What makes memories last—dynamic ensembles or static synapses?

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses?

Both. One problem is that we conceive of memory as something the external world imposes on the brain, but the reality is more bidirectional—an interdependence exists between the bona fide stimuli and how the brain is organized to receive those stimuli. In many ways, the brain is projecting something akin to beliefs onto the stimuli, in order to process and store the experience. Another facet of this problem is the false dichotomy of the question. Information is the product of a dynamical process. It is not stored in any single element, and the storage is intimately related to the readout. Endel Tulving formalized this concept with the term ecphory (originally from Richard Semon), meaning memory performance (mem) ≈ I*C*M, where I is stored information, C is the cues that trigger the recall, and M is mindset.

In a dynamical system, it may not be practically possible to separate the process of storage from the access. For example, many of the same hippocampal cells seem to be engaged in both the encoding and recall of information in memory. Consider CA1 hippocampal place cells, which have at least two major anatomically distinct inputs, one from CA3 via the synapses at the Schaffer collateral terminals, and the other from entorhinal cortex (EC) via the synapses at the temporoammonic pathway terminals. The CA1 population discharge tends to represent the current location, as expected for place cells. But a study my group published in 2018 showed that on the infrequent occasions that the CA3 inputs dominate the EC inputs, the CA1 population discharge represents locations remote from the current location, such as where the mouse recollects being shocked. (In the experiment, the mouse will actively avoid this place for the next two seconds.) The place information for the entire environment was presumably stored in a distributed set of CA1 synapses accessible via presynaptic activity that could originate from CA3 or from EC, each activating different subsets of CA1 synapses.

The fact that CA1 output represents remote or current places depending on the relative CA3 and EC input suggests both the synapses and the input activity contribute to the output representation. The information stored in synapses is analogous to how an artificial neural network stores information in a covariance matrix that describes the interactions among its elements. Activity of the network can generate a version of the covariance matrix to encode or read out information. The information being stored and the circuit that is operating on the information are strongly interdependent.

We all know that biological processes operate and interact simultaneously across different levels of biological organization. But because experiments that span many levels are hard (but not impossible) to do, we tend to dichotomize instead of integrate. To better understand neurobiological systems, the field has to learn to rigorously integrate across levels with both data collection and data analysis.

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts?

Yes. The population of synapses that defines a circuit can reliably generate a particular and reliable output pattern from variable inputs. Think of the population of 100 recorded cells as creating a state space where the activity of each cell is a dimension of the 100-dimensional state space. If the activity of some cells is correlated, then they don’t signal information independently. Dimensionality reduction techniques, such as principal component analysis, can transform variable high-dimensional inputs into stable outputs by projecting them into one or more lower-dimension subspaces, where each subspace dimension is a composite of many correlated cells.

This video from a 2023 paper in Cell Reports offers a more intuitive visualization of how different patterns of activity can produce the same meaning, so long as the patterns are similar in the relevant subspace. Conversely, similar population activity patterns can generate distinct representations if they pass through distinct synaptic networks, each emphasizing different correlation patterns and thus distinct subspaces. Furthermore, drift in a high-dimensional space need not produce a read-out problem, because the high-dimensional space is so large that two drifting population patterns will only rarely come close by chance. This is analogous to two families in one apartment building. The families remain distinct even if the individuals move about each apartment.

What experimental data would be helpful to reconcile these observations to help bring these theories together?

Simultaneous neural ensemble activity recordings during the manipulations of so-called engram cell recordings will go a long way toward clarifying the apparent contradiction. We have made such recordings and have a preliminary report on bioRxiv. (We are adding a place cell study to the work and will then send it out for review.) Researchers have generally assumed that in a dynamical system, such as the CA1 network, stimulated neurons will reliably respond to the stimulation and not to the dictates of the network. But we found this assumption is wrong—the whole network of cells responds to the stimulation, whether or not specific cells are directly stimulated. This is similar to how collective behavior is organized in interacting populations of individuals, such as flocks of birds and schools of fish. (For more on this, see my review, “Remapping revisited: How the hippocampus represents different spaces,” in Nature Reviews Neuroscience.)

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses? 

The answer here is likely “yes,” in that cellular changes that enable memory storage are likely to occur not only at individual synapses but also at the level of individual neurons whose cell-biological features may change as memories are stored. As an example, it might be that changes in gene expression lead to changes in activity levels, although at the moment we really don’t understand the scope of these changes. 

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts? 

