Is Your AI Memory Private? What Each Tool Remembers

"Private" means three different things

• Private from the vendor: Your data is not on their servers. For all three major assistants, the answer is no — they store and process memory on their infrastructure.
• Private from you: Can you see everything stored? Partially — each offers some visibility, but synthesized memory may not show the full picture.
• Private from others: For a link-based system, anyone with the URL can read it. Unguessable ≠ encrypted.

What ChatGPT remembers and where it lives

ChatGPT stores chats, a Memory Summary, saved memories, Custom Instructions, uploaded files, and connected app data (such as Gmail on eligible plans). Data is processed on servers in the US and other jurisdictions per OpenAI's privacy policy.

• Training:Consumer content may train models if "Improve the model for everyone" is enabled. Business and Enterprise tiers are excluded by default.
• Deletion: Personal data removed within 30 days after a user deletion request.
• Sensitive info: May appear in memory if you share it. Temporary Chats do not create memories.

What Claude remembers

Claude stores chats, global memory synthesis, per-project memory summaries, and feedback. Memory synthesis is included in standard data exports. Enterprise admins can disable organization-wide memory, which deletes all user memory synthesis.

• Training: Consumer model training is opt-in via model improvement settings. Incognito chats are excluded.
• Retention: Deleted chats removed from backend within 30 days. Safety-flagged content may persist up to two years.
• Focus: Memory emphasizes work-related context — role, projects, communication and technical preferences.

What Gemini remembers

Gemini stores Gemini Apps Activity (chats), imported memory and chat history, connected-app data when Personal Intelligence is enabled, and standing instructions.

• Retention: Activity auto-deletes after 18 months by default (3, 36, or indefinite options available). A 72-hour minimum retention applies even with activity off.
• Human review: A subset may be retained up to three years, disconnected from your account.
• Imports:Data imported from ChatGPT or Claude is saved in Activity and, when activity is on, may be used to improve Google's services including model training.
• Temporary chats: Not used to train; 72-hour operational retention still applies.

Who can see your data

Vendors process memory on their servers. Gemini may use human reviewers on a subset. Claude Enterprise admins can access org-wide memory settings. None of this means your memory is visible to other users — but it is not invisible to the provider.

How to audit and delete

• ChatGPT: Settings → Personalization → Memory Summary; Manage memories for legacy list; Data Controls for export and deletion
• Claude: View and edit memory; Settings → Privacy for export; Incognito for no-training sessions
• Gemini: Gemini Apps Activity; Personal Intelligence → Memory toggle; delete individual chats or adjust auto-delete period

A privacy-first alternative: authored memory you control

AIMemLink stores plain text you write — nothing inferred behind your back. Your private link is an unguessable capability URL: anyone with the link can read your memory, but the URL is not guessable from your username alone. Rotate it from the dashboard or Telegram /rotate to revoke all prior links immediately.

Private links are unguessable, not encrypted. Do not store passwords, API keys, or authentication codes. Telegram alerts you when AIMemLink receives a request to your private link.

AIMemLink does not claim zero vendor exposure — you still share context when an AI reads your link in a chat. It gives you inspectability, authorship, and revocation that inferred vendor memory does not.

What not to put in any AI memory

Passwords, API keys, financial account numbers, regulated health or legal data, and third-party private information without consent. If you would not put it in a shared Google Doc link, do not put it in AI memory — vendor-hosted or link-based.