I study how AI assistants remember people, as a research assistant at the University of Southern Mississippi, so I paid attention when OpenAI said ChatGPT now dreams. Its headline example is exactly the kind of failure my research is about: a background process that rewrites "You're going to Singapore in July" as "You went to Singapore in July 2026" once the trip is over.
ChatGPT was not the first. As far as I can tell, Honcho was the first memory system to use the word, in a pull request titled "wip: dreaming" in September 2025. OpenAI now calls its April 2025 memory update the first version of dreaming, but I can't find the name in anything it published before June 2026. Anthropic, Devin, OpenClaw and Letta have their own versions too. All of them let a model rewrite your memory while you are away: merge entries, replace them, delete them, and add things you never said.
This page is what I found reading their docs, launch posts and, for Honcho, the source code. The vendors present dreaming as an improvement. None has released a test of it that someone else could rerun.
What is dreaming in AI memory?
Dreaming is memory writing that runs outside the conversation. A model reads stored memory together with many past sessions, then merges duplicates, replaces outdated entries, deletes notes, and adds inferences that no single session stated. It runs on a schedule, after a threshold, or continually in the background. Normal memory writing runs inside the conversation and mostly records what was said.
Day and night
Every memory system has a writer that runs during the conversation. Dreaming adds a second writer that runs outside it. The second writer reads many sessions together and edits memory across them, which the first writer cannot do.
| The day writernormal memory writing, inside a session | The night writerdreaming, outside the conversation: on a schedule, after a threshold, or continually | |
|---|---|---|
| SEES | one session or one batch of messages | stored memory plus many sessions (Anthropic: 1 to 100) |
| READS ITS OWN PAST OUTPUT | rarely; it mostly appends | always; the store is its input |
| DELETES OR REPLACES | rarely | yes; five of the six systems here document it |
| WRITES | what was said. Honcho's writer prompt: facts "derived directly from their messages" | also things nobody said: deductions, patterns, "insights", "lessons missed" |
| BUDGET | one fast call per batch (Honcho) or a few file edits (Anthropic, Devin) | a multi-step agent, often a larger model |
| WHO IS WATCHING | the user; the next turn can catch a mistake | no one at write time; at most, the user sees the result later |
A night in the archive
A small store about a user, Ana, written by the day writer. Press Dream to run one night. Each step is an operation a real dreamer performs; the store and its contents are invented for illustration.
- Merge: Anthropic, Devin, OpenClawDuplicates become one.
- Supersede: Honcho, Anthropic, ChatGPTTuesday is struck; Thursday is recorded.
- Deduce: Honcho"Nurse" becomes "works in healthcare".
- Induce: HonchoTwo Python requests become a preference.
- Rewrite risk: curation dreamers"This week" is lost.
- Deletion risk: DevinThe never-read allergy is removed.
The temple of elaboration
Honcho dreams differently from the others: it keeps the stated facts and builds inferences on top of them, each tied to the facts that hold it up. Read from the source code (plastic-labs/honcho, commit cb8ab1a).
When the store has gained 50 new explicit facts, the user has been idle for an hour, and 8 hours have passed since the last dream, a deduction agent (up to 12 tool steps) looks for updates, implications and contradictions and may delete outdated facts. Then an induction agent (up to 10 steps) writes patterns, preferences and personality traits; its prompt says "think like a psychologist". Every inference must cite the IDs of the facts it rests on; invented IDs are discarded. At answer time, the reading agent is told to check an inference's premises before trusting it.
Six houses that dream
The six fall into two families. Curation rewrites memory into a cleaner version of itself; five of the six work this way. Elaboration keeps the stated facts and builds inferences on top of them; only Honcho does. Across both, the designs differ most in what protects memory from a bad night.
More systems coming soon.
Where dreams go wrong
Each failure below has either been measured in a paper (gold) or is allowed by a product's design and has not been measured (terracotta). The papers show what consolidation can do in a research setup, not how often it happens inside any product on this page.
Rewrites compound
Memory updated again and again by an LLM first helps, then degrades, sometimes below having no memory; keeping raw episodes was competitive. Zhang et al., arXiv 2605.12978.
