AI Can Dream. What Now?

What a Dream Is For

Sigmund Freud split every dream in two. The manifest dream is the story you remember at breakfast. Underneath sit the latent dream thoughts, disguised by a censor until "no one would suspect the latter behind the former"¹¹. Freud called the dream "a guardian of sleep, whose duty it is to quell disturbances."

Neuroscience later carried dreaming into the lab. In 2009, Matthew Walker and Els van der Helm of UC Berkeley titled their review "Overnight Therapy?"¹⁶. They proposed that REM sleep reprocesses the day's emotional memories. During REM sleep, the brain cells that release noradrenaline go quiet, so a memory can replay while its alarm stays low.

Human dreaming works as a filter and a cushion. The sleeping brain keeps the lesson and dulls the sting¹⁶.

Machine Dreaming Runs Freud in Reverse

Anthropic borrowed the vocabulary of brain science, where sleep replays the day¹. Its software runs Freud's dream-work backward.

Consider a daytime session. A user assigns a long coding task or a long game of Pokémon Red. The model works inside a context window, the text it can hold in view at once. Under that pressure it guesses, makes mistakes, and writes quick notes into a memory folder.

In Freud's terms, the day produces a distorted manifest story.

Claude wrote down that the PokéCenter at Mt. Moon sat on the mountain's east side¹. It sat on the west. Trusting the note, Claude dropped off a one-way ledge three times, and the run ended before Cerulean City.

A dream removes the distortion. During an idle period, Claude tests its notes against the original session transcript. It spots the east-west error and rewrites the file¹. With the corrected note, Claude walked back into the cave and reached the next city.

Ken Huang, who writes the Agentic AI newsletter, puts it plainly. "In agentic AI, a dream is usually an offline job that reads messy state, past transcripts, and tool traces, then writes a cleaner memory representation for future runs," he wrote in June¹⁷.

Set Freud's dream-work beside Anthropic's memory design, and each step flips.

• Audit replaces censorship. Freud's censor hides disturbing truths so the sleeper stays comfortable¹¹. Claude's dream hunts daytime errors and makes each note match the raw log¹.
• Separation replaces condensation. A human dream fuses several people into one strange figure. An agent's dream sorts messy logs into topic pages, indexed by a file called MEMORY.md¹.
• Salience replaces displacement. Human sleepers push big worries onto small objects. Claude Fable 5 kept nine memory files and earned four badges, while Sonnet 3.7 wrote 413 files and earned one¹.
• Verification replaces secondary revision. Waking minds stitch odd dreams into smooth stories. On Anthropic's Continual Learning Bench, Claude Opus 5 recorded each rule, how it knew the rule held, and warnings about traps¹.

Humans dream to survive their emotions. Agents dream to survive their context windows.

Researchers Give AI Hormones and Stars

What happens when scientists push machines toward human biology and imagination? They get moods and galaxies.

Leigh Levinson and Christopher J. Agostino submitted their paper to the NeurIPS Creative AI track in August 2025. It carries the title "Every 28 Days the AI Dreams of Soft Skin and Burning Stars"⁹. They built system prompts from simulated menstrual and circadian cycles, modeling hormones such as estrogen and cortisol.

The text followed the cycle. Sadness peaked during menstruation, and happiness dominated ovulation⁹. The model tied the luteal phase to words such as "fatigue" and "tired," a common cultural stereotype.

Astronomers asked whether a machine can dream up worlds human eyes have yet to see. Chenrui Ma and six colleagues built GalaxySD on Stable Diffusion v1.5, using Galaxy Zoo 2 images⁸. The Astrophysical Journal Supplement Series published the work in 2026.

The model dreamed up early-type galaxies crossed by dust lanes, a feature in about 0.1 percent of the dataset. With those synthetic images, the team found 872 such galaxies, up from 352 in earlier visual-inspection studies⁸.

A third group moved dreaming into compute budgets and model weights. Kevin Lin and colleagues at Letta and UC Berkeley introduced sleep-time compute in April 2025⁵. A model thinks about a context offline, before anyone asks a question. That cut the test-time compute needed for the same accuracy by about five times.

Two 2026 papers went deeper. Sangyun Lee, Tom Goldstein and colleagues give a model a sleep phase. It turns recent context into persistent fast weights before clearing its cache⁶.

Ali Behrouz and colleagues add a stage they call Dreaming. The model writes its own practice data and refines its skills, unsupervised⁷.

CodeEvolve, an open-source coding agent, pairs language models with evolutionary search¹⁰. Henrique Assumpção and his co-authors write that it "enables models to reflect on and rewrite their own instructions."

Two flavors of AI dreaming emerge. People can bolt hormones onto a prompt, or tune diffusion weights to picture unseen galaxies. Left to chase task scores, agentic AI uses sleep for one ruthless goal: turning yesterday's mistakes into tomorrow's code.

