Memory is not an archive, but a filter. A study

Remembering is almost always regarded as an advantage: those who retain the past have access to more information, recognise what they have already encountered and can use that experience to find their bearings in the present. Memory thus appears as a storehouse in which a greater quantity of data should produce better knowledge – an idea that has become even more natural since we entrust it to photographs, conversations and documents, or devices whose quality is also measured by their storage capacity. Living organisms, however, face a different problem, because every memory must be inscribed within a physical structure that consumes resources, introduces errors and continues, for some time, to represent a world that may have changed in the meantime.

An animal searching for food by following its scent receives signals that are constantly distorted by the wind. At a certain point, the scent becomes strong; a few moments later it disappears, then reappears in a different direction; a single perception describes only the present situation, whilst a comparison with what was perceived shortly before allows a trend to be recognised. To make this comparison, it is necessary to retain a record of the past, by having a system that records it with sufficient precision and retains it for long enough to make it usable.

Even a cell, despite lacking a brain, constantly faces problems of the same kind. The concentration of an external substance fluctuates, molecular signals are disrupted by random fluctuations, and a complex response may only be worthwhile when the stimulus persists. The cell must therefore attribute meaning to the events it encounters over time, using biochemical circuits in which a preceding condition continues to influence the subsequent one for a few seconds or a few minutes. Biological memory begins right here, in the ability of a present physical state to retain some of the information received in the past.

Takehiro Tottori, of the RIKEN Centre for Brain Science, and Tetsuya J. Kobayashi, of the Institute of Industrial Science at the University of Tokyo, have sought to establish when this ability becomes advantageous. In their paper, published in Physical Review Letters, they describe a theoretical organism that must estimate a changing environmental condition using imperfect signals, with the option of relying solely on current perception or combining it with a record of previous observations. The system has limited resources, and memory comes at a cost, as the record must be updated and protected from fluctuations that tend to corrupt it.

The question therefore concerns the value of memory. The organism can minimise its errors by comparing the present with the past, provided that the stored information is reliable enough to justify the resources expended in maintaining it. When this condition is not met, memory adds to the noise from the environment that produced by its own internal mechanisms, perpetuating signals of little value and influencing decisions that would have been better had the past been allowed to fade away.

The first finding by the two researchers concerns the way in which the transition from one strategy to another takes place. When available resources are scarce, the theoretical organism achieves the best result by completely ignoring memory and reacting to the perception of the moment. Even a rudimentary memory requires an investment, whilst the trace it manages to retain remains too weak and noisy to adequately improve knowledge of the environment. The emergence of a small capacity for memory therefore does not necessarily yield a small advantage: it may result in an expenditure that lacks sufficient utility.

As resources are gradually increased, the system continues for a certain period to rely solely on the present; once a threshold is reached, the memory becomes stable enough to genuinely alter the quality of decisions, and the optimal strategy changes abruptly. The authors define this shift as a ‘phase transition’, using the language of physics to indicate a qualitative transformation brought about by exceeding a critical value. Below the threshold, the past is ignored; above the threshold, it becomes a fully integral part of the process by which the organism interprets its environment.

The reason for this discontinuity lies in the very structure of memory, which requires a mechanism capable of recording the signal and an internal state stable enough to retain it. A trace that is easily written and quickly erased disappears before it can guide behaviour; a very stable state that is poorly updated remains largely insensitive to what is happening. The advantage becomes apparent when recording and retention together achieve sufficient efficiency, allowing memory to correct errors in current perception. In the model, the intermediate solution may exist without becoming the optimal choice: a weak memory remains disadvantageous until the entire system begins to function in a coordinated manner.

This passage suggests a possible answer to a general evolutionary question. A complex function may depend on components which, when considered in isolation or present in a too rudimentary form, offer limited benefit; gradual changes may accumulate to produce a configuration whose biological value changes rapidly. The model does not reconstruct the historical emergence of memory, nor does it demonstrate that a particular species has crossed the threshold identified by the calculations; rather, it shows how a continuous variation in resources can generate a discontinuity in the most advantageous strategy. Natural selection would then have two different ways of processing information at its disposal, each advantageous in a different range of biological and environmental conditions.

Memory does not become useful in simple proportion to the uncertainty of the senses. When the signal from the environment is very precise, present-moment perception already contains almost everything that is needed, and memory adds little. If the signal is moderately uncertain, comparing successive observations allows us to distinguish a persistent variation from an occasional fluctuation, thereby improving our assessment of the situation. When uncertainty increases further and observations become unreliable, retaining them merely prolongs the noise, attributing a longer duration to data that do not allow us to reconstruct what is happening. Memory is thus particularly advantageous in the intermediate region, where information still bears a meaningful relationship to reality and immediate perception remains insufficient.

An animal following a scent trail derives little benefit from its memory when the scent clearly indicates the direction, as it need only react to the current signal. In conditions of moderate turbulence, previous perceptions make it possible to smooth out the current fluctuations and reconstruct the overall movement towards the source. If the wind disperses the scent in an almost random manner, a longer memory records a more extensive sequence of events lacking any usable order. The amount of past information available increases, whilst its ability to guide the animal remains minimal.

The value of memory also depends on the speed at which the environment changes. Previous information is only helpful as long as it remains relevant to the present situation; in a highly unstable world, memories become outdated rapidly and may cause the organism to respond to conditions that no longer exist. The model replicates the shift towards a memory-free strategy as environmental volatility increases, consistent in qualitative terms with experiments in which people reduce the weight they attribute to previous observations when the context becomes more changeable.

An animal that has repeatedly found prey in the same place may focus its subsequent searches there, saving time and energy; after the prey has migrated, the same memory continues to point to a location that has become unproductive. The trace retains its content intact but loses its value, because the link between past and present has been severed. Forgetting thus reduces the influence of old information and allows new signals to modify behaviour more rapidly.

This results in a very different conception of memory from that of a repository in which the benefit increases in line with the quantity of data accumulated. Tottori and Kobayashi’s model does not compare archives of different capacities, as it examines the advantages of using or ignoring an internal state that preserves the past; the general implication concerns the predictive quality of the content. A memory is valuable when its contents enable us to better assess the present or anticipate what is about to happen. A larger collection of signals that are degraded, random or relate to conditions that no longer exist merely increases the volume of stored material without improving our understanding.

This idea links memory to the selection of information. Retaining everything would require resources and would confer permanence even on that which arises by chance; an effective system is not measured by the quantity of traces it manages to retain, but by its ability to let go of that which no longer maintains a connection with the regularities of the world. Memory thus becomes an active filter, which does not accumulate indiscriminately but decides, through its own instability, what can still serve as a guide and what, on the other hand, must be discarded.

The optimal duration of a memory is not an intrinsic property of the memory itself, but emerges from its relationship with the persistence of the phenomena that generated it: a few seconds may be sufficient for a bacterium comparing successive concentrations, whilst more complex behaviours require greater persistence when the environment retains its characteristics for longer periods. It is in this gap between what remains and what changes that the value of memory is defined; this value does not lie in its duration but in its ability to remain synchronised with the world it seeks to describe.