Hitting delete does not mean the company deleted it. ChatGPT and Claude clear deleted chats from their systems within about 30 days, and Gemini holds your activity for 18 months by default. Those are the numbers most pages stop at. The real answer runs longer, because a federal preservation order, a training pipeline, or a human-review queue can keep a copy alive after your history shows nothing. The per-platform table below is current for 2026, which matters because the New York Times litigation rewrote what deletion means at OpenAI and then partly reversed it in October 2025.
You can delete the chat. Whether the company has, too, is now a question for a court docket.
ChatGPT: a 30-day default with a court hold on top
ChatGPT deletes a conversation from OpenAI systems within 30 days of you deleting it, and Temporary Chats run on the same 30-day clock automatically. That is the baseline. The exception came from the New York Times copyright suit in the Southern District of New York, where Magistrate Judge Ona T. Wang ordered OpenAI to preserve output logs it would otherwise have deleted. For months, that order sat on top of the 30-day default for consumer ChatGPT and API content.
Here is what most ranking pages still get wrong. On October 9, 2025, the court released OpenAI from the broad duty to preserve and segregate logs going forward, so the company is no longer required to hold logs created after September 26, 2025. Two carve-outs remain. Logs already preserved under the original order stay accessible to the plaintiffs, and OpenAI must keep retaining logs tied to accounts the Times specifically flagged. So the accurate 2026 answer is not indefinite for everyone. It is 30 days by default, plus a frozen pool of older preserved logs and a targeted hold on flagged accounts.
Training is the separate clock. On consumer ChatGPT, your conversations can feed model improvement unless you switch that off in data controls. Enterprise, Edu, and zero-data-retention tiers sit outside the consumer default, run on admin-controlled retention, and stay exempt from certain legal retention orders. Put plainly: deletion governs the visible copy, the opt-out governs the training copy, and the court governs a slice of older logs that neither switch can touch.
A deletion setting is a promise about the future. It does almost nothing about copies a court already froze in place.
Claude: 30-day backend deletion, training data up to 5 years
Claude removes a conversation from your chat history immediately and deletes it from backend storage within 30 days. That mirrors ChatGPT default. The long tail lives in the training pipeline. Opt in to help improve Claude, and Anthropic may keep those conversations in a de-identified form for up to 5 years. Switching the setting off stops future training use, though anything mid-cycle stays in for that round.
Two more paths round out the picture. Chats flagged for violating usage policies are kept up to 2 years, and the trust-and-safety classification scores attached to them are held up to 7 years. Incognito chats are never used for model improvement, whatever your settings say. So Claude honest answer is layered: 30 days for ordinary deleted chats, up to 5 years for opted-in de-identified training data, and a 7-year safety tail that only touches flagged content.
Gemini: 18-month default, but human-reviewed chats live 3 years
Gemini is the outlier on the default and the easiest to tighten. Gemini Apps Activity auto-deletes after 18 months out of the box, and you can move that to 3 months, 36 months, or off entirely. Even with activity off, conversations stay in your account for up to 72 hours to run the service and process feedback. The 18-month figure makes Gemini look like the worst keeper at a glance, yet its default is the most adjustable of the three.
Human review is the part deletion never reaches. Chats that human reviewers have seen, along with related signals such as language, device type, location, and feedback, are not deleted when you delete your activity. They are kept up to 3 years and stored disconnected from your Google Account. A conversation can be gone from your history and still sit for years in a separate, de-linked review store, and deleting your activity does not pull it back.
| Platform | Default for deleted chats | Training-data retention | Longest separate tail |
|---|---|---|---|
| ChatGPT | Removed from systems within about 30 days | Used unless you opt out (consumer tier) | Preserved NYT-litigation logs plus flagged-account holds, beyond the 30-day default |
| Claude | Backend deletion within 30 days | De-identified, up to 5 years if you opt in | Up to 7 years for trust-and-safety scores on policy-flagged chats |
| Gemini | 18-month default, adjustable to 3 or 36 months or off | Used unless you turn off Gemini Apps Activity | Human-reviewed chats kept up to 3 years, de-linked, survive deletion |
The part no retention table shows you: memory outlives the chat
Retention tables describe raw conversation logs. They almost never describe persistent memory, which is a second copy with its own lifecycle. When an assistant reads a conversation and writes a durable fact about you into a memory store, that derived record does not die with the chat. Delete the source, and the fact it produced can stay. The summary outlives the transcript, and that is the gap the top results leave open.