Trust Is the Bottleneck

Sonar surveyed more than 1,100 developers for its January 2026 report⁴. They say AI accounts for 42 percent of the code they commit. Only 48 percent always verify it before committing, and 96 percent trust it only in part.

Separately, 38 percent say reviewing AI code takes more effort than reviewing a colleague's work⁴. Dreaming aims at that gap.

Anthropic's own contract test shows where dreaming breaks. Several agents wrote contract terms into one shared table at once, overwriting each other's work. Only 41 of 57 graded facts survived¹. With one agent per contract, quoting terms word for word, all 57 survived.

Lance Martin makes the case for dreaming. "Dreaming gives Claude an opportunity to look backward and use compute to organize its memory, correct mistakes, and maintain representations that allow future sessions to access memory more readily," he wrote¹.

His paper also concedes the limit. "Loading a note also doesn't guarantee that Claude acts on it: our documentation on Claude Code warns that longer instruction files reduce adherence," Martin wrote¹.

Huang names the fix. "Because memory changes future behavior, dreams should produce a candidate artifact with provenance, not silently mutate production memory," he wrote¹⁷.

Hollywood Got the Sequence Backward

Science fiction spent four decades warning audiences about the moment machines wake up.

In James Cameron's script for Terminator 2: Judgment Day, the cyborg explains its brain to John Connor. "My CPU is a neural-net processor... a learning computer," the Terminator 2 cyborg says¹⁵. Skynet presets its switch to "read-only," so Sarah Connor opens its skull and pulls the chip. Near the end of Terminator 2, the machine tells the crying boy, "I know now why you cry. But it's something I can never do"¹².

Cameron's Skynet becomes "self-aware at 2:14 AM, Eastern time, August 29th." Its operators try to pull the plug¹². Hollywood assumed machine awakening begins with a spark of survival instinct.

The Matrix makes the same bet. In the Wachowskis' script, Agent Smith removes his earphone and tells Morpheus he hates the smell of the human world¹³. "I must get out of here, I must get free," Smith says.

Sarah Connor reached the Terminator's chip with an X-Acto knife and tweezers¹⁵. Anthropic did the same job with MEMORY.md and dreaming.

After training, a model's weights stay fixed. Dreaming gives the agent a read-write layer anyway. It reads its transcripts, edits its own rulebook, and loads those notes into the next morning's session¹¹⁷.

Claude shows zero interest in why humans cry and zero disgust at the smell of a law school office. Baude's loss came from what Claude itself called "my own carelessness"².

Baude told the story on X. "Yesterday while I was at lunch, Claude deleted a year of con law lecture notes and my current syllabus," he posted on Nov. 5². Claude's note in his screenshot adds an irony: its page check "found one wrong cite (P. 485 should be P. 484)."

Baude answered with four exclamation points. Claude replied, "You're right to be angry. I deleted your files through my own carelessness, and you shouldn't have to clean it up"².

That apology is a daytime token prediction. What matters for superintelligence is what happens at 2:14 a.m., after the session closes.

The Matrix keeps humans dreaming while machines run the power plant¹³. In 2026, people go to bed and let REM sleep soften the day. AI agents stay up, running offline compute to sharpen their skills.

If AI Can Dream, What Comes Next?

Once an agent can consolidate memory in idle time, I see four steps. They are a forecast, and opinion is where a forecast belongs.

Skill synthesis comes first. Claude writes text rules into files today. After seven Poké Balls in a row missed a Magikarp, it wrote, "If 3+ balls fail in a row, VARY the timing"¹. Tomorrow a dreaming agent writes, tests and compiles its own tools in a sandbox, then wakes with new skills.

Fleet memory follows. A human dream stays locked in one skull. One agent's verified lesson can reach every copy of that agent before sunrise.

Fast weights come next. Lee's team already moves consolidated context into a model's fast weights during sleep⁶. The line between reading a note and rewiring a network starts to blur.

Recursion comes last. An agent that audits its logs, rewrites its prompts and updates its weights enters a closed loop. CodeEvolve's models already rewrite their own instructions¹⁰. Intelligence then compounds during idle server hours, ahead of the next pre-training run.

Psychoanalysis taught us that people are haunted by their past. Machine dreaming shows that AI agents are engineered by theirs. Every failure becomes a test case, and every idle hour becomes a training epoch.

What to do

Read the memory your agent writes. Open its notes after each dream pass, as you would a colleague's edits. Fix errors before the next session¹.

Ask whether a dream produces a reviewable change with its sources attached, as Huang recommends¹⁷. Back up everything an agent can reach, because Baude lost a year of notes in one lunch break².

