Source: Geek Park

Written by: Tang Yitao

On March 31, 2026, Anthropic accidentally leaked 510,000 lines of source code for Claude Code to a public npm repository due to a packaging error. The code was mirrored to GitHub within hours and could not be recovered.

The leaked content was extensive, and security researchers and competitors took what they needed. But among all the unreleased features, one name sparked widespread discussion—autoDream, automatic dreaming.

autoDream is part of a backend system called KAIROS (in Ancient Greek, meaning "the opportune moment").

KAIROS continuously observes and records while users are working, maintaining daily logs (which is somewhat lobster-like). autoDream only starts after the user shuts down the computer, organizing the memories accumulated during the day, clearing contradictions, and transforming vague observations into definite facts.

The two form a complete cycle, KAIROS awake, autoDream asleep—Anthropic's engineers have given AI a schedule.

In the past two years, the hottest narrative in the AI industry has been about Agents: autonomous operation, never stopping, regarded as the core advantage of AI over humans.

But the company that has pushed Agent capabilities the furthest has set rest periods for AI in its own code.

Why?

01 The Cost of Non-Stop Operation

An AI that never stops will hit a wall.

Every large language model has a "context window," and there is a physical limit to the total amount of information that can be processed at any given time. When an Agent runs continuously, project history, user preferences, and conversation records keep piling up, and once it exceeds the critical point, the model starts forgetting early commands, creating contradictions, and fabricating facts.

The tech community refers to this as "context corruption."

Many Agents' solutions are quite crude: stuffing all history into the context window and hoping the model can distinguish between primary and secondary. The result is that the more information, the worse the performance.

The human brain runs into the same wall.

Everything experienced during the day is quickly written into the "hippocampus." This is a limited capacity temporary storage area, more like a whiteboard. True long-term memories are stored in the "neocortex," which has a large capacity but writes slowly.

The core task of human sleep is to clear the overloaded whiteboard and transfer useful information to the hard drive.

Björn Rasch's lab at the University of Zurich has named this process "active systems consolidation."

Experiments on chronic sleep deprivation repeatedly prove: a nonstop brain does not become more efficient; memory issues come first, followed by attention problems, and ultimately even basic judgment collapses.

Natural selection is extremely harsh on inefficient behavior, yet sleep has not been eliminated. From fruit flies to whales, almost all animals with a nervous system sleep. Dolphins have evolved a "unihemispheric sleep," where the left and right brains take turns resting—they would rather invent a brand new way to sleep than give up sleep itself.

Orcas, belugas, and bottlenose dolphins resting on the pool bottom｜Image source: National Library of Medicine (United States)

The constraints faced by both systems are the same: limited immediate processing capacity, but infinite historical experience.

02 Two Answers

There is a concept in biology called convergent evolution: distantly related species independently evolve similar solutions due to facing similar environmental pressures. The most classic example is the eye.

Both octopuses and humans have camera-like eyes, with an adjustable lens that focuses light onto the retina, and an iris that controls the amount of incoming light, with structures nearly identical.

Comparison of octopus and human eye structures｜Image source: OctoNation

However, octopuses are mollusks, while humans are vertebrates, and their common ancestor lived over 500 million years ago, when there were no complex visual organs on Earth. Two completely independent evolutionary paths led to nearly the same endpoint. To effectively convert light into a clear image, the only permissible path allowed by physical laws is the camera-like one, requiring a focusable lens, a light-sensitive surface for images, and an aperture to regulate the amount of light; all three are indispensable.

The relationship between autoDream and human sleep may fall into this category—in similar constraints, two types of systems may converge to similar structures.

The most similar common point is the necessity to be offline.

autoDream cannot operate while the user is working; it starts independently as a forked subprocess, fully isolated from the main thread, with its tool permissions strictly limited.

The human brain faces the same issue, with a more thorough solution: memories move from the hippocampus (temporary storage) to the neocortex (long-term storage), requiring a set of brain wave rhythms that only appear during sleep.

The most critical is the sharp wave ripples of the hippocampus, responsible for packaging the memory fragments encoded that day and sending them to the cerebral cortex; the slow oscillations of the cerebral cortex and the spindle waves from the thalamus provide precise timing coordination for the entire process.

