Suppressed or deleted? What actually happens when you remove an AI memory
You tell your AI assistant to forget something. It says done. What actually happened?
Depending on the assistant, one of three quite different things: the detail was hidden but kept, the record was deleted but can be re-learned from your chat history, or, in the best case, a row in a database is actually gone. The vendors' own documentation distinguishes these poorly, users conflate them constantly, and the difference matters most exactly when it matters at all: when the thing you removed is something you needed gone.
I went through the current help docs for ChatGPT, Claude, Gemini, and Copilot to pin down what "remove" does in each, then through what GDPR actually demands. Calmara, my product, ships both a retract and a hard delete as separate buttons, so the last section explains why I think that split is the right answer. Everything vendor-specific below is sourced from their documentation as of mid-July 2026; this area moves fast.
Quick answer: "Removing" an AI memory can mean suppression (the assistant stops mentioning it but keeps it), record deletion (the entry is removed but may be re-derived from your chat history), or true erasure (the record is gone). ChatGPT offers suppress and per-item delete, but deleted details can be re-learned unless you also delete the source conversations. Claude added a per-item delete in July 2026, alongside editing and a full irreversible reset, though it keeps no record of the deletion. Gemini and Copilot delete items on request but keep data in separate retention and improvement pipelines. No major assistant offers a documented choice between "retract but keep the record" and "erase entirely."
The three things "remove" can mean
It helps to name them, because vendor UIs use forget, dismiss, delete, and reset almost interchangeably.
Suppression stops the assistant from surfacing a detail. The underlying data stays. ChatGPT's "Don't mention this again" is the clearest example: it is a visibility control, not a deletion.
Record deletion removes the stored memory entry itself. This is what most people assume every "delete" button does. The catch is that if the assistant continuously re-learns from your chat history, deleting the derived record without deleting the source is temporary by design.
Erasure means the information is gone from everything: the memory store, the retention pipelines, and, in the strictest reading, anything trained on it. That last part is where everyone should be honest about the limits: removing knowledge already absorbed into model weights is called machine unlearning, and it remains an open research problem that is computationally expensive and tends to degrade the model. No consumer assistant offers it, and none claims to. Which is one more reason to care whether your data feeds training at all before it ever gets there.
What ChatGPT actually does
Credit where due first: ChatGPT gives you three separate controls on the memory page, and one of them is a real per-item delete. You can correct a memory by editing it, dismiss it with "Don't mention this again," or delete the entry outright from its menu. A claim you will still see repeated, that post-Dreaming ChatGPT only suppresses and cannot delete individual memories, is out of date, and I have corrected it in my own earlier piece on the Dreaming update too.
The real problem is durability. Since the June 2026 Dreaming update, ChatGPT's memory is a summary continuously re-synthesized from your chat history rather than a static list. Delete one line from the summary, and nothing stops the same detail from being re-derived from the conversations that produced it. OpenAI's guidance says as much: to fully remove something, you delete the memory entry and the source conversations and any connected files. Deletion works, but it is deletion of a derivative, and the source keeps regenerating.
Two more details from OpenAI's docs worth knowing. Deleting a conversation does not delete memories that came from it; they live separately and must be removed on the memory page. And deleted memories are kept in an internal log for up to 30 days "for safety and debugging purposes," which is a normal retention practice, but it does mean "deleted" and "gone" are 30 days apart even in the best case.
What Claude actually does
Claude's memory changed on July 10, 2026. It now stores "a set of individual entries that are organized into categories," updated in real time, replacing the daily-refreshed summary (Free, Pro, and Max are on the new model; Team and Enterprise are still migrating). You can edit it, directly or by telling Claude what to change, and you can reset it. The reset is genuinely absolute: Anthropic's help page says it "permanently deletes all memories including project memories" and "cannot be undone." That is real erasure, and it is more than some competitors offer.
Claude now does have a per-item delete: you select an entry and remove it, which was not true before July 2026 and is a real improvement. What it does not have is any record that the deletion happened. Anthropic's memory documentation states plainly that "Individual user memory edits are not logged." And the source-and-derivative problem remains: an entry removed while the conversation it came from is still in your history can be learned again, structurally the same dance as ChatGPT. The link is broken in the other direction too: as of July 2026, Anthropic's docs state that deleting a conversation does not remove the memory entries generated from it, so chat deletion is not memory cleanup either. That inversion gets its own article.
Retention behind the scenes follows Anthropic's consumer terms: deleted chats leave backend storage within 30 days by default, but if you opted into model improvement, data can persist de-identified in training pipelines for up to five years, and content flagged for safety review is retained regardless of deletion: the messages themselves for up to two years, and the safety classification scores for up to seven. The delete button and the retention schedule are different systems.
What Gemini and Copilot actually do
Gemini's "Saved info" entries persist until you delete them, and deleting a chat stops Gemini from using it for personalization. But Google's privacy hub contains a sentence worth reading twice: "Your Gemini settings don't control processing of your chats to create anonymized data to improve Google services." Deletion reaches the personalization layer, not the improvement pipeline. Retention is also nothing like a single number: 72 hours for temporary or history-off chats, an 18-month default auto-delete for activity (configurable to 3 or 36 months), and up to three years for chats a human reviewer looked at.
