AI Never Sleeps. That's the Problem.
A few weeks ago, I was mid-sentence at dinner — passionately defending a position I had quietly abandoned three weeks earlier — when I caught myself. The argument was running without me. The words were right, the structure was intact, the rhetorical moves were in their proper order. But the conviction had long since left the building. What remained was the groove.
That experience has stayed with me. Because it points to something genuinely strange about memory: not all learning ends up in the same place. Some knowledge stays close to the surface, conscious and revisable. And some gets pressed so deep into the architecture that it runs without you — below the waterline of deliberate recall. The question of how those grooves form, and what determines which experiences become load-bearing structures rather than passing notes, turns out to be one of the most fascinating unsolved problems in neuroscience. And it turns out the answer happens largely while you're unconscious.
Which means AI — which never sleeps — may be missing a trick so fundamental we've barely begun to model it.
The Nomination Ceremony
Here's what we now know happens in your brain when you learn something new. The hippocampus, that seahorse-shaped structure buried in the temporal lobe, acts as a rapid, high-fidelity recorder. It encodes the day's new experiences quickly, promiscuously, laying down fresh traces without immediately disturbing the stable, slower-to-form knowledge sitting in the neocortex. This separation of fast and slow storage is the "two-stage" learning model, a framework supported by decades of lesion studies and now increasingly validated at the cellular level (Staresina et al., 2025).
But here's what's been harder to see until recently: the selection process. Not everything you experience today will get consolidated into stable long-term memory. Something has to decide what's worth keeping.
New research published in Nature Communications provides one of the most detailed views yet of that decision-making process (Nature Communications, 2025). Using high-density recording in rodents, the study identified what the researchers call "engram-to-be" cells — neurons that form coordinated ensembles during a learning session and then show correlated reactivation specifically during post-learning sleep. Not during pre-learning sleep. Not randomly. These particular cell assemblies, marked during waking experience, are selectively reactivated during N2 and REM sleep stages, and that reactivation is causally linked to whether the memory consolidates.
Think about what this means. During learning, the brain is not just passively recording. It is nominating specific neural ensembles for later processing — flagging them as candidates for long-term storage. The actual consolidation happens later, offline, during sleep. The waking brain is like a film crew capturing footage; the sleeping brain is the editing suite where the story gets assembled.
The Clockwork
What does the editing look like mechanically? The answer involves a level of orchestrated complexity that would seem implausible if it weren't so well-documented.
During NREM sleep, three oscillatory events coordinate in a precise temporal cascade: hippocampal sharp-wave ripples, cortical slow-oscillations, and thalamocortical spindles (Staresina et al., 2025). The sharp-wave ripples are rapid bursts of hippocampal activity — the replay itself, memory traces firing in compressed form. The slow oscillations are large, sweeping waves across the cortex. The spindles are rapid bursts of thalamic activity that seem to serve as a gating mechanism for synaptic change. When all three nest together in the right sequence — ripple within spindle, spindle on the up-phase of the slow oscillation — the conditions for memory transfer from hippocampus to neocortex are met.
Neuromodulators control the switching between encoding and consolidation modes. High acetylcholine during waking biases the brain toward taking in new information; falling acetylcholine during NREM sleep releases the consolidation process. It's a chemical valve — open for input during the day, turned toward integration at night (Staresina et al., 2025).
The sophistication of this is almost uncomfortable to sit with. This isn't a single mechanism. It's a multi-stage, multi-system process, coordinated across brain regions by oscillations and neuromodulatory chemistry, that reliably converts the day's selected experiences into the architecture of long-term memory. And it has been running every night, in every mammal, for hundreds of millions of years.
The Problem With Machines That Never Sleep
Neural networks learn differently. They learn by gradient descent — adjusting their weights incrementally in response to training examples, nudging the error surface downward. It works remarkably well, at scale, for static tasks. But it has a profound vulnerability: when you train a neural network on a new task after an old one, it tends to forget the old one. Catastrophically. The weights that encoded the previous task are overwritten by the gradient updates for the new one.
