a, We examined five different possible determinants of trial-by-trial contextual memory retrieval in the Squircle. For each possible determinant, we tested whether that determinant predicted contextual memory retrieval on each Squircle trial (that is, a Square- or Circle-consistent replace location) and computed classification accuracy across all trials for each participant (% correct, compared to 50% theoretical chance baseline dotted line). The five determinants included: (i) The modal Square (Sq) or Circle (Ci) context during the N odd testing trials prior to each individual Squircle trial. Although trial order was balanced across all Squircle trials, for each individual trial, participants were tested more often in either the Sq or Ci every odd number of prior Sq and Ci trials, which may bias contextual memory. For example, for a given Squircle trial, if the three preceding trials were two Sq and one Ci, a participant may be more likely to retrieve a Sq-consistent memory. We thus tested whether Squircle memory retrieval on each individual Squircle trial was predicted by the modal context of the N = 1, 3, 5 prior Sq and Ci trials. However, replace locations in the Squircle were not predicted by the previous trials’ modal context (t23 = 0.98, 1.04, 1.75, pFWE = 0.508, 0.462, 0.140 for lags 1,3, and 5, respectively). (ii) Distance error during training. Although distance error during training was matched on average between the Sq and Ci, for each target object each participant had numerically lower distance error for that object in one of the two training contexts. For each participant, we tested whether Squircle contextual memory was determined by whichever training context had the lowest distance error for each object. Classification accuracy was not significantly higher than expected by chance (t23 = 0.985, p = 0.335). (iii) Squircle trial start location. Each Squircle trial started a random arena location, which may have biased contextual memory. For each Squircle trial, we tested whether contextual memory retrieval was predicted by whichever context-consistent location was nearest to that trial’s starting location. Prediction accuracy was not significantly higher than expected by chance (t23 = 1.41, p = 0.175). iv) Contextual memory stability (Mem Stability) across Squircle trials within (WI) the same participant. For each target object, did each participant retrieve the same contextual memory for that object across all Squircle trials? For each Squircle trial, we tested whether memory was predicted by the modal contextual memory retrieved on all other Squircle trials for the same target object. We repeated this separately for each participant. For each target object, participants tended to retrieve the same contextual memory for that object across Squircle trials (t23 = 5.28, p = 2.34 × 10−5). v) Mem Stability across Squircle trials between (BW) participants. For each target object, did each participant retrieve the same contextual memory for that object as the other participants? For each Squircle trial, we tested whether memory was predicted by the modal contextual memory retrieved for that trial’s target object by a different participant across all Squircle trials, jackknifed across all participants. For each target object, participants tended to retrieve the same contextual memory as other participants (t23 = 9.58, 1.70 × 10−9). However, choice stability was lower across participants than within the same participant (t23 = 4.73, p = 9.22 × 10−5), indicating that although there was some choice stability between participants, each participant tended to have a unique contextual preference. b, Participants replaced target objects in the Squircle closer to a context-consistent location than to the location intermediate between the two context-consistent locations (t = 10.86, p = 1.57 × 10−10), and this was the case in 23/24 (96%) individual participants, providing further evidence that participants retrieved either the Square or Circle-consistent contextual memories in the Squircle. Throughout the figure, error bars indicate ±1 SEM; All t-test two-tailed; Dots denote individual participants (n = 24). nsp > 0.05, ***p < 0.001.