
The mind remembers in two ways: when we go looking, and when we are asleep and not looking at all. The second may matter more. And the newest AI systems are now being built to remember the same way.
You are washing a cup, and suddenly your grandmother's kitchen is there. The blue enamel tin. The smell of mustard oil heating. You did not ask for it. It simply arrived.
Other times you try hard to recall a name, a street, the colour of a dress, and get nothing. Then at 3 a.m. it surfaces on its own, whole and certain. Anyone who has lived long enough knows both experiences. Few of us stop to ask what they reveal about how memory actually works.
Two doors into the past
Deliberate retrieval is effortful and often disappointing. Spontaneous recall is effortless and often startlingly rich. They are two routes into the same store.
Deliberate retrieval is a search. We supply a cue, a question or a place, and the brain tries to reassemble a pattern that matches. It works well for facts we rehearse. It struggles with the faded, emotional, autobiographical material we most want back.
Spontaneous recall works differently. Psychologists call these involuntary autobiographical memories, and research suggests they arrive as often as the ones we chase. A smell, a song or a slant of light matches a stored pattern, and the memory unlocks without permission. No effort, no searching, just a door that opens because the right key brushed against it.
The lesson is quietly radical. Memory is not a filing cabinet we open at will. It is a living system that keeps working when we are not watching.
What the sleeping brain is doing
Sleep is not downtime for memory. It is when most of the sorting happens.
During the day, the hippocampus captures experience quickly and in rough form, like a reporter's notebook. Only some of it is kept. Chemical signals released around an emotional or surprising moment act like a bookmark, flagging that experience as worth holding on to (Frey & Morris, Nature, 1997).
Then, in sleep and in quiet rest, the hippocampus replays the day. Flagged moments are replayed more often. Through this replay, experiences are gradually distilled and woven into the neocortex, where long-term knowledge lives (McClelland, McNaughton & O'Reilly, 1995; Klinzing, Niethard & Born, Nature Neuroscience, 2019).
Dreams seem to be the felt edge of that work. Fragments of the recent past get stitched to much older memories. The result is strange, but the function appears practical: link the new to the old, keep what matters, and let the rest fade.
This is why a forgotten name can arrive at 3 a.m. The brain kept searching after you stopped.
Machines that dream
In November 2026, Anthropic published The components of agent memory, and its central idea is borrowed straight from the sleeping brain. They call it dreaming.
The piece opens with Kierkegaard: life is understood backward but lived forward. An AI agent faces the same bind. While working, it has to guess what will be worth remembering before it knows what comes next. So the design splits memory into two kinds of work, just as the brain does.
In the moment, Claude takes notes. Like the hippocampus, it captures salient events as they happen, writing them to a memory folder. Anthropic frames each note as a quiet prediction: by writing a rule down, the agent bets it will matter again.
Offline, Claude dreams. Between sessions, it revisits the raw transcripts of its past work and reorganises what it knows. It fixes wrong notes, merges duplicates, builds linked wiki pages, and promotes the most important points to a small set of notes loaded at the start of every new session. The idea builds on sleep-time compute, proposed by researchers at Letta and UC Berkeley (arXiv, 2025).
The experiments are vivid.
- Knowing what to forget. Playing Pokémon Red, an older model saved everything: 413 files across 225 folders, one badge earned. A newer model kept nine files plus a single dashboard, and earned four. It stored only what the screen could not tell it, such as strategy, plans and rules it had tested.
- Dreaming fixes false memories. One run wrongly noted which side of a mountain a landmark sat on. Trusting that note, the agent walked off a one-way ledge three times and never got through. After a dreaming pass checked the note against the original transcript and corrected it, the agent went straight back into the cave and reached the next city.
- Dreaming builds understanding. Asked to build a wiki over roughly 350,000 internal company files, Claude's wiki-only score rose from 13% to 48% as dreaming scaled from two sessions to 128. Like a new employee becoming a seasoned one, its map of the place grew richer each night.
One finding is especially human. Telling stronger models exactly what to record, such as "record only failures", made them worse. They did best when told what memory was for and trusted to judge salience themselves.
The parallels, side by side
Put the two systems next to each other and the rhymes are hard to miss.
| What memory needs | In the human brain | In an AI agent |
|---|---|---|
| Fast capture | Hippocampus records the day in rough form | Session transcripts and in-the-moment notes |
| Marking what matters | Emotion and surprise chemically tag a moment | The agent judges salience and writes a note |
| Offline sorting | Replay during sleep and quiet rest | Dreaming between sessions |
| Linking new to old | Dreams stitch recent fragments to older memories | Wiki pages link new notes to existing knowledge |
| Correcting errors | Recall reopens a memory and updates it | Dreaming checks notes against original transcripts |
| Forgetting well | Unrehearsed detail fades | Low-value detail is pruned or pushed down |
| What's always at hand | Core knowledge and sense of self | A small set of notes loaded into every session |
| Search vs. surfacing | Effortful recall vs. a smell that unlocks a memory | Searching transcripts vs. notes already in context |
The deepest parallel is the last row. The memories that serve us best are not the ones we must hunt for. They are the ones that arrive when needed, because something already organised them while we were not looking.
Where the parallel breaks
The machine has something we never will: the original transcript.
When Claude dreams, it can check a note against an exact record of what happened. We cannot. Every time a human memory is recalled, it is rebuilt from fragments and quietly rewritten, a process neuroscientists call reconsolidation. A memory can feel vivid and certain and still be wrong. Vivid is not the same as accurate.
That cuts both ways. Our memories drift, but they also deepen. A grandmother's kitchen recalled at forty carries meanings it did not have at ten. The rewriting is not only error. It is how a life makes sense of itself.
An agent's memory serves a task: the next question, the next level, the next brief. A human memory serves a self. The architecture rhymes. The purpose does not.
Recover, don't generate
What the machines are learning, we already knew in our bodies: memory is not stored once and fetched later. It is tended.
The best AI agents now remember less, sort more, and revisit the past to correct it. That is good advice for people too. The memories we care about most, of parents, of childhood rooms, of voices we can no longer hear, rarely come back through effort alone. They return through the right cue, a patient setting, and time to let the mind do its offline work.
The machine dreams over a perfect transcript. We dream over something more fragile and more precious. That is exactly why our memories are worth recovering with care, not generating, and not leaving to chance.
— M
References
- Anthropic (2026, 6 November). The components of agent memory. www.anthropic.com/research/components-of-agent-memory
- Lin, K., Snell, C., Wang, Y., Packer, C., Wooders, S., Stoica, I., & Gonzalez, J. E. (2025). Sleep-time compute: Beyond inference scaling at test-time (preprint). arXiv:2504.13171. arxiv.org/abs/2504.13171
Show all 5 referencesShow fewer
- McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419–457. doi.org/10.1037/0033-295X.102.3.419
- Frey, U., & Morris, R. G. M. (1997). Synaptic tagging and long-term potentiation. Nature, 385(6616), 533–536. doi.org/10.1038/385533a0
- Klinzing, J. G., Niethard, N., & Born, J. (2019). Mechanisms of systems memory consolidation during sleep. Nature Neuroscience, 22(10), 1598–1610. doi.org/10.1038/s41593-019-0467-3