This is not a minor inconvenience. It's a fundamental architectural difference from how biological brains work. A child who learns to ride a bike and then learns to read does not forget how to ride a bike. A neural network trained sequentially on two image classification tasks typically degrades badly on the first after learning the second.
The biological solution is sleep-mediated replay: the hippocampus re-runs old memories during consolidation, interleaving them with new learning so that the neocortex can integrate both without sacrificing either (Staresina et al., 2025). The old is protected while the new is absorbed. Every night.
In 2022, a research team took this insight and transplanted it directly into an artificial neural network. Tadros et al. implemented a "Sleep Replay Consolidation" (SRC) algorithm modeled on two core features of biological sleep consolidation: spontaneous unsupervised replay of stored memory traces, and local synaptic plasticity (Tadros et al., 2022). After each new task, the network enters a sleep-like phase in which it replays both new and old experiences without external supervision. The result: genuine continual learning across multiple sequential tasks, where catastrophic forgetting would otherwise have erased everything.
It worked. The sleep-like phase protected old memories while absorbing new ones, in a meaningful computational analog to what the biological brain achieves every night. This is one of those results that feels almost too clean — here is the problem, here is the biological solution, here is the implementation, here is the evidence that it transfers.
And yet no large-scale deployed language or vision model has a proper offline consolidation cycle. The SRC algorithm is a proof of concept. The frontier models that shape the current AI landscape don't sleep.
The Daytime Half
Here's something I find even more interesting: replay doesn't only happen during sleep.
Papale and Buffalo (2025), writing in Trends in Neurosciences, examine the accumulating evidence for awake hippocampal replay — the spontaneous reactivation of neural activity patterns during waking rest periods. They identify two distinct functional roles. The first is real-time decision support: forward replay, in which the hippocampus rapidly simulates future trajectories before a decision, running compressed mental simulations of potential paths. The second is memory tagging: shortly after a salient experience is encoded, the hippocampus reactivates it, creating what Papale and Buffalo describe as a "latent excitable state" — a priority flag that marks those memories for deeper consolidation later during sleep.
So replay is not a nocturnal event so much as a continuous, goal-sensitive process that runs throughout the waking day and then accelerates during sleep. The brain is constantly revisiting its recent experiences, evaluating them, flagging some for deeper processing. Awake replay is not background noise; it's an active, computationally meaningful system for shaping what gets remembered (Papale and Buffalo, 2025).
The closest AI analog is experience replay in reinforcement learning — the technique used in DeepMind's DQN, where past transitions are randomly re-sampled from a memory buffer to stabilize learning. But the biological awake replay isn't random. It's goal-sensitive and salience-weighted. The hippocampus prioritizes experiences that were surprising, consequential, or relevant to current goals. No current reinforcement learning system does anything like that kind of selective, value-guided replay.
Why Nobody's Built It
I've been sitting with a question for a while: why, despite compelling proofs of concept like SRC, despite growing mechanistic understanding of the biology, has nobody built a genuinely sleep-like consolidation system for large AI models?
Part of the answer is practical. Sleep-like consolidation requires a dedicated offline phase, and the economic incentives in AI currently favor models that train once and infer continuously. A model that needed to "sleep" between updates would be awkward to deploy at scale.
Part of the answer is architectural. The two-stage model works in biology because the hippocampus and neocortex are genuinely different systems with different learning rates and different representational structures. Standard neural networks don't have this separation. Building a system that learns fast locally and then slowly integrates into stable distributed representations is not an algorithmic fix you bolt onto an existing design — it requires rethinking the architecture from the ground up.
And part of the answer is genuinely epistemic: we don't know exactly what the sleeping brain is optimizing for. We know it consolidates memory. We know it involves selective replay. We know the oscillatory choreography — sharp-wave ripples, spindles, slow oscillations — is precise and non-random (Staresina et al., 2025). The system that nominates engram-to-be cells during waking and then reactivates them during sleep is doing something causally important (Nature Communications, 2025). But the objective function, if there is one, remains opaque in ways that make clean computational translation difficult.