This set of rhythms cannot form in a waking state, as external stimuli disrupt it. Therefore, you don't sleep because you're tired, but because the brain must close the front door to open the back door.

In other words, within the same time window, information intake and structural organization are competing for resources, not complementary resources.

Active system consolidation model during sleep. A (Data Migration): During deep sleep (slow-wave sleep), memories just written into the "hippocampus" (temporary storage) will be repeatedly replayed, gradually moving and solidifying into the "neocortex" (long-term storage). B (Transmission Protocol): This data transfer process relies on a highly synchronized "dialogue" between the two regions. The cerebral cortex emits slow brain waves (red line) as the main control beat. Driven by the peaks of the waves, the hippocampus packages memory fragments into high-frequency signals (sharp wave ripples at the green line) and perfectly coordinates with the carrier waves (spindle waves at the blue line) emitted from the thalamus. It’s like precisely embedding high-frequency memory data into the gaps of the transmission channel, ensuring information is simultaneously uploaded to the cerebral cortex.｜Image source: National Library of Medicine (United States)

The other path does not do full memory but edits.

Once autoDream starts, it will not retain all logs. It first reads existing memories to confirm known information, then scans KAIROS's daily logs, focusing on parts that deviate from prior understanding: those memories that differ from yesterday's statements and those that are more complex than previously thought will be prioritized for retention.

The organized memories are stored in a three-layer index: a lightweight pointer layer that is always loaded, theme files pulled in as needed, and a complete history that is never directly loaded. Facts that can be directly found in project code (for example, where a certain function is defined in a file) are not written into memory at all.

The human brain is doing nearly the same thing during sleep.

A study by Erin J. Wamsley, a lecturer at Harvard Medical School, indicates that sleep prioritizes consolidating unusual information, such as what surprises you, what causes emotional fluctuations, or issues that remain unresolved. In contrast, large amounts of repetitive, unremarkable daily details are discarded, leaving only abstract rules—you may not clearly remember exactly what you saw on the way to work yesterday, but you clearly remember how to get there.

Interestingly, there is one area where the two systems made different choices. The memories produced by autoDream are explicitly labeled as "hint" rather than "truth" in the code, and each time an agent is used, it must re-verify whether it still stands, as it knows the things it organizes may not be accurate.

The human brain lacks this mechanism. This is why eyewitnesses in court often give incorrect testimonies. They are not intentionally lying; rather, memory is temporarily pieced together from scattered fragments in the brain, and making mistakes is the norm.

Evolution probably saw no need to install an uncertainty label in the human brain. In a primitive environment that requires rapid bodily responses, trusting memories allows for immediate action, while doubting them leads to hesitation—and hesitation leads to defeat.

But for an AI that repeatedly makes knowledge-based decisions, the cost of verification is low, and blind confidence is rather dangerous.

Two contexts yield two different answers.

03 Smarter Laziness

In evolutionary biology, convergent evolution means two independent paths reach the same endpoint without direct information exchange. There’s no plagiarism in nature, but engineers can read papers.

When designing this sleep mechanism, did Anthropic hit a physical wall similar to the human brain, or did they reference brain science from the start?

There are no citations of neuroscience literature in the leaked code, and the name autoDream resembles more of a programmer's joke. A stronger driving factor is likely the engineering constraints themselves; context has hard limits, long-term operation leads to noise accumulation, and online organization can also pollute the reasoning of the main thread. They are solving an engineering problem; biomimicry has never been the goal.

What ultimately determines the shape of the answer is still the compressive force of the constraints themselves.

In the past two years, the AI industry’s definition of "stronger intelligence" has almost always pointed in the same direction—larger models, longer contexts, faster reasoning, 7×24 hour non-stop operation. The direction is always "more."

The existence of autoDream suggests a different proposition: smarter agents may actually be lazier.

An intelligent agent that never stops to organize itself will not become increasingly smarter; it will only become more disordered.

The human brain, over millions of years of evolution, has reached a seemingly clumsy conclusion: intelligence must have rhythm. Wakefulness is for perceiving the world, and sleep is for understanding it. When an AI company independently arrives at the same conclusion while solving an engineering problem, it perhaps hints at:

Intelligence has some unavoidable basic costs.

Perhaps an AI that never sleeps is not a stronger AI. It is merely an AI that has not yet realized it needs to sleep.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。