Memories survived the loss of half the brain's synapses in hibernating mice

A team headed by the Okinawa Institute of Science and Technology put trained mice into artificial hibernation for two days and found their memories survived the temporary loss of more than half the brain's synaptic connections ^[1][2][5]. The result, published Aug. 13 in Science, cuts against the assumption that has organised decades of memory work: that a long-term memory is held in particular strengthened synapses with large, stable dendritic spines ^[3][4].

The experimental logic borrows a stress test from biology. Hibernating animals shrink their brains and pare down cell-to-cell connections during long stretches of low metabolism, apparently as an energy-saving measure, and yet alpine marmots and European ground squirrels emerge remembering food locations and members of their own species ^[15][13]. Rather than wait for winter, the group used a switch: in 2020 a University of Tsukuba team led by Takeshi Sakurai, a coauthor on the new paper, identified the circuitry that induces hibernation in mice ^[14]. Animals were first trained on contextual fear, a setting paired with a mild paw shock, and on the locations of sugar pellets in a maze, both hippocampal episodic tasks ^[9]. Then the switch went on. Pruning began within 30 minutes, and Tanaka says the degree of remodeling was much larger than the team expected ^[10]. That puts the onset inside roughly the first one percent of the two-day episode ^[17].

What tracked recall was not the individual strong synapse. According to the authors, small clusters of engram-to-engram synapses were protected through the widespread remodeling, and those preserved motifs are what allowed accurate recall and the rebuilding of functional networks afterwards ^[8][6]. "This topological architecture of the broader network seems to be more important" than sturdy individual connections, Tanaka told Live Science ^[7].

Tanaka is explicit about the boundary. He says the work does not argue against long-term potentiation, which many studies support as critical for forming new memories, only against LTP being the essential ingredient for keeping them ^[11]. The finding also lands in existing company: spines appear and disappear, engram representations drift within days, and memories remain retrievable after previously strengthened connections are disrupted ^[12].

The consequence for anyone who reads synapse counts as a proxy for memory: in this model, a greater-than-50-percent drop in synaptic connections coexisted with intact recall, so raw connection density is not by itself a sufficient readout of whether a memory is still there ^[16].

Three things to watch. First, whether the protected clusters can be identified before the pruning rather than inferred after it, and whether ablating them selectively abolishes the memory, which is the test that converts a correlation into a substrate. Second, whether the account holds beyond the hippocampus and beyond 48 hours, since episodic memories are thought to migrate out of the hippocampus over weeks to months for longer-term storage, a window this experiment did not probe ^[19]. Third, whether any of it transfers: the material covers mice, the hippocampus, and a two-day induced shutdown, and makes no claim about neurodegeneration or humans ^[18].