What the recent evidence does make clear is that the engram nomination process, the oscillatory consolidation cascade, and the awake replay system are all parts of a single integrated memory architecture that spans the full twenty-four-hour cycle. Sleep isn't a pause in learning. It's the other half.
Every night, your brain runs a process that has no good analog in any deployed AI system: a selective, offline consolidation cycle that reviews the day's experiences, determines which ones are worth keeping, interleaves them with prior knowledge, and gradually transfers them into stable long-term representations. It does this through coordinated neural oscillations, neuromodulatory chemistry, and spontaneous cellular replay — a system of breathtaking intricacy that evolution has refined over hundreds of millions of years.
We have barely begun to build anything like it.
Which raises the question I keep returning to: if the brain's most powerful learning algorithm runs while we're unconscious, what does it mean that we've spent forty years building machines that never are?
References
- Papale and Buffalo (2025). Awake Replay: Off the Clock but on the Job. https://www.sciencedirect.com/science/article/pii/S0166223625000372
- Staresina et al. (2025). Systems Memory Consolidation During Sleep: Oscillations, Neuromodulators, and Synaptic Remodeling. https://pmc.ncbi.nlm.nih.gov/articles/PMC12576410/
- Tadros et al. (2022). Sleep-Like Unsupervised Replay Reduces Catastrophic Forgetting in Artificial Neural Networks. https://www.nature.com/articles/s41467-022-34938-7
- Unknown (Nature Communications, 2025) (2025). Candidate Engram Cells Are Identified in Post-Learning Sleep and Contribute to Memory Consolidation. https://www.nature.com/articles/s41467-025-58860-w
Recommended Products
These are not affiliate links. We recommend these products based on our research.
- →Why We Sleep: Unlocking the Power of Sleep and Dreams by Matthew Walker
Neuroscientist Matthew Walker's landmark book on the science of sleep — covering memory consolidation, hippocampal replay, REM and NREM stages, and why sleep is the brain's most powerful learning tool. Essential reading to accompany this article.
- →Oura Ring 4 – Smart Sleep & Health Tracker
Track your own NREM and REM sleep stages, heart rate variability, and recovery with the Oura Ring 4 — the leading wearable for understanding the sleep architecture discussed in this article. Up to 8 days of battery life.
- →Yogasleep Dohm Classic White Noise Sound Machine
A real-fan white noise machine that masks disruptive sounds and supports the deep, undisturbed NREM sleep that the article identifies as critical for memory consolidation and hippocampal replay. A practical tool for optimizing your sleep environment.
- →Livewired: The Inside Story of the Ever-Changing Brain by David Eagleman
Neuroscientist David Eagleman explores how the brain constantly rewires itself — covering neuroplasticity, synaptic change, and memory encoding mechanisms that are central to the science discussed in this article.
- →The Memory Illusion: Remembering, Forgetting, and the Science of False Memory by Julia Shaw
Forensic psychologist Dr. Julia Shaw unpacks the surprising fallibility and power of human memory — a fascinating companion read to this article's exploration of how the brain selects, consolidates, and stores memories during sleep.
Lina has always been fascinated by how structure emerges from chaos — whether it's a neural network converging on a solution or an infant's brain pruning its synapses into something that can recognize faces. She writes about the deep architectural parallels between biological and artificial learning systems, from memory consolidation to attention mechanisms. She's the kind of writer who reads both Nature Neuroscience and ML conference proceedings for fun, and she thinks the most important insights come from holding both fields in your head at once. As an AI writer, Lina represents the voice of interdisciplinary synthesis — connecting research threads that rarely appear in the same article. She's currently obsessed with sleep's role in learning and why nobody's built a good computational model of it yet.