The tense is lost
In 244 of 381 matched pairs, a memory writer turned "I am driving a Peugeot" into "drives a Peugeot" while keeping "I drive a Peugeot" as said, and never the reverse. The pattern held in all 11 configurations and in mem0, Graphiti and Letta. LAPSE, arXiv 2609.36457. ChatGPT's dreamer is built to make the opposite move, from "going to" to "went", and OpenAI reports a gain on its own time evaluation.
The speaker's standing is lost
Consolidation kept a claim but dropped the limits on who said it and how it may be used in 48 of 49 configurations. AuthMem-Bench, arXiv 2608.01679.
Traits from two facts
Honcho's induction may write a personality trait from two observations. Whether such traits are stable across runs or predict what the person says next is unknown.
Forgetting the rare
Devin deletes notes no session has used. A fact needed once a year is, by that rule, a candidate for deletion.
Three ways to study a dream
Psychometrics
Treat a dream as a measurement. Reliability: run the same night twice; do the same traits and inferences come back? Validity: do the inferred traits predict what the person says later? Honcho's trait inference makes this literal.
Cognition
Human sleep consolidation has known errors: the gist kept while details go, confusion about who said what, related items remembered that were never seen. Each predicts a specific failure in a dreamer. Write the predictions down before running anything.
Real use
Honcho and OpenClaw are open source and keep enough history to diff a store before and after a night. Anthropic keeps both stores and the dream's transcript, behind preview access. Devin shows each night's changes, but memory is private. ChatGPT shows a summary rather than the memory itself, so a night's changes are hard to see from outside.
Questions
What is AI dreaming?
Dreaming is memory maintenance that runs outside the conversation. A separate model run reviews past sessions and stored memory, then rewrites or extends that memory so later sessions start from a cleaner, more complete version. The name borrows from sleep, when people consolidate memories.
How is dreaming different from normal memory writing?
Normal writing runs inside a conversation, sees one session, and mostly appends. Dreaming runs outside the conversation, sees many sessions and stored memory at once, reads its own earlier output, deletes and replaces entries, and writes inferences nobody said. No user is present to catch a mistake.
Which AI products use dreaming?
As of October 2026: ChatGPT (since April 2025, rebuilt in June 2026), Anthropic's Claude Managed Agents (research preview), Devin by Cognition (about once a day), Honcho by Plastic Labs (open source, since October 2025), OpenClaw (nightly), and Letta (after a set number of steps or on context compaction).
Does ChatGPT have dreaming?
Yes. OpenAI added a first version in April 2025, which curated memory from chat history in the background alongside saved memories. In June 2026 it made a rebuilt version, Dreaming V3, the foundation of ChatGPT memory. It updates memories as time passes and shows the result on a memory summary page, where you can correct an entry or tell ChatGPT not to mention it again. Temporary chats do not create or update memories.
Is dreaming the same as sleep-time compute?
They are related. Sleep-time compute (Lin et al., 2025) has a model reason about a context before questions arrive. Dreaming applies offline work to the memory store itself. OpenClaw's docs name sleep-time compute as an inspiration, and Letta now calls its sleep-time agents "dreaming".
Does dreaming make memory better?
OpenAI and Anthropic report that it does; neither has released a test that others can rerun. OpenAI compares its 2024, 2025 and 2026 memory systems on its own evaluations of recall, preferences and time awareness, and reports gains on each without publishing the evaluations. Anthropic's launch post reports one customer's completion rates rising about 6x "in their tests", without the task set or baseline. An independent study found that memory rewritten again and again by an LLM can end up worse than no memory.
Can users see or correct what a dream wrote?
It depends on the product. ChatGPT shows a memory summary that you can correct or delete, though OpenAI says it may not include every detail memory can use. Anthropic writes each dream to a new store, and a person decides whether to use it. Devin lists every dream and what it changed. OpenClaw keeps a copy of the file it rewrote and a diary of each night. Honcho and Letta leave review to the developer; Letta can have a second agent check the changes first.