The main point of this new Science paper (Lin et al., August 13, 2026) is surprisingly radical:
Long-term memory may be stored less in the persistence of particular strong synapses than in the architecture of a surviving pattern of connections.
The experiment gets at this unusually cleanly. The researchers induced an artificial hibernation-like state in mice. Hippocampal neuronal activity fell dramatically, and large numbers of dendritic spines and synapses disappeared. Yet when the animals recovered, the previously learned memories remained, as did the hippocampal neuronal representations associated with those memories.
What survived was not simply the traditionally favored large, stable dendritic spines. Instead, a particular subset of spines survived disproportionately: spines connected to multisynaptic boutons—presynaptic terminals that contact more than one postsynaptic element. The authors therefore propose that what matters is an engram architecture, a specific relational configuration among synapses, rather than some privileged individual synapse carrying the memory by itself.
I would reduce the finding to:
memory ≠ particular synapse
memory ≈ resilient topology of connections.
That distinction is important because neuroscience has long had a problem: if memories persist for years, how can they do so when proteins turn over, dendritic spines appear and disappear, and neuronal representations themselves “drift”? This paper offers one possible answer. The components can change while the organizational relation survives. Science News aptly summarizes the implication: preservation of every individual connection may not be necessary; the wider network architecture may preserve the memory.
And this is where I think the paper becomes especially interesting in relation to our recent discussions. It suggests something resembling a Ship of Theseus model of memory:
物可以換，關係可以重組，而某種 pattern 仍然在。
Not “this molecule remembers,” nor even “this neuron remembers,” but something closer to:
this configuration remains reproducible.
There is nevertheless an important qualification. The paper does not show that memories are abstract information completely independent of material substrate. Quite the opposite: there is still a physical trace. The novelty is that the physical trace seems to be architectural and distributed, particularly involving those preserved multisynaptic arrangements, rather than simply “big spine = memory.”
So I would formulate its deepest implication as:
The engram may be a relation before it is a thing.
And that produces an intriguing paradox in terms of your recurring question about self and memory: perhaps autobiographical continuity does not require the preservation of its constituent pieces. What must persist is merely enough relational architecture to regenerate the pattern.
In other words, 記憶不是保存碎片，而可能是保存「碎片如何重新找到彼此」的方法。
That, to me, is much more interesting than the artificial-hibernation result itself.