The answer here is also “yes,” but to get there we first need to step back and think about what it means to store and then retrieve a memory. Focusing on memories for the events of daily life, our current conception is that the events themselves drive activity across the brain, engaging specific neurons whose activity represents the various sights, sounds, smells and feelings that are part of the experience. These activity patterns converge on the hippocampus, a brain structure critical for the formation of memories of events, and rapid changes in synapses (and perhaps other cellular features) within the hippocampus modify hippocampal activity and create a new pattern associated with the event.

This first stage of memory formation must also be accompanied by a second stage in which the hippocampal pattern is itself associated with activity outside the hippocampus. This association, which presumably also involves synaptic plasticity, would enable the hippocampal pattern to “retrieve” the memory by reactivating neurons outside the hippocampus that encode aspects of the original experience. The key point here is that this loop of associations, from information processed outside the hippocampus to a hippocampal index and back again to the rest of the brain, is what makes memory formation and subsequent retrieval possible.

And now we can return to the question. If the hippocampal and extra-hippocampal patterns evolve together and remain connected, retrieval remains possible. The details of what is retrieved might themselves evolve as the extra-hippocampal patterns change, a feature that might help us understand the malleability of memory over time.

What experimental data would be helpful to reconcile these observations to help bring these theories together?

At a circuit level, the key prediction of the framework described above is that drift occurs concurrently across regions to maintain the links that enable retrieval. Direct measurements of cross-regional drift would be critical here. Moreover, we’d expect these changes to be happening most prominently in “offline” periods, including quiet wakefulness and sleep. If so, we need to make these measurements continuously over long time periods to be able to understand what is happening to these representations and when it happens.

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses?

Yes. First, I agree with the premise: Information is stored in the brain at some level(s). It’s worth noting that not everyone agrees with this. Information can be “stored” at multiple levels: synapses, cells, circuits, the body. It can even be fully externalized, back into the environment, as cues or changes we create, as complex as photos and writing and as simple as a reminder string around a finger. All can involve information in service of memory. I would submit that much of the heavy lifting is done at both the synaptic and circuit/ensemble level. Which levels dominate depends on factors such as memory type, when information was acquired and how it is integrated with the existing structures, themselves reflecting changes from earlier experiences. Granted, our methods have been biased to evaluate changes at the synaptic level, or at the level of cellular changes as proxy for synaptic changes.

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these memory engram drifts?

Why not? Whoever said memory needs to be supported by a single mechanism or at a single level? That said, we may need to be careful in using the term “these memories” or “these memory engrams.” Such terms suggest that experience creates biological bins to hold discrete memories, that memories exist as entities that are created “de novo,” and that neural modifications must reside at only one level, all of which are positions that are not or may not be true.

I don’t think a coexistence is any more troublesome than other accounts of brain function that cross levels of scale. For example, we accept that synapses may exhibit a variety of short-term plasticity mechanisms such as facilitation and adaptation. This doesn’t preclude long-term synaptic modifications. Rather, the two coexist. For example, changes in postsynaptic receptors alter the strength of the synapse over longer time frames. Although this effect could, in principle, eclipse any short-term modifications, it doesn’t necessarily have to. The two processes could be superposed, whereby short-term effects could ride on top of the new baseline created by the latter.

Similarly, groups of cells may become newly recruited to become co-active, and some of the original “cast members” rotated off, with the net effect of preserving a pattern that arose from previous experience. As this happens, the change would supersede the effects of strengthened synapses of the original cast members. So the notion of a fixed “engram” over time increasingly misses the true target.

Importantly, certain kinds of systematic, structured change in the members of an assembly can nevertheless preserve the decodablility, or identity, of a pattern. This would allow the organism to respond adaptively, based on knowledge (synthesized memories) accrued over time. Furthermore, fluctuations among constituent members of the many patterns needed for structured memories could help reallocate those cells best suited physiologically to contribute to new patterns, by criteria such as current excitability, clustered synapses or homeostatic plasticity, for example.

In the extreme, this reallocation may even involve a gradual change in the dominant “ownership” realized across whole brain regions. On one hand, this notion is highly speculative, but on the other hand, considering the long-standing evidence for memory-guided responses that arise from wholly separate brain regions and circuits over time, I would suggest the question is “how” and not “whether” such dynamics are in place. The time is ripe for modeling the conditions necessary for maintaining patterns under conditions of drift, to balance increased capacity and generalization with energy efficiency and to use empirical investigation to inform the biological constraints that need to be in place.

What experimental data would be helpful to reconcile these observations to help bring these theories together?

I’m not sure the goal is to bring them together, as they may coexist (see above). I’ll show my bias: We know substantially less at the ensemble level than at the IEG activity-tagged or synaptic plasticity level. Even the long-standing replay literature—the OG, time-resolved “engrams,” if you like—has evaluated cell assemblies using methods that work by assuming mostly stationary membership. To confidently reject ideas around the “ensemble doctrine” for memories requires measuring ensembles and constituent members’ activities over time for multiple different remembered experiences.