Copilot lets you delete saved memories, and details inferred from chat history are purged within 7 days of deleting all the source chats, or within 30 days of turning personalization off. The sharper issue with Copilot is not deletion mechanics but accountability: memory actions generate no audit log entries in Microsoft's own compliance suite, which means there is no record of what existed before you deleted it, or that you deleted it. That gap is its own article: your AI has an audit log now.
What does GDPR actually require?
The right to erasure, Article 17, clearly applies to the records: your chats, your memory entries, inputs and outputs. The EDPB's ChatGPT taskforce report noted that OpenAI steers users toward erasure rather than correction precisely because fixing an LLM's beliefs is technically hard. If a vendor cannot delete a memory entry on request, that is a straightforward compliance problem.
Model weights are where it gets genuinely contested. The Hamburg data protection authority took the position that an LLM's weights do not store personal data in a retrievable form, so the model itself is not subject to erasure. The EDPB's Opinion 28/2024 pushed back: models trained on personal data "cannot, in all cases, be considered anonymous," meaning obligations can attach to the model, case by case. Regulators are split, and anyone telling you GDPR definitively does or does not require unlearning is ahead of the law.
The practical takeaway is cleaner than the legal one: the erasure right you can actually exercise today is record-level. Which makes the quality of a product's record-level deletion, what exactly gets removed, from where, verifiably, the thing worth evaluating. It is one of the six controls I would check before approving any assistant for regulated work.
Why retract and erase should be two different buttons
Here is the design conclusion I drew from all of this, and built. Disclosure: Calmara is my product.
The vendors above each collapse two legitimate but opposite needs into one vague "forget." Sometimes a memory is wrong or outdated, and what you want is to retract it while keeping the record that it was believed, because the history is how you audit what the AI was doing. And sometimes a memory should never have existed, and what you want is erasure, the GDPR kind, where the row is gone.
In Calmara those are separate operations on every fact. Retracting a fact marks it invalidated with a reason and keeps it in the fact's revision history, visibly labeled, with its validity period and what replaced it. Deleting permanently removes the row itself, history included, and the confirmation dialog spells out which of the two you are choosing. One deliberate limitation: assistants connected over MCP can retract facts but cannot hard-delete them. Irreversible erasure is a human-only button, because "the AI erased the evidence" is not a sentence I ever want a user to say.
There is a second copy that any honest version of this claim has to account for, and I only thought it through properly in August 2026. Calmara logs every tool call an assistant makes, storing the exact input it sent, which means a fact an assistant asserted also exists inside that log entry. Deleting the fact used to leave that copy sitting there. It now erases it: the stored payload is replaced by a tombstone marking that content was erased at the user's request. The log row itself survives, deliberately. An audit log whose entries can quietly disappear is not evidence, and the surviving row still shows that the AI acted, when, from where and how it authenticated, which is the part erasure was never meant to take away. If you want the log itself gone, that is a separate, explicit action with its own button.
The retraction side is what makes the history auditable rather than just mutable: every superseded or withdrawn fact stays on the record with its dates and provenance, which is how you can ask what the AI believed last month and get an answer. The erase side is what makes the right to be forgotten real. Neither substitutes for the other, and as far as I can find, no major assistant documents offering both. That surprised me enough to double-check it. It held.
FAQ
Does "Don't mention this again" delete a ChatGPT memory?
No. It suppresses the detail so ChatGPT stops bringing it up; the memory entry persists. Deleting is a separate action on the memory page, and even a deleted detail can be re-derived from your chat history unless the source conversations are deleted too.
Can I delete one specific memory in Claude?
Yes, since July 2026. Claude stores memory as individual entries and you can select one and delete it, alongside editing and a full reset that permanently and irreversibly deletes everything. Two caveats: the deletion itself is not logged, and a fact can be re-learned from the source conversation unless you delete that too.
If I delete a memory, is it gone immediately?
Rarely. OpenAI retains deleted memories in an internal log for up to 30 days, Anthropic removes deleted chats from backend storage within 30 days, Copilot purges chat-derived details within 7 to 30 days, and Google uses windows from 72 hours to 3 years depending on the data. "Deleted from your view" and "deleted from the vendor's systems" are separated by days to years.
Does GDPR force AI companies to remove data from trained models?
Unsettled. Erasure clearly applies to stored records like chats and memory entries. Whether it reaches model weights is disputed between regulators: Hamburg's DPA says weights are not retrievable personal data, while the EDPB says models trained on personal data are not automatically anonymous. Record-level erasure is the right you can reliably exercise today.
What is the difference between retracting and hard-deleting a fact in Calmara?
Retraction invalidates the fact but preserves it in the revision history with a reason, its validity dates, and what replaced it, so the memory stays auditable. Hard delete removes the record entirely, including from history, and is only available to a human in the UI, never to a connected assistant.
---
If you are moving memory between assistants and want it to land somewhere these distinctions exist, the migration guide covers getting your memory out of ChatGPT and into a store you can audit.
Written by Dan Hagen