The standard model of memory replay during sleep comes from rodent studies. Rats learn a maze. During subsequent sleep, their hippocampal place cells fire in the same sequence as during navigation — the same cells, in the same order, compressed into a fraction of the original time. The replay preserves the trajectory. The brain, it seemed, consolidates by re-running the tape.
Researchers at the University of Bonn recorded from 1,433 neurons in 17 epilepsy patients who had learned a visual story — a Fotonovela with sequential slides, each featuring a named character, a picture, and linking sentences. Patients practiced until they could reproduce the correct sequence six times. They then slept while their neurons were recorded. The study (bioRxiv 2026, Niediek, Bierling, et al.) identified 87 “concept cells” — neurons that fire selectively for specific semantic content regardless of presentation format. A neuron for Lois Griffin fires whether you see her picture or read her name.
During slow-wave sleep, these concept neurons reactivated together at rates above chance, particularly during sharp-wave ripples — the brief electrical bursts associated with memory consolidation. Content was preserved. Neurons representing characters from the same story co-fired more than neurons from different stories.
But when the researchers checked the order of firing, only 39% of reactivation sequences matched the story's temporal sequence — statistically indistinguishable from chance. The neurons remembered who but forgot when. The associations survived. The timeline didn't.
This is lossy compression with a specific compression strategy: discard the dimension most cheaply reconstructed from context, retain the dimension hardest to reconstruct. If you know the characters and their relationships, plausible orderings can be inferred. If you know only the order, you cannot reconstruct the characters. The hippocampus compresses by dropping time and keeping structure — the opposite of what the rodent replay model predicted.
The discrepancy between rats and humans matters. Rat place cells encode spatial sequences, and spatial sequence is intrinsic to the content — you cannot have a trajectory without an order. Human concept cells encode semantic associations, where sequence is incidental to the content. A story about Lois Griffin and Professor Elger has a temporal ordering, but the relationship between the characters exists independently of when each appeared. The hippocampus preserves what's intrinsic and discards what's incidental. It's not choosing content over sequence as a general rule. It's choosing the structural over the decorative, and which dimension is structural depends on what kind of information is being consolidated.
The compression principle extends beyond neuroscience. Version control systems store diffs, not snapshots, because temporal sequence is reconstructible from ordered diffs but the content of changes is not. Lossy audio codecs discard frequencies that fall below masking thresholds — sounds that are reconstructible by the auditory system from surrounding frequencies. In each case, the system discards whatever can be cheaply regenerated and preserves whatever cannot.
What the sleeping brain reveals is that this principle operates not at the level of bits, but at the level of meaning. The question it answers during consolidation is not “what order did things happen in?” but “what was connected to what?” The timeline is dropped because it can be rebuilt. The web of associations is kept because it can't.