One prediction from an evolutionary lens is that species that live longer and have larger, more complex cortices may benefit disproportionately from the allocation of patterns at the “fluid ensemble” level. One possibility is that phenomena associated with dynamic ensembles, such as “representational drift,” have conserved topological patterns that support long-term flexible memory performance. Whether that’s the case remains to be seen.

TL;DR—One promising directive: Measure more manifolds of monkey memories.

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses?

I consider that information is stored at the systems level, and that it consists of both assemblies of cells and the synapses that connect them. It is, however, too early to say that those cells and synapses are sites of stored memory per se, as they may simply function to gain access to the memory. In addition, one should consider the existence of functionally distinct active ensembles within the engram circuits. Finally, it is also possible that there aren’t specific sites for memory storage; cells and synapses may be part of a brain-wide code for memory expression.

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts? 

To reconcile these observations, it is important to understand the biological meanings behind discrete engram circuits and their drifts. It is perhaps easier to reconcile if we consider that the “engram” circuits identified by current technology are merely vessels to fetch memories.

What experimental data would be helpful to reconcile these observations to help bring these theories together? 

Experiments addressing why the “engram” neurons drift and how they drift are much needed. To start with, I would suggest thorough monitoring of the drifting “engram” neurons, coupled with careful examination of their synaptic connectivity.

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses? 

Synapses in circuits! The field has held synaptic plasticity up as the main mechanism for information storage in the brain for several decades now, and I haven’t heard any good reasons to start doubting it yet.

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts?

Yes. We know from the work of Eve Marder and others, and more recently from studying deep neural networks, that degeneracy is likely to be ubiquitous in the brain. Many, many different cellular configurations can achieve the same circuit function. If and how brains evolved to use this degeneracy to maintain long-term memories is unknown. But it seems both possible and extremely useful.

What experimental data would be helpful to reconcile these observations to help bring these theories together?

The dream experiment would be to track the strengths of many thousands of synapses and record the electrical activity and gene expression of their associated pre- and postsynaptic neurons, longitudinally in vivo, before and during encoding of a memory in a living animal and then again later during memory recall a few days or weeks later. This will, of course, need to be done in a brain circuit that we know stores an essential component of the engram underlying the initial memory encoding. These data would let us ask if synaptic changes correlate with neural activity changes, and in turn ask if these correlate with any changes in the behavioral expression of memory. As a humble theorist, I will naturally be leaving these straightforward experiments as an exercise for my wet-lab colleagues.

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses?

Both. For the brain to learn new information, there needs to be a history-dependent change in its internal properties that ultimately manifests in behavioral changes. This means that neural circuits alter their properties, and we know for sure that synaptic changes are involved in many situations. Even if all the changes were concentrated at synapses (and we know that isn’t always the case), the “information” is distributed because it can be accessed only through the aggregate dynamics of a circuit.

Why does this question persist? Perhaps due to some of the most influential studies in the field, most obviously Eric Kandel’s work on the gill withdrawal reflex in Aplysia. This work is famous for identifying changes at a single synapse that underlie habituation—i.e., a change in behavior. This finding was possible because it exploited an unusually narrow bottleneck in the architecture of a specific circuit. So, we must pay attention to the fact that what made this circuit experimentally tractable is the same thing that makes it an outlier among neural circuits generally. Therefore, although some features of this work generalize—for example, some of the biochemistry involved—we shouldn’t expect such peculiar wiring in other circuits and species, and such a precise relationship between synaptic efficacy and behavior.

This is a paradox of “model systems”: By selecting models that permit manipulations that are impossible in general, we can end up with insights that can’t possibly generalize.

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts?

Yes. Let’s remind ourselves of the granularity in the evidence here. The classic engram experiments demonstrate broad correlations between behavior and reactivation of subpopulations of “engram” neurons that are tagged during some kind of conditioned learning. This could be freezing behavior, triggered by the association of a foot shock with a visual cue. Is there some degree of variability in the behavior during engram reactivation? Yes. Is there some degree of variability in the state of the network during engram reactivation? Undoubtedly. Do we believe that the “memory” is recapitulated exactly as if the animal is remembering the experience? I don’t think we do. For one thing, the means of reactivating engram cells produces activity patterns that the brain never generates endogenously. But as long as there is sufficient overlap in the stimulated population and any vestige of a “true” engram, the behavior can be biased toward the conditioned behavior.

