Our results indicate that when a visual conflict reoccurs, traces of past perception increase the probability that the visual system will assume a previous state of interpretation rather than a different one. Despite its apparently simple nature, such a mechanism could have great merit for visual function. Visual input quite generally contains ambiguities, and in normal conditions only one perceptual interpretation is veridical. Selecting the one correct interpretation often requires the combination of multiple information sources 
and engages extensive regions of the brain 
. The current observations suggest that by biasing the system toward previous perceptual interpretations-be it the most recent one or one that consistently dominated longer ago-visual memory automatically enforces the outcome of previous perceptual conflicts, and thereby eliminates the need for the same conflicts to be resolved repeatedly.
Our model work shows that in neural terms this memory could be carried by the adaptation state of sensory neurons, provided adaptation occurs on more than one timescale. Indeed, adaptation in known to occur on a wide range of timescales in sensory cortex 
, and it has been argued on theoretical grounds that this holds promise for functional forms of history-dependence in neural systems 
. The present work thus suggests that the perceptual memory observed here constitutes one such functional correlate.
Our observations seem to conflict with an earlier study 
where percept choice at stimulus reappearance was interpreted as an attempted continuation of the preceding percept sequence. For instance, a sequence of percepts ABA would cause percept B at reappearance, forming the regular sequence ABAB. This is inconsistent with our data mainly because it involves suppressive effects of past dominance on subsequent percept choice (for instance when ABA causes B whereas BBA causes A, the initial percept stimulates opposite perception at the end of the sequence), whereas we find only facilitation. The discrepancy may be due to the fact that the stimuli used by 
were not fully ambiguous but biased toward one interpretation. The use of ambiguous stimuli throughout our experiments has allowed us to characterize memory of the internally constructed interpretation of a stimulus in isolation. Effects of unbalanced stimuli are probably a combination of the present effects and differential gain control at stages prior to where the percept arises 
. This is known to cause different effects on subsequent perception 
We modeled our findings by adding a longer adaptation timescale to an existing model of perceptual stabilization 
. An alternative model 
of stabilization has also been proposed. Both models are essentially standard oscillator models expanded with an additional interaction 
or storage mechanism 
to allow a trace of previous perception to bias the next percept choice at stimulus reappearance. Both models in their original form have the limitation of lacking multiple timescales of storage. Regardless of the number of timescales, a drawback of 
is that is predicts stabilization of a percept that has dominated briefly before stimulus offset but no stabilization of a percept that has dominated longer [Figure 7 in 23]
. This is opposite to the experimental finding that brief dominance will prevent stabilization, and longer dominance is required for a percept to recur [Figure 2b in the present work, Figure 3c in 3]
. Models of the type of 
do reproduce this feature. Arguably a second objection to 
is that it entails a binary memory, where the system is in one of two states of ‘remembering’ either percept. Experiments indicate that, instead, the system's bias toward one or the other percept varies over time in a continuous fashion [Figures 2b, 2e and 3a in the present work, also 24]
, consistent with the model we used.
Our view of ambiguous figure memory suggests a relation to visual memory in other situations. Previous notions that attributed perceptual stabilization to prolongation of a perceptual state during stimulus absence seemed to imply that it is a specifically ambiguity-related phenomenon. The present view of perceptual stabilization as a bias in a decision network–in this case regarding a perceptual decision at stimulus onset–allows more room for extensions beyond ambiguous perception. Specifically, the accumulation of a bias during perceptual dominance that we observe here is reminiscent of the progressive decrease in response time that is observed when subjects direct their attention 
or eye fixation 
to a similar search target appearing on several consecutive trials. This type of attention priming occurs automatically, independent of conscious recollection. It has been attributed to progressive use-related changes that build up in the neural structures activated when the target is attended 
, so that every allocation of attention or gaze to an item simultaneously acts to stimulate reorientation to that item in the future. This is analogous to the accumulating bias that facilitates repeated perceptual dominance in our paradigm, a similarity that is particularly remarkable considering the numerous other parallels between attentional selection and perceptual dominance 
Our findings bear directly on the question asked at the outset, how long a history to incorporate into current processing. Functionally, the answer depends on the liability for the conditions to change. If they change every few seconds it is useless to incorporate a minute-scale history because what happened a minute ago bears little relation to the present situation. If, in contrast, the conditions remain relatively stable for minutes, incorporating a longer-term history prevents unfavorable sensitivity to seconds-long (noisy) excursions. Our findings suggest how just such a strategy is implemented in vision, by use of parallel biasing traces on several timescales. In case of ambiguity resolution, if recent perception was highly stable, slow biases have built up sufficiently to outweigh the fast bias due to the most recent percept. If perception was variable, however, no slow biases have accumulated and the most recent percept becomes the main driving factor. This organization therefore ensures automatic adjustment of the effective memory timescale, dependent on the changeability of the situation at hand.