 230 permutations of a number of correct trials that equals the mean percentage of 231 incorrect trials = 10%); the effect was equally present for the subset of correct trials 232 Fig 4d). This suggests that timely information flow, as indexed by GC, is 233 relevant for producing a correct response. 234

235
Working memory (WM) describes our capacity to represent sensory input for 236 prospective use. Our findings suggest that this cognitive function is subserved by 237 bidirectional oscillatory interactions between the hippocampus and the auditory 238 cortex as indicated by phase synchrony and Granger causality. In our verbal working 239 memory task, the encoding of letter items is isolated from the maintenance period in 240 which the active rehearsal of memory items is central to achieve correct 241 performance. First, analysis of task-induced power showed sustained oscillatory 242 activity in cortical and hippocampal sites during the maintenance period. Second, 243 analysis of the inter-electrode phase synchrony and the directional information flow 244 showed task-induced interactions in the theta band between cortical and 245 hippocampal sites. Third, the directional information flow was from auditory cortex to 246 hippocampus during encoding and, during maintenance, the reverse flow occurred 247 from hippocampus to auditory cortex. This pattern was dominant on the left cortical 248 hemisphere, as expected for a language related task. Fourth, the comparison 249 between correct and incorrect trials suggests that the participants relied on timely 250 information flow to produce a correct response. Our data suggests a surprisingly 251 simple model of information flow within a network that involves sensory cortices and 252 hippocampus ( Fig. 4 g): During encoding, letter strings are verbalized as melody. 253 The incoming information flows from sensory cortex to hippocampus (bottom-up). 254 During maintenance, participants actively recall and rehearse the melody in their 255 phonological loop [1,2]. The Granger causality indicates the information flow from 256 hippocampus to cortex (top-town) as the physiological basis for the replay of the 257 memory items, which finally guides action. 258 The current study is embedded in previous studies using the same or similar 259 tasks. Persistent firing of hippocampal neurons indicated hippocampal involvement in 260 the maintenance of memory items [19,22,23]. An fMRI study reports salient activity 261 in the auditory cortex during maintenance in an auditory working memory task [33], 262 which indicates that sensory cortical areas are involved in the maintenance of WM 263 items. During encoding, the activity of local assemblies was associated with gamma 264 frequencies and local processing (Fig. 2 a b c, Fig. 3 g) while GC inter-areal 265 interactions took place in theta frequencies, in line with previous reports [29,34]. 266 Parietal generators of theta-alpha EEG indicated involvement of parietal cortex in 267 WM maintenance [3,5,7,19,35]. The hippocampo-cortical phase synchrony (PLV) 268 was high during maintenance of the high workload trials [19]. Building on these 269 previous studies, the current study focused on high workload trials and extended 270 them by the analysis of directional information flow. 271 In the design of the task we aimed to separate in time the encoding of memory 272 items from their maintenance. In the choice of the 2s duration for the encoding period 273 were guided by the magic number 7±2, which may correspond to "how many items 274 we can utter in 2 seconds" [1, 2]. The median Cowan's k = 6.1 shows that high-275 workload trials were indeed demanding for the participants, where both encoding and 276 maintenance may limit performance. We therefore presented the letters both as a 277 visual and an auditory stimulus. Certainly, maintenance processes are likely to 278 appear already during the encoding period as maintenance neurons ramp up their 279 activity already during encoding [19]. Furthermore, encoding may extend past the 280 visual stimulus (t = -3 s). We therefore focused our analysis on the last two seconds 281 of maintenance [-2 0] s. With this task design, we found patterns of GC that were 282 clearly distinct between encoding and maintenance. 283 The interaction between recordings from different brain regions has to be 284 discussed with respect to volume conduction [36]. Of note, there was a strong 285 frequency dependence of GC from hippocampus to ECoG (Fig. 2 h, Fig. 3 d, i). 286 Likewise, GC to EEG sources showed a strong frequency dependence (Fig. 4 a). 287 This speaks against volume conduction because the transfer of signal through tissue 288 by volume conduction is independent of frequency in the range of interest here [37, 289 38]. Furthermore, there was a strong task dependence of GC (Fig. 2 h, Fig. 3 d, i,  290   Fig. 4 a) We used a modified Sternberg task in which the encoding of memory items 310 and their maintenance were temporally separated (Fig. 1a). Each trial started with a 311 fixation period ([−6, −5] s), followed by the stimulus ([−5, −3] s). The stimulus 312 consisted of a set of eight consonants at the center of the screen. The middle four, 313 six, or eight letters were the memory items, which determined the set size for the trial 314 (4, 6, or 8 respectively). The outer positions were filled with "X," which was never a 315 memory item. The participants read the letters and heard them spoken at the same 316 time. After the stimulus, the letters disappeared from the screen, and the 317 maintenance interval started ([−3, 0] s). Since the auditory encoding may have 318 extended beyond the 2 s period, we restrict our analysis to the last 2 s of the 319 maintenance period ([−2, 0] s). A fixation square was shown throughout fixation, 320 encoding, and maintenance. After maintenance, a probe was presented. The 321 participants responded with a button press to indicate whether the probe was part of 322 the stimulus. The participants were instructed to respond as rapidly as possible 323 without making errors. After the response, the probe was turned off, and the 324 participants received acoustic feedback regarding whether the response was correct 325 or incorrect. The participants performed sessions of 50 trials in total, which lasted 326 approximately 10 min each. Trials with different set sizes were presented in a 327 random order, with the single exception that a trial with an incorrect response was 328 always followed by a trial with a set size of 4. The task can be downloaded at 329 www.neurobs.com/ex_files/expt_view?id=266. 330