People may quibble about details here, but the point I think most will agree on is that identifying and manipulating engrams is a statistical, population-level kind of analysis—and this is where drift comes in. At the population level, a circuit can drift significantly and still allow binary discrimination between encoded associations, for example. This is the fundamental fact that reconciles these observations, and it has been pointed out many times in the drift literature.

What experimental data would be helpful to reconcile these observations to help bring these theories together?

We already have them. In the drift literature, multiple analyses show that associations can be decoded far above chance in a drifting population, and even quantitative predictions of the animal’s behavior, such as its running trajectory through a maze, can be decoded despite drift. And let’s keep in mind that this decoding is done with a small subset of the neurons in any given circuit and with fairly crude statistical models. The brain can probably do better.

More data will help, of course. Personally, I would like to see much longer-term studies of neural representations in multiple neural circuits, so we can track engrams as they evolve and track the co-evolution of different circuits. This will let us tease out components that are related to behavioral change and understand how much the brain evolves in unison.

So, in summary: Drift happens. Not all circuits drift to the same extent, but even in those that do, there is overwhelming evidence that behaviorally relevant information is preserved over behaviorally relevant timescales. That’s all that is required of the biology. What we expect to see in an experiment is another thing and is often as much a product of our textbook paradigms as it is a reflection of reality.

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses?

It depends on what we mean by “information.” In many discussions, “information” simply refers to plasticity or a correlation of neuronal activity with some stimulus presentation. In that sense, a memory could be said to be stored at multiple levels, but what we would really be saying is that biological activity at those levels enables memory function. However, when I am pursuing the question of “information,” I want to know the essential level at which information that an animal knows about the world is stored as an engram. It seems to me that the plausible level for the storage of long-term memories is in the topography of the connectome. So, the information is engraved through stable changes in the brain’s microanatomical circuit. Although specific molecules, organelles, synapses and neurons are certainly all involved in the formation of such an engram, they may be dispensable for the retention and coding of the information itself.

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts?

I don’t see these two things as conflicting observations. On the one hand, we have a theory of engrams that has been tested (and so validated) by behavioral experiments in which engram ensembles are manipulated. On the other, we have a set of observations of ensemble activity and flux over time, popularly referred to as “representational drift.” A clearer question might be to ask how a comprehensive engram theory can incorporate the facts of what we call “drift.” How can this extend our understanding of engrams? And what predictions would such a theory make? 

What experimental data would be helpful to reconcile these observations to help bring these theories together?

The obvious question is, “Where do these (drifting) representations come from?” And, “What happened to the original representation?” We already know, from engram experiments, that the original representation is still present and functional in the brain, but inactive. Currently, we know less about the origin of the drifting representations. Were they there at the time of learning, but inactive? Do they represent degeneracy or redundancy in how engrams are formed? Or are they simply newer engrams that were learned as the animal has progressed through life? In a sense, representational drift may simply be a name we give the phenomenology of different engrams competing for expression in the brain.

Is information stored in the brain at the level of cells (or circuits) or at the level of synapses? 

I actually slightly disagree with the premise of the question—I’m convinced by some of the evidence on all of these scales, so I don’t see the question so much as identifying which scale is correct, as I see it as trying to understand how the evidence on these scales fits together. I find the evidence pretty convincing that activity in many parts of the brain, including canonical learning and memory centers such as the hippocampus, change over time. As a consequence, I don’t think information can be stored in cells or synapses in the hippocampus in a way that is stable over a lifetime. In other parts of the brain, this may not be the case. But I think that for these circuits that drift, we’re forced to think about how the brain can remember things despite the lack of permanent engrams. To me, this comes down to a question about compensatory learning rules. If the brain is constantly updating memories to account for drift, that could be a more robust solution that allows for ongoing learning without corrupting old memories. 

Can we reconcile observations that show distinct engram circuits seem to store memories versus observations that show the neuronal activity of these engram neurons drifts?

I think it’s possible that the apparent conflict between observations of engrams and drift might be due to issues of scale and subtleties in experimental design. On a timescale of a few days, memories seem pretty stable. On a timescale of a few weeks, there’s less evidence for stability. The degree of stability also seems to depend on factors such as how many times the animal experiences the same stimulus, so in that sense I think the apparent conflict may also depend in part on experimental design.

What experimental data would be helpful to reconcile these observations to help bring these theories together?

I think standardizing experimental paradigms across labs that are seeing different things would be helpful. If the discrepancies are really an issue of timescale or the details of experimental design, having two labs agree on the details and run a side-by-side comparison could help. Even if we were to find that evidence for stability or drift depends on the details, that alone wouldn’t bring the theories together, but it would provide useful constraints for computational work to try to unify the theories.