The replay

A rat is asleep in a corner of the recording room. The amplifier is still streaming, the LFP rolls slowly, the cells have gone quiet. Then, without warning, a fifty-millisecond burst of high-frequency oscillation rides up out of the noise, and in those fifty milliseconds the place cells that ran the maze ten minutes ago fire again, in the same order they fired during the run, only twenty times faster. The animal does not move. The brain has just rehearsed the morning’s experience.

This is replay, and it is the central phenomenon of the rest of this site. Fragments of recent waking activity are reproduced inside short, stereotyped LFP events during quiet states — sleep, drowsy stillness, the moments between trials. Replay was first reported by Wilson and McNaughton in 1994, sharpened eight years later by Lee and Wilson, and has held up across hundreds of follow-up studies. The events compress real waking sequences into a fraction of their original duration, preserve the order, and recur thousands of times per night.

The packaging matters as much as the content. Replay rides on top of a particular LFP signature called a sharp-wave ripple — SWR for short — produced when tens of thousands of CA1 neurons synchronize for a brief, structured burst. Pull a wire into stratum pyramidale during slow-wave sleep and you will hear, alongside the slow drone of the cortex, a sequence of discrete electrical pops about once a second. Each pop is one SWR. Each one is, by current best evidence, the brain rehearsing something.

The two-stage memory model

The reason replay deserves a chapter on its own is the framework around it. The hippocampus is famously good at fast learning — episodic memory is a single-shot affair — but it is also famously small, and the experiences it can hold are limited to the recent past. The cortex, by contrast, learns slowly, holds vast amounts, and resists forgetting. The two-stage memory model, due originally to Marr in 1971 and developed by Buzsáki, McClelland, McNaughton and O’Reilly through the 1990s, proposes that this division of labor is solved by replay: the hippocampus learns once during waking, then quietly rehearses the trace during quiet states, allowing the cortex to integrate the new information across many low-rate exposures. Without replay, in the model, you cannot safely turn a hippocampal one-shot into a cortical long-term memory.

The model also predicts that disrupting SWRs during post-task sleep should impair memory the next day. That prediction has been tested and held. Girardeau et al. (2009), Ego-Stengel and Wilson the same year, and several follow-ups all showed that closed-loop electrical or optogenetic shutdown of SWRs during sleep degrades subsequent spatial behavior. Replay is not just suspicious correlation; it is causally involved in the consolidation of recent experience.

What compression looks like

Two states, the same eight neurons, the same behavioural sequence. Watch the cells fire in order across four seconds of running, then watch the entire pattern reappear inside a single sharp-wave ripple lasting under a fifth of a second. Press Play to sweep the timeline.

Two things are worth noticing. First, the order is preserved — the cell that fired earliest in the run also fires earliest in the replay. The replay is not just a burst of activity; it is a sequence. Second, the time compression is real and dramatic. A traversal that took several seconds in waking is rehearsed in roughly a hundred and fifty milliseconds. The factor is roughly twentyfold across most of the literature, though the exact compression depends on the cell, the run speed, and whether the replay runs forward or backward. Yes — Foster and Wilson (2006) showed that some replays run in reverse, the run as if rewound, which is its own delightful complication and probably tells us something about reward-driven evaluation.

Two facts that do all the work

Two properties of replay drive the rest of the experiments in this site, and stacking them is what makes the design possible:

(a) Different memories produce different replay patterns. A rat that has learned two arenas will, during a single sleep, replay both. A decoder built on the waking place maps from each room can usually tell which arena a given event is rehearsing. Remapping (Chapter 3) is what makes the patterns distinguishable in the first place.

(b) Disrupting replay impairs memory. A laser flash through ArchT-expressing pyramidal cells during a detected SWR will shut that SWR down within milliseconds, and the impact on next-day spatial behavior is robust across the literature.

Stack (a) and (b) and you get the experiment Igor’s PhD was built around. Train a rat on two arenas. Watch the post-task sleep. For each detected replay event, decide in real time which arena it represents. Disrupt only the events that look like Room A; let Room B replays pass untouched. If everything works, the animal should later remember Room B normally and search badly in Room A — same neurons, same animal, same night, same brain chemistry, two different memories with two different fates. The control is the rat’s own head.

The math, the surgery, the decoder, the sequence preservation, the SWRs, the two-room paradigm — all of it has been pointing toward this single experimental move. From here on, the chapters are about getting it to work.

The first complication is one of timing. By the moment the LFP shows you what we just rendered above as a clean sharp-wave ripple, several tens of milliseconds of replay have already elapsed; the burst you are reacting to is mostly already over. If you want to interrupt a replay while it is happening, the SWR signature itself is too late. Igor used a different signal entirely — and that is the next chapter.