# A Bayesian model of memory in a multi-context environment

- Dave Kleinschmidt, Psychology, Rutgers University, New Brunswick, New Jersey, United States
- Pernille Hemmer, Rutgers University, Piscataway, New Jersey, United States

**Abstract** In a noisy but structured world, memory can be improved by enhancing limited
stimulus-specific memory with statistical information about the context. To
do this, people have to learn the statistical structure of their current
environment. We present a Sequential Monte Carlo (particle filter) model of
how people track the statistical properties of the environment across multiple
contexts. This model approximates non-parametric Bayesian clustering of
percepts over time, capturing how people impute structure in their perceptual
experience in order to more efficiently encode that experience in memory.
Each trial is treated as a draw from a context-specific distribution, where
the number of contexts is unknown (and potentially infinite). The model
maintains a finite set of hypotheses about how the percepts encountered thus
far are assigned to contexts, updating these in parallel as each new percept
comes in. We apply this model to a recall task where subjects had to recall
the position of dots [@Robbins2014]. Unbeknownst to subjects, each dot
appeared in one of a few pre-defined regions on the screen. Our model
captures subjects' ability to learn the inventory of contexts, the statistics
of dot positions within each context, and the statistics of transitions
between contextsâ€”as reflected in both recall and prediction.