Fix your AI memory in 10 minutes.

Why your AI is getting worse

Every time you use ChatGPT, Claude, or Gemini, it saves details about you into a memory profile. The problem is these systems have no concept of what actually matters — a throwaway question about a pasta recipe gets the same weight as your core work preferences. Over time, your AI builds a