Fig. 3 |. Memory updating depends on contextual inference.

a, Participants experienced two contexts defined by a sensory cue (right or left target) paired with a perturbation sign (P⁺ or P⁻). Participants moved a control point (right vs. left, grey disk) on a virtual bar to the corresponding target ¹⁹. For clarity, schematic not to scale. The colours of cues and perturbations indicate the context to which they are associated (blue and red for context 1 and 2, respectively). b, Training: cues (background colour) were consistently paired with perturbations (black line) randomly selected on each trial (only a few trials shown for clarity). c, Triplets: two channel trials (both with cue 1, ${\text{P}}_1^{\text{c}}$) bracket an ‘exposure’ trial that uses one of the four possible cue-perturbation combinations. Single-trial learning for the memory associated with context 1 is measured as the difference (Δ) in adaptation across the two channel trials. d Single-trial learning for context 1 before (top) and after (bottom) training. Experimental data (mean ± SEM, column 1) across n = 24 participants (dots). Positive values indicate single-trial learning consistent with the exposure trial perturbation (increase following P⁺ and decrease following P⁻). The average (± SEM across participants, column 2) of the individual COIN model fits (8 parameters each, Extended Data Fig. 3).