331
The participants in the study were patients with drug resistant focal epilepsy. 332 To investigate a potential surgical treatment of epilepsy, the patients were implanted 333 with intracranial electrodes. The participants provided written informed consent for 334 the study, which was approved by the institutional ethics review board (PB 2016-335 02055). The participants were right-handed and had normal or corrected-to-normal 336 vision. For nine participants (4 -13), the PSD and PLV has been reported in an 337 earlier study [19]. 338

339
The depth electrodes (1.3 mm diameter, 8 contacts of 1.6 mm length, spacing 340 between contact centers 5 mm, ADTech®, Racine, WI, www.adtechmedical.com) 341 were stereotactically implanted into the hippocampus. Subdural grids and strips were 342 placed directly on the cortex according to the findings of the non-invasive presurgical 343 evaluations. Platinum electrodes with 4 mm 2 contact surface and 1 cm inter-electrode 344 distances were used (ADTech®). In addition, scalp EEG electrodes were placed at 345 the sites of the 10-20 system with minor adaptations to avoid surgical scalp lesions. 346

347
To localize the ECoG grids and strips, we used the participants' postoperative 348 MR, aligned to CT and produced a 3D reconstruction of the participants' pial brain 349 surface. Grid and strip electrode coordinates were projected on the pial surface as 350 described in [41] (Fig. 2a, Fig. 3a,f). 351 The stereotactic depth electrodes were localized using post-implantation 352 computed tomography (CT) and post-implantation structural T1-weighted MRI 353 scans. The CT scan was registered to the post-implantation scan as implemented in 354 FieldTrip [42]. A fused image of CT and MRI scans was produced and the electrode 355 contacts were marked visually. The hippocampal contact positions were projected on 356 a parasagittal plane of MRI (Fig. 1b). 357 Some of the electrodes contacts were found in tissue that was deemed to be 358 epileptogenic and that was later resected. Still, neurons in this tissue have been 359 found to participate in task performance in an earlier study [19]. 360

361
All recordings were performed with Neuralynx ATLAS, sampling rate 4000 Hz, 362 0.5-1000 Hz passband (Neuralynx, Bozeman MT, USA, www.neuralynx.com). ECoG 363 and LFP were recorded against a common intracranial reference. Signals were 364 analyzed in Matlab (Mathworks, Natick MA, USA). We re-referenced the 365 hippocampal LFP against the signal of a depth electrode contact in white matter. We 366 re-referenced the cortical ECoG against a different depth electrode contact. The 367 choice of two separate references for LFP and ECoG has been shown to avoid 368 spurious functional connectivity estimates [39]. The scalp EEG was recorded against 369 an electrode near the vertex and was then re-referenced to the averaged mastoid 370 channels. All signals were downsampled to 500 Hz. All recordings were done at least 371 6 h from a seizure. Trials with large unitary artefacts in the scalp EEG were rejected. 372 We focused on trials with high workload (set sizes 6 and 8) for further analysis. We 373 used the FieldTrip toolbox for data processing and analysis [30]. 374

375
We first calculated the relative power spectral density (PSD) in the time-376 frequency domain (Fig. 2 b). Time-frequency maps for all trials were averaged. We where PLVi,j is the PLV between channels i,j, N is the number of trials, X(f) is the 389 Fourier transform of x(t), and (•)* represents the complex conjugate. 390 Using the spectra of the two-second epochs, phase differences were calculated for 391 each electrode pair (i,j) to quantify the inter-electrode phase coupling. The phase 392 difference between the two signals indexes the coherence between each electrode 393 pair and is expressed as the PLV. The PLV ranges between 0 and 1, with values 394 approaching 1 if the two signals show a constant phase relationship over all trials. 395 In our description of EEG frequency bands, we used theta (4-8 Hz), alpha (8-396 12 Hz), beta (12-24 Hz) and gamma (> 40 Hz), while the exact frequencies may differ 397 in individual participants. 398 399

400
We reconstructed the scalp EEG sources using linearly constrained minimum 401 variance (LCMV) beamformers in the time domain. To solve the forward problem we 402 used a precomputed head model template and aligned the EEG electrodes of each 403 participant to the scalp compartment of the model. We then computed the source grid 404 model and the leadfield matrix wherein we determined the grid locations according to 405 the brain parcels of the automated anatomical atlas (AAL) [43]. We solved the 406 inverse problem by scanning the grid locations using the LCMV filters separately for 407 encoding and maintenance. The EEG sources were baselined with respect to the 408 fixation period and presented as a percent of change from the pre-stimulus baseline. 409 We defined cortical areas from multiple parcels since AAL is a parcellation based on 410 sulci and gyri. We performed all the steps of the source reconstruction with FieldTrip 411 [30] and projected the sources onto an inflated brain surface. 412