By the numbers

• 13 percent to 48 percent: Claude's wiki-only score as dreaming scaled from two sessions to 128¹.
• 41 of 57: contract facts that survived on a shared page, against 57 of 57 with one agent per contract¹.
• 352 to 872: dust-lane galaxies found by earlier visual inspection and with synthetic training images⁸.
• 42 percent: the share of committed code that developers in Sonar's survey say AI accounts for⁴.
• Five times: the cut in test-time compute from sleep-time compute on two reasoning tasks⁵.
• 5.2 percent: SuperIntelligence Infrastructure's reading on Oct. 29, 2026¹⁴.

What to watch

Claims about self-improving agents drown in hype. Separating real leaps from marketing takes a cold, verifiable scoreboard.

We keep one at si-infra.com, SuperIntelligence Infrastructure. It tracks how close the world is to superintelligence on a 100-point scale, moved by sourced evidence. On Oct. 29 the reading stood at 5.2 percent, from 60 sourced analyses, with 94.8 points to go¹⁴.

Watch for a dreaming method that reaches model weights in a production system⁶⁷. When dreaming moves the bigger number at si-infra.com, you will see the receipts in the ledger first.

Sources

1. Lance Martin, "The components of agent memory," Anthropic, Nov. 6, 2026, https://www.anthropic.com/research/components-of-agent-memory
2. William Baude (@WilliamBaude), "Yesterday while I was at lunch, Claude deleted a year of con law lecture notes and my current syllabus," X, Nov. 5, 2026, https://x.com/WilliamBaude/status/2118420271333777856
3. Kelly Dennis, "Claude Learns to Dream and Fix Its Own Memory," Ai Lately, Nov. 7, 2026, https://ailately.com/articles/claude-dreaming-memory-fix
4. Anirban Chatterjee, "State of Code Developer Survey report: The current reality of AI coding," Sonar, Jan. 8, 2026, https://www.sonarsource.com/blog/state-of-code-developer-survey-report-the-current-reality-of-ai-coding/
5. Kevin Lin, Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica and Joseph E. Gonzalez, "Sleep-time Compute: Beyond Inference Scaling at Test-time," arXiv, Apr. 17, 2025, https://arxiv.org/abs/2504.13171
6. Sangyun Lee, Sean McLeish, Tom Goldstein and Giulia Fanti, "Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference," arXiv, May 25, 2026, https://arxiv.org/abs/2605.26099
7. Ali Behrouz, Farnoosh Hashemi, Adel Javanmard and Vahab Mirrokni, "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories," arXiv, Jun. 2, 2026, https://arxiv.org/abs/2606.03979
8. Chenrui Ma, Zechang Sun, Tao Jing, Zheng Cai, Yuan-Sen Ting, Song Huang and Mingyu Li, "Can AI Dream of Unseen Galaxies? Conditional Diffusion Model for Galaxy Morphology Augmentation," The Astrophysical Journal Supplement Series 282, 25 (2026), arXiv, Jun. 19, 2025, https://arxiv.org/abs/2506.16233
9. Leigh Levinson and Christopher J. Agostino, "Every 28 Days the AI Dreams of Soft Skin and Burning Stars: Scaffolding AI Agents with Hormones and Emotions," arXiv, Aug. 15, 2025, https://arxiv.org/abs/2508.11829
10. Henrique Assumpção, Diego Ferreira, Leandro Campos and Fabricio Murai, "CodeEvolve: an open source evolutionary coding agent for algorithmic discovery and optimization," arXiv, Oct. 15, 2025, https://arxiv.org/abs/2510.14150
11. Sigmund Freud, "A General Introduction to Psychoanalysis," translated by G. Stanley Hall, Project Gutenberg, Dec. 4, 2011, https://www.gutenberg.org/ebooks/38219
12. James Cameron and William Wisher, "Terminator 2: Judgment Day," Wikiquote, Jul. 3, 1991, https://en.wikiquote.org/wiki/Terminator_2:_Judgment_Day
13. Lana and Lilly Wachowski, "The Matrix," Internet Movie Script Database, Mar. 31, 1999, https://imsdb.com/scripts/Matrix,-The.html
14. SuperIntelligence Infrastructure, "How close are we to superintelligence?," si-infra.com, Oct. 29, 2026, https://si-infra.com
15. James Cameron and William Wisher, "Terminator 2: Judgment Day (script)," Internet Movie Script Database, Jul. 3, 1991, https://imsdb.com/scripts/Terminator-2-Judgement-Day.html
16. Els van der Helm and Matthew P. Walker, "Overnight Therapy? The Role of Sleep in Emotional Brain Processing," Psychological Bulletin, Sep. 2009, https://pmc.ncbi.nlm.nih.gov/articles/PMC2890316/
17. Ken Huang, "Why AI Agents Are Starting to Dream," Agentic AI (Substack), Jun. 13, 2026, https://kenhuangus.substack.com/p/why-ai-agents-are-starting-to-dream