How do I delete one specific thing ChatGPT remembers about me?

Open ChatGPT → Settings → Personalization → Manage Memory. You'll see a list of stored entries. Click the trash icon next to the one you want gone. It removes that entry, but doesn't prevent the model from re-inferring the same fact from past chats.

How do I delete a specific fact in Claude?

Claude doesn't have a long-term Memory store the way ChatGPT does. Facts persist via Projects (a folder-level notes file you control) and Custom Instructions. To delete, open the relevant Project and edit the project knowledge directly, or update your global Custom Instructions. There's no per-claim audit trail.

How do I delete a specific fact in Gemini?

Gemini Apps Activity holds your conversation history; Memory is rolled out gradually and surfaced differently per region. Open myactivity.google.com → Gemini Apps to delete specific past conversations. The model's stored personalization is less transparent than ChatGPT's, so deletion is more about wiping source chats than editing claims.

What does 'surgical delete' actually mean?

Surgical delete means removing one claim while preserving everything else — and being able to see what that claim was, where it came from, and what the assistant will stop saying as a result. Native provider tools let you delete entries, but not surgically: you can't see the link from a bad reply back to the entry that caused it.

Should I just clear all my AI memory and start fresh?

Sometimes. If you've radically changed jobs, moved cities, or want a clean slate, nuking is fine. The cost is months of personalization. Most people do better with a periodic surgical cleanup than a full reset every quarter.

What's the GDPR / privacy angle here?

Under GDPR and CCPA you have the right to deletion of personal data held by the provider. OpenAI, Anthropic, and Google all offer data-deletion requests in their privacy portals. That's broader than the in-product Memory panel and is the right tool for 'please erase everything about me from your servers.'

How does Konshus handle surgical delete?

Every stored fact is an atom with a source and an audit ID. You can delete one atom and see which replies will change as a result. The deletion is logged (auditable) and you can hard-delete, soft-delete, or supersede with a corrective atom. That per-claim editing is the Brain Surgery patent pending claim (/patent).

A practical guide · ~8 min read

How to Delete Specific Things ChatGPT, Claude, and Gemini Know About You

Most "how to delete ChatGPT memory" guides tell you to wipe everything and start fresh. That's the bazooka option. Most of the time you don't want a bazooka — you want a scalpel for the one weird thing the assistant keeps bringing up. Here's how to do it per provider, and what the per-claim alternative looks like.

ChatGPT (OpenAI)

Open ChatGPT in a browser → click your name → Settings.
Go to Personalization → Memory → Manage.
You'll see a list of memory entries. Click the trash icon next to any one to delete it individually.
To wipe everything: Clear ChatGPT's memory button at the top.
For a full data deletion request (server-side, GDPR/CCPA): privacy.openai.com → request data deletion.

What this doesn't do: it doesn't stop the model from re-inferring the same fact next time you start a chat that touches the topic. The entry is gone; the pattern isn't.

Claude (Anthropic)

Claude doesn't have a long-term Memory store like ChatGPT does — most "what Claude knows about you" lives in Custom Instructions and Projects.
Open Settings → Profile to edit global Custom Instructions.
Open each Project and edit the project knowledge / instructions to remove the specific fact.
For server-side deletion: privacy.anthropic.com → submit a data deletion request.

Claude's model is less leaky between sessions, which means surgical deletion is more tractable — but only if you've kept your context in Projects rather than relying on the model's working memory across chats.

Gemini (Google)

Open myactivity.google.com → Gemini Apps Activity.
You can delete individual past conversations or set Auto-delete (3, 18, or 36 months).
Gemini's personalization (where rolled out) is surfaced in Settings → Personalization.
For broader account-level deletion: takeout.google.com to export, then delete via your Google Account → Data & privacy.

Gemini's memory layer is the least transparent of the three, and changes by region. If a specific weird claim keeps coming up, the most reliable lever is deleting the conversation that taught it.

Short answer

Each major provider lets you delete individual memory entries or past conversations, but none of them shows you the link between a bad reply and the entry that caused it. Per-claim editing — every fact pinned to its source so you can surgically delete one without losing the rest — is the Brain Surgery patent pending claim (/patent).

Never lose your AI again

Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.

Meet Konshus

When to nuke instead

Surgical delete is the right default. But there are cases where wiping makes sense:

Major life change. You've moved countries, changed careers, ended a long relationship. Half your stored memory is now stale; rebuilding from a blank slate is cleaner than editing one at a time.
A breach or shared account. Someone else used your account and you don't trust what's in there. Wipe and restart.
You're leaving a provider. Export, then submit a server-side deletion request. The in-product Memory panel doesn't reach the backups; the privacy portal does.

For everything else, scalpel beats bazooka. (We argue the same point about contradictions specifically in why your AI contradicts itself.)

The per-claim alternative

A memory layer built on atoms with sources makes deletion almost boring. The assistant says something weird; you click the reply, see which atoms it used and which source they came from, and remove or supersede the offending atom. The audit log shows the delete. Next reply, the claim is gone. No guessing, no nuking, no months of re-onboarding.

That's the bar. We think it should be the default. Until it is across providers, surgical cleanup via the steps above is the realistic best practice.

Frequently Asked Questions

Delete one thing, keep the rest

Konshus pins every stored fact to its source so you can surgically delete a single claim — with a real audit trail. Brain Surgery, patent pending.

Meet Konshus