The predominant focus in the neurobiological study of memory has been on remembering (persistence). However, recent studies have considered the neurobiology of forgetting (transience).

Here we draw parallels between neurobiological and computational mechanisms underlying transience. We propose that it is the interaction between persistence and transience that allows for intelligent decision-making in dynamic, noisy environments.

Specifically, we argue that transience (1) enhances flexibility, by reducing the influence of outdated information on memory-guided decision-making, and (2) prevents overfitting to specific past events, thereby promoting generalization. According to this view, the goal of memory is not the transmission of information through time, per se. Rather, the goal of memory is to optimize decision-making. As such, transience is as important as persistence in mnemonic systems.

In the practical use of our intellect, forgetting is as important as remembering.

William James (The Principles of Psychology)

Above, we reviewed a number of neurobiological mechanisms that can promote mnemonic transience. The most intuitive explanation for why the brain possesses these mechanisms is that they help to “make room” for new memories. However, when we consider the sheer number of neurons and synapses in the brain, it would seem that there is ample capacity to store many more memories than we actually do. For example, the human brain is estimated to have roughly 80–90 billion neurons (Azevedo et al 2009). If we were to reserve only a tenth of those for memories of specific events, then according to computational estimates of capacity in auto-associative networks, we could reliably store approximately one billion individual memories (Amit et al 1985). Furthermore, when we consider sparsely encoded memories this number can increase by several orders of magnitude (Amari1989). Given that it is apparently possible to remember far more than most of us actually do, why did evolution endow most individuals with brains that work to prevent faithful transmission of information through time? In other words, is there a utility to memory transience, given the seemingly obvious benefits of memory persistence?

We propose that memory transience is required in a world that is both changing and noisy. In changing environments, forgetting is adaptive because it allows for more flexible behavior. In noisy environments, forgetting is adaptive because it prevents overfitting to peculiar occurrences. According to this perspective, memory persistence is not always useful. For example, persistence of memory for aspects of the world that are either transient or uncommon would be detrimental since it might lead to inflexible behavior and/or incorrect predictions. Rather, persistence is only useful when it maintains those aspects of experience that are either relatively stable and/or predictive of new experiences. Therefore, it is only through the interaction of persistence and transience (persistence × transience) that memory actually serves its true purpose: using the past to intelligently guide decision-making (for related viewpoints, see Dudai & Carruthers2005, Schacter et al 2007). Below, we review the computational case for using transience to increase behavioral flexibility and promote generalization. In addition, we identify the parallels between how transience is used computationally and how it appears to be implemented in the brain.

New learning represents large challenges for neural networks that use distributed representations (French1999, Lewandowsky & Li1995, McCloskey & Cohen1989, Ratcliff1990). The challenges are two.

First: new learning might overwrite previous memories (ie. catastrophic interference), and in turn, new learning is impeded by existing, stored memories (ie proactive interference) (Burgess et al 1991, McCloskey & Cohen1989, Palm2013, Siegle & Hasselmo2002). This is the “stability versus plasticity” dilemma in neural networks (Abraham & Robins2005, Carpenter & Grossberg1987). As such, according to the traditional view, memory persistence is incompatible with behavioral flexibility because a network that is good at maintaining persistent memories will be poor at learning new information, especially if it conflicts with previous experiences. However, recent neural network models that use external memory devices or synapses that change over multiple timescales challenge the universality of this dilemma (Graves et al 2016, Kirkpatrick et al 2017, Santoro et al 2016a). Moreover, another strategy the brain can use to solve this dilemma is to sparsely encode experiences using orthogonal representations, which may potentially arise from pattern separation processes (see Yassa & Stark2011 for a review). The contextual dependence of memory is one example of this strategy: by maintaining orthogonal patterns, memories that are encoded in a particular context are more likely to be expressed in that context, but not other contexts (Maren et al 2013). This type of strategy maximizes the number of patterns that can be stored within a neural network without interference (Amari1989).

Seocnd: however, in dynamic environments it might also be important to discard outdated information regardless of any capacity constraints (Kraemer & Golding1997). If the environment changes, but our memories do not, then we may persevere to our own detriment. Therefore, transience may facilitate decision-making by eliminating outdated (and potentially misleading) information, allowing an organism to respond more efficiently to changes in its environment.

Consistent with this idea, recent studies provide evidence that forgetting is necessary for flexible behavior in dynamic environments (Dong et al 2016, Epp et al 2016, Shuai et al 2010). As introduced above, Shuai and colleagues trained Drosophila flies to discriminate two odors (odor A, paired with shock [A+] versus odor B, not paired with shock [B−]) and found that Rac1 inhibition slowed forgetting (Shuai et al 2010_14ya). They then asked, to what extent slower forgetting would now interfere with reversal learning. Accordingly, they retrained the flies but reversed the odor-shock contingencies (ie. A− and B+). Flies in which Rac1 was inhibited (ie. flies displaying slower forgetting) exhibited impaired reversal learning, indicating that increased persistence of odor-shock memories interfered proactively with new learning (thereby reducing flexibility). Conversely, flies in which Rac1 was activated had the opposite phenotype. They exhibited accelerated forgetting, and this increased forgetting facilitated reversal learning (thereby increasing flexibility). This pattern of results extended to 5 different lines of flies engineered to express mutations linked to autism spectrum disorder that also interfere with Rac activity. All these lines of flies with disrupted Rac function exhibited impaired forgetting, and this, in turn, impaired reversal learning (Dong et al 2016).