In a previous post we spoke of memory as a method for expressing intelligence. But is it possible to conceive of intelligence in a way that is not structured around the development, preservation, and employment of memory? As our team began considering this question, we first noticed the wide range of contexts within which the notion of memory is brought to bear. Here we share a few preliminary reflections on the nature of memory and the usage of the word. The article takes the form of a rudimentary general critique, aiming to make references to memory less opaque by acknowledging the constructive and creative character of this phenomenon. We begin by looking for common sense among humans, and from there seek to move toward a more transparent understanding of memory that would be equally applicable in both human and machine contexts.
A Simple Definition of Memory
People may talk about personal memories, cultural memories, sub- or unconscious memories, evolutionary memories stored in genes, memory as access to preserved data, environmental memories left as imprints in an ecosystem, etc. Since appeal to the idea of “memory” is so frequent and widespread, it is hard to say which of the word’s connotations might be the more straightforward and which the more abstract. Still, episodic memory of past events—as when for example we recall yesterday’s walk together—is a good candidate for a commonsensical understanding of “memory.” Something that happened in the past is nonetheless still kept in mind now.
Pure Recollection Does Not Occur
But if we accept that such a recollection of the past is possible—as indeed it would be hard not to, on at least some level—such an acceptance covers the fact that the use of memory is a constructive exercise, not the reflecting of past facts. If it were possible to fully remember the past event in all its rich specificity, we would literally be reliving the past when remembering. This would make memory time travel. Memory would enable us to enter into the past and, importantly, as the persons we were then. Otherwise we could not be said to have genuine access to the nature of the given past experience. If we possessed such a capacity for memory employment our lives would no longer be subject to the forward progression of time. Alternatively, if instead memory is about viewing the past from the perspective of our present self, or from a perspective that is by necessity different from the original one, then our recollections and memories can of course no longer be a mirror on the past, no matter how vivid, clear, and detailed they might otherwise be.
Memory Processing among Humans
This recognition that actual “re-minding” is impossible is easy to overlook, because it is so incontestable that it has the appearance of being already well-known and widely acknowledged. Indeed, the fact that the concept of “memory” is made to perform so many different functions within our intellectual landscapes can be seen as evidence of a generally shared acknowledgement that pure and mirror-like recollection does not occur. Looking at the ways we use the term, it seems we generally agree that memory entails acting constructively on what was before, and that such activity creates avenues for shaping the future. Knowledge is associated with processes of accumulation, and hence the storage of memory, and it is then compelling to think of knowledge accumulation as the compilation of data that should remain available for processing. But human minds, we may say, excel precisely at the extrapolation of meaning from myriads of discrete or loosely compounded data to which we neither can have, nor need to have, direct access. When we humans “remember” we partake of such processes of meaning construction out of incomprehensibly complex fields. Memories are in this way not only the meaningful building blocks of our individual stories, but also the stuff of collective stories—shared, represented, and reinforced through specific terms and symbols. All the while, the substratum for this process of meaning construction remains our physical genome and biology—itself the processed product of astronomical amounts of evolutionary memory.
Keeping Both Facts and Memory Meaningful
But if we allow ourselves to nevertheless dwell a bit on the impossibility of pure, mirror-like recollection there may still be interesting ramifications to notice. Consider this example. Humans, we may say, continually make stories from the history they lived. These are summaries, models of the past, that are developed with a view to predicting the future. When we say that such concise memories are imperfect, we might do so thinking in terms of their failure to capture all of the relevant facts. But it is then worthwhile to recall the just stated absurdities that follow from the notion of pure recollection of past facts. Keeping those in mind, we can conclude: If there were past facts they could not be recollected; if something is recollected now, it cannot be past facts. What then is memory and what does it mean to remember?
One way to answer those questions is along the following lines. When we disqualify a certain recollection or narrative by appeal to the notion that there are data that our story fails to capture, we may speak as if exhaustive data could, in principle, be available for recollection if only our cognitive capacities were powerful enough. Yet no intelligent power is capable of bringing the data of the past into the present. Such a feat would require turning into the past cognizer of that data, and such a transformation would not be memory, but something altogether different. Therefore, we may suggest, both the data we refer to and the idea of their recollection are themselves nothing more or less than meaningful constructs derived from collective memory processing. This is how both facts and their recollection get their traction—dynamically and in context. Not by virtue of an underlying transcendent reality that is so complex that it remains inaccessible to our limited cognitive capacities.
Such an understanding of memory and reality may come across as anthropocentric, or perhaps even indicative of a solipsistic orientation. But since both facts and meanings require communities for their construction, qualms about solipsism are obviously unfounded. As for anthropocentrism, there is no reason to limit the social contexts within which meaning emerges to human society alone. On the contrary, understanding both facts and their temporal processing as inextricably symbolic and meaning-laden goes well in hand with the idea of evolutionary memory that is accumulated across biological species and within the physical universe.
What Then Might Electronic Data Storage Be?
The idea of recollection of stored data is in this way problematic, if not plain absurd, when understood in a literal sense. At the same time, the construction and preservation of memory are functions that are crucial for intelligent survival in the world that we know. For those of us not actively engaged in AI engineering or the maintenance of networks, it may seem odd to straightforwardly propose, in the context of electronic data storage, consequences that are similar to those we would otherwise draw with the regard to the storage and accessing of human memories. Superhuman access to truly vast data is, after all, both the hallmark and great promise of machine intelligence. But if we go along with an understanding of data and memory as emergent, collective constructs, that sort of understanding must be as relevant to artificial intelligent agents as it would be to biological intelligences, such as humans. It is hard to think of a reason that AI memory should be qualitatively different in this regard. Indeed, our point above was precisely that the idea of actual recollection of past facts remains unintelligible regardless of the agent’s capacity.
The consequences of such an alignment of the workings of AI and human intelligence are then there to be explored. At CSAS we aim to do so, and we expect and invite others do so as well. For now, a simple conclusion we may want to draw is that unintended consequences appear to be ineliminable. If the frameworks meant to circumscribe the unfolding of human intelligence and AI are made of meaning-transmitting, collective constructs there is, in the end, just as little reason to expect perfect predictability in terms of AI as there is with respect to good old, human consciousness. In a follow-up post to this, we will then present a Buddhist perspective on memory and its relation to intelligence.