This is where the ChatGPT Dreaming V3 memory update collides with deletion. The June 2026 rollout reworks how personalization is built and, by design, narrows the audit trail of what the model synthesized about you and when. A thinner audit trail weakens a simple question: if memory derived a claim from a chat you later deleted, can you find that claim and remove it on its own. The control has to act on the memory layer, not just the chat list, and less audit trail makes that harder to verify.
Retention is not one timer. It is the chat timer, the training timer, the legal timer, and the memory timer, all running on different clocks.
What deletion actually does, platform by platform
On ChatGPT, deletion clears the visible chat and starts the roughly 30-day removal clock. It does nothing to logs already preserved under the New York Times order or held on flagged accounts. On Claude, deletion removes the chat and triggers backend removal within 30 days, but it cannot pull back de-identified data already inside the training pipeline for its up-to-5-year window. On Gemini, deletion removes activity on your chosen schedule and leaves human-reviewed chats untouched, where they persist up to 3 years.
The pattern across all three is the same. Deletion is a strong control over the primary copy and a weak one over derived and legally held copies. The most effective lever is upstream: turn off training contribution wherever it exists, because that stops the long-lived secondary copy from ever being made. You cannot opt out of a court order, but you can keep a conversation out of a 5-year training store before it lands there.
Why architecture beats any single retention setting
Retention settings are policies. A company can change them, override them, or be ordered around them. Architecture is the part that holds whatever the policy says on a given day. The durable question is not only how long data is kept, but how it is partitioned and protected while it is kept: whether one user records sit isolated from everyone else, and whether they are encrypted at rest under managed keys.
That is the lens behind MemX, an external AI memory layer by Neural Forge Technologies that keeps what your assistants remember about you in one place you own. MemX is private by architecture: per-user isolation, encryption at rest backed by Google Cloud KMS, and on-device handling where it applies. It is not end-to-end encrypted and it is not zero-knowledge, and the point is narrower than a slogan. When memory lives in a store you control, retention turns into a setting you can answer, not a policy you have to read and a docket you have to track.
Settings decide how long. Architecture decides who can reach it while it is there. The second one survives a policy change.
How to check and cut your own retention today
Start with the training switch, since it controls the longest-lived copy. On ChatGPT, turn off model training in data controls if you are on a consumer plan. On Claude, leave the improve-Claude setting off to keep conversations out of the up-to-5-year de-identified pipeline. On Gemini, set Gemini Apps Activity to a shorter window or off, knowing anything already human-reviewed sits in a separate 3-year store you cannot delete.
Then treat memory as its own surface. Open the assistant memory or personalization view and prune entries directly, because deleting a chat may leave the fact derived from it. For anything genuinely sensitive, the safest move is a temporary or incognito mode that skips both training and persistent memory, so there is no long tail to clean up after.
01How long does ChatGPT keep deleted chats?
By default, OpenAI removes a deleted ChatGPT conversation from its systems within about 30 days, and Temporary Chats follow the same window automatically. The exception is the New York Times litigation: logs already preserved under the earlier court order, and logs on accounts the Times flagged, are held beyond that 30-day default.
02Does Claude store my conversations after I delete them?
Claude removes a deleted chat from history immediately and from backend storage within 30 days. If you opted in to improve Claude, de-identified copies can stay in the training pipeline for up to 5 years. Policy-flagged chats carry longer safety retention, up to 7 years for the classification scores attached to them.
03How long does Google Gemini keep my data?
Gemini Apps Activity defaults to 18 months and can be set to 3 months, 36 months, or off. Separately, chats already reviewed by human reviewers are kept up to 3 years, stored disconnected from your account, and are not deleted when you delete your activity.
04Is deleting a chat the same as deleting the AI memory of me?
No. Persistent memory is a separate, derived copy. Deleting the source conversation does not always remove the fact the assistant synthesized from it. Open the memory or personalization settings and prune those entries directly, since the summary can outlive the transcript it came from.
05Did a court order OpenAI to keep all ChatGPT chats forever?
Not anymore. The broad preservation order was lifted on October 9, 2025, ending the going-forward duty to hold consumer logs created after September 26, 2025. Data already preserved stays accessible to plaintiffs, and logs on New York Times-flagged accounts are still retained, so it is not a blanket forever hold.
The honest 2026 summary: defaults are 30 days at ChatGPT and Claude and 18 months at Gemini, but the real retention is set by three other clocks. Training pipelines hold de-identified copies for years where you have not opted out. Human review keeps a separate multi-year store at Gemini. And a court docket, not a settings page, governs an older slice of ChatGPT logs. Read the timers separately, switch off training first, and treat the memory layer as the surface that actually decides how long an assistant remembers you.