413
In order to evaluate the direction of information flow between the hippocampus 414 and the cortex, we calculated spectral non-parametric Granger causality (GC) as a 415 measure of directed functional connectivity analysis [30]. We evaluated the direction 416 of information flow in the [4 20] Hz frequency range. To compute the GC we first 417 downsampled the signals to the Nyquist frequency = 40 Hz. We then computed the 418 GC between hippocampal contacts and ECoG grid contacts. We also computed GC 419 between the same hippocampal contacts and EEG sources located over the regions 420 of interest. GC examines if the activity on one channel can forecast activity in the 421 target channel. In the spectral domain, GC measures the fraction of the total power 422 that is contributed by the source to the target. We transformed signals to the 423 frequency domain using the multitaper frequency transformation method (2 Hann 424 tapers, frequency range 4 to 20 Hz with 20 seconds padding) to reduce spectral 425 leakage and control the frequency smoothing. 426 We used a non-parametric spectral approach to measure the interaction in the 427 channel pairs at a given interval time [44]. In this approach, the spectral transfer 428 matrix is obtained from the Fourier transform of the data. We used the FieldTrip 429 toolbox to factorize the transfer function H(f) and the noise covariance matrix Σ. The where S_(xx(f)) is the total power and S _(xx(f)) the instantaneous power. To 436 improve legibility, we present GC as Granger % = GC*100. To average over the 437 group of participants, we calculated the Granger spectra for the selected channel 438 pairs and averaged these spectra over participants (Fig 4a). 439 To illustrate the time course of GC over time, we calculated time-frequency 440 maps with the multitaper convolution method of Fieldtrip [30]. 441

442
To analyze statistical significance, we used cluster-based nonparametric 443 permutation tests. To assess the significance of the difference of the Granger 444 between different directions, we compared the difference of the true values to a null 445 distribution of differences. We recomputed GC after switching directions randomly 446 across trials, while keeping the trial numbers for both channels constant. Then we 447 computed the difference of GC for the two conditions. We repeated this n = 200 times 448 to create a null distribution of differences. The null distribution was exploited to 449 calculate the percentile threshold p = 0.05. In this way, we compare the difference of 450 the dark and light spectra against a null distribution of differences. We mark the 451 frequency range of significant GC with a blue bar for encoding (dark blue spectrum 452 exceeds light blue spectrum, information flow from cortex to hippocampus) and with a 453 red bar for maintenance (dark red spectrum exceeds light red spectrum, information 454 flow from hippocampus to cortex). 455 To test the statistical significance of the spatial spread of contacts with high 456 PSD, PLV, or ΔGranger, we calculated the spatial collinearity on the grid contacts 457 against a null distribution. First, we transform the activation on grid contacts into a 458 grid vector. We then performed 200 iterations of random trial permutations. For each 459 iteration we selected two subsets (50%) of trials and we calculated the scalar product 460 between the vectors corresponding to the two subsets. The null distribution was 461 created by randomly mixing trials from the two task periods fixation and encoding. 462 We finally tested the statistical significance of the scalar product. The true distribution 463 was established to be statistically distinct from the null distribution if it exceeded the 464 95th percentile of the null distribution. 465 We assess if the reconstructed EEG sources during encoding and 466 maintenance are significantly different from the pre-stimulus baseline (fixation). We 467 use the FieldTrip's method ft_sourcestatistics [30] wherein we apply a non-468 parametric permutation approach to quantify the spatial activation pattern during the 469 encoding of the memory items and their active replay. Competing interests 491 All authors declare that they have no competing interests. 492

493
The participants provided written informed consent for the study, which was 494 approved upfront by the institutional ethics review board (PB 2016-02055). 495

496
All data needed to evaluate the conclusions in the paper are present in the paper. All 497 codes used to produce the results in the paper can be requested from the authors. 498 The task can be downloaded at www.neurobs.com/ex_files/expt_view?id=266. Part 499 of the data has been published earlier [35]. Additional data and code are indexed in 500 www.hfozuri.ch. 501   size (4, 6 or 8 letters) determines WM workload. In each trial, presentation of a 656 letter string (encoding period, 2 s) is followed by a delay (maintenance period, 3 s). 657

Author contributions
After the delay, a probe letter is presented. Participants indicate whether the probe 658 was in the letter string or not. during, fixation (black) encoding (blue) and maintenance (red). The PLV spectra 679 show a broad frequency distribution. The PLV during maintenance is higher than 680 during fixation. Red bars: frequency ranges of significant PLV difference (p<0.05, 681 cluster-based nonparametric permutation test against a null distribution with 682 scrambled trials during fixation and maintenance). 683 g) Phase locking value (PLV) between hippocampus and cortex in theta (4-8 Hz) 684 during maintenance ([-2 0] s) is highest to contacts over auditory cortex. 685 b) The relative power spectral density (PSD) in the temporal scalp EEG electrode 724