The target article by Barzykowski and Moulin (B&M) argues that involuntary autobiographical memory (IAM) and déjà vu are the result of an integrated system of memory and that they naturally arise from recognition and memory retrieval processes. While we agree with this stance, we find the proposed memory architecture lacking in two respects. First, the focus of the authors on the phenomenology of IAM and déjà vu neglects the functional role of these phenomena, the justification for why they might exist, and how they might be used by agents. Second, the memory architecture, as summarized in Figure 1, is only imprecisely specified, leaving room for alternate theories, potential inconsistencies, and omitted details. Here, we consider how performing a rational analysis of, and building computational models of, IAM and déjà vu can mitigate these problems. We use our work on how IAM can support prospective memory as an example of addressing both problems (Li & Laird, Reference Li and Laird2015), then extend that reasoning to the familiarity judgments that underlie déjà vu.
The rational analysis framework assumes that cognitive processes are optimally adapted to the functional goals of the agent, while subject to ecological constraints and limits on biological and cognitive resources (Anderson, Reference Anderson1990; Lieder & Griffiths, Reference Lieder and Griffiths2020). For memory, we take its primary function to be to “bring past experience to bear on present action” (Anderson, Reference Anderson1994), operating within a small working memory capacity, fixed bandwidth to long-term memory, and other cognitive constraints. This serves as the starting point for understanding the functional role of phenomena such as IAM.
A hypothesis about the structure of memory can be tested via its implementation in a computational model. These models force researchers to be precise in their definitions of the computational representations and processes that underlie their theories and ensure that hypothesized theoretical models of memory are consistent both internally and with broader theories of cognition. This is particularly true in cognitive architectures such as ACT-R (Anderson, Reference Anderson2007), which integrate multiple cognitive processes and potential neural correlates into a single system. This enables the evaluation of their combined performance across multiple tasks, thus ensuring that a hypothesis is compatible with the same mechanisms used to model other phenomena.
Consider, for example, the hypothesis that IAM is the result of automatic matching of sensory and abstract cues with items in memory in order to “quickly rais[e] pertinent information to consciousness without effort” (target article, sect. 4, para. 3). A rational analysis of IAM would start by considering the limits of deliberate retrieval and situations where those limits are exceeded. One such situation is in the prospective memory for future goals, when there may not be the intention to initiate a deliberate retrieval. For example, if one were previously asked to pass a message to a colleague, nothing about seeing the colleague later in the day would necessarily prompt a deliberate retrieval to bring that task to mind, especially if the encounter is otherwise routine. It is in this context that IAM provides a functional benefit, and indeed, this is known as spontaneous retrieval in the prospective memory literature and is one of several possible strategies for achieving such a goal (McDaniel & Einstein, Reference McDaniel and Einstein2007). This application of rational analysis showcases how IAM can play a role in problem solving: Beyond the passive role suggested in the target article, people can learn to take advantage of IAM to reduce cognitive load.
We have implemented a spontaneous retrieval mechanism in a cognitive architecture, modeled its use in prospective memory, and shown that the conditions under which it succeeds qualitatively resemble results from human experiments (Li & Laird, Reference Li and Laird2015). More than that, the model requires a fully specified theory of how IAM arises and how it interacts with other memory processes such as deliberate retrieval; in our case, to prioritize problem solving, involuntary retrieval only occurs when no deliberate retrievals are taking place. This decision, which follows from the assumption that memory is used to support goal-driven behavior, suggests an explanation for why researchers have found that IAMs occur most commonly during “relaxed or non-focused state[s] of awareness” and how “being focused would inhibit the activation of knowledge units that are inconsistent with the individual's current goals” (Berntsen, Reference Berntsen and Mace2008; quoting Mandler, Reference Mandler, Umiltà and Moscovitch1994). In this case, the computational model built on the rational analysis framework aligns with the psychology literature.
As for déjà vu, although we know of no existing theory of its functional role in cognition, we agree that it results from false positives in familiarity judgments. Familiarity judgments – or at least its simplest form, recognition – have long been a subject of study via rational analysis, with recognition probability following the optimal Bayesian solution (Shiffrin & Steyvers, Reference Shiffrin and Steyvers1997). That familiarity is faster than recall allows it to be used to guide the strategic search for knowledge, as per the cognitive-heuristic account of metamemory (Schwartz & Metcalfe, Reference Schwartz and Metcalfe2011). We have implemented recognition judgments in a cognitive architecture and used it to trigger deliberate retrievals, which led to situations where a false-positive recognition resulted in retrieval failure (Li, Derbinsky, & Laird, Reference Li, Derbinsky and Laird2012). Although we have not modeled déjà vu explicitly, the retrieval failure could suggest that recognition was “implausible.” This interaction between deliberate retrieval and plausibility was not explored in the target article, nor how familiarity and recollection interact with each other over time.
In sum, the target article by B&M presents a compelling proposal for how IAM and déjà vu arise. However, their description is missing details that would clarify the relationships between memory mechanisms and could be improved by accounting for the functionality of these phenomena. Applying the rational analysis framework, and considering how the proposed system may be modeled computationally, would resolve these issues.
You have
Access
The target article by Barzykowski and Moulin (B&M) argues that involuntary autobiographical memory (IAM) and déjà vu are the result of an integrated system of memory and that they naturally arise from recognition and memory retrieval processes. While we agree with this stance, we find the proposed memory architecture lacking in two respects. First, the focus of the authors on the phenomenology of IAM and déjà vu neglects the functional role of these phenomena, the justification for why they might exist, and how they might be used by agents. Second, the memory architecture, as summarized in Figure 1, is only imprecisely specified, leaving room for alternate theories, potential inconsistencies, and omitted details. Here, we consider how performing a rational analysis of, and building computational models of, IAM and déjà vu can mitigate these problems. We use our work on how IAM can support prospective memory as an example of addressing both problems (Li & Laird, Reference Li and Laird2015), then extend that reasoning to the familiarity judgments that underlie déjà vu.
The rational analysis framework assumes that cognitive processes are optimally adapted to the functional goals of the agent, while subject to ecological constraints and limits on biological and cognitive resources (Anderson, Reference Anderson1990; Lieder & Griffiths, Reference Lieder and Griffiths2020). For memory, we take its primary function to be to “bring past experience to bear on present action” (Anderson, Reference Anderson1994), operating within a small working memory capacity, fixed bandwidth to long-term memory, and other cognitive constraints. This serves as the starting point for understanding the functional role of phenomena such as IAM.
A hypothesis about the structure of memory can be tested via its implementation in a computational model. These models force researchers to be precise in their definitions of the computational representations and processes that underlie their theories and ensure that hypothesized theoretical models of memory are consistent both internally and with broader theories of cognition. This is particularly true in cognitive architectures such as ACT-R (Anderson, Reference Anderson2007), which integrate multiple cognitive processes and potential neural correlates into a single system. This enables the evaluation of their combined performance across multiple tasks, thus ensuring that a hypothesis is compatible with the same mechanisms used to model other phenomena.
Consider, for example, the hypothesis that IAM is the result of automatic matching of sensory and abstract cues with items in memory in order to “quickly rais[e] pertinent information to consciousness without effort” (target article, sect. 4, para. 3). A rational analysis of IAM would start by considering the limits of deliberate retrieval and situations where those limits are exceeded. One such situation is in the prospective memory for future goals, when there may not be the intention to initiate a deliberate retrieval. For example, if one were previously asked to pass a message to a colleague, nothing about seeing the colleague later in the day would necessarily prompt a deliberate retrieval to bring that task to mind, especially if the encounter is otherwise routine. It is in this context that IAM provides a functional benefit, and indeed, this is known as spontaneous retrieval in the prospective memory literature and is one of several possible strategies for achieving such a goal (McDaniel & Einstein, Reference McDaniel and Einstein2007). This application of rational analysis showcases how IAM can play a role in problem solving: Beyond the passive role suggested in the target article, people can learn to take advantage of IAM to reduce cognitive load.
We have implemented a spontaneous retrieval mechanism in a cognitive architecture, modeled its use in prospective memory, and shown that the conditions under which it succeeds qualitatively resemble results from human experiments (Li & Laird, Reference Li and Laird2015). More than that, the model requires a fully specified theory of how IAM arises and how it interacts with other memory processes such as deliberate retrieval; in our case, to prioritize problem solving, involuntary retrieval only occurs when no deliberate retrievals are taking place. This decision, which follows from the assumption that memory is used to support goal-driven behavior, suggests an explanation for why researchers have found that IAMs occur most commonly during “relaxed or non-focused state[s] of awareness” and how “being focused would inhibit the activation of knowledge units that are inconsistent with the individual's current goals” (Berntsen, Reference Berntsen and Mace2008; quoting Mandler, Reference Mandler, Umiltà and Moscovitch1994). In this case, the computational model built on the rational analysis framework aligns with the psychology literature.
As for déjà vu, although we know of no existing theory of its functional role in cognition, we agree that it results from false positives in familiarity judgments. Familiarity judgments – or at least its simplest form, recognition – have long been a subject of study via rational analysis, with recognition probability following the optimal Bayesian solution (Shiffrin & Steyvers, Reference Shiffrin and Steyvers1997). That familiarity is faster than recall allows it to be used to guide the strategic search for knowledge, as per the cognitive-heuristic account of metamemory (Schwartz & Metcalfe, Reference Schwartz and Metcalfe2011). We have implemented recognition judgments in a cognitive architecture and used it to trigger deliberate retrievals, which led to situations where a false-positive recognition resulted in retrieval failure (Li, Derbinsky, & Laird, Reference Li, Derbinsky and Laird2012). Although we have not modeled déjà vu explicitly, the retrieval failure could suggest that recognition was “implausible.” This interaction between deliberate retrieval and plausibility was not explored in the target article, nor how familiarity and recollection interact with each other over time.
In sum, the target article by B&M presents a compelling proposal for how IAM and déjà vu arise. However, their description is missing details that would clarify the relationships between memory mechanisms and could be improved by accounting for the functionality of these phenomena. Applying the rational analysis framework, and considering how the proposed system may be modeled computationally, would resolve these issues.
Financial support
The authors did not receive any outside funding for this work.
Competing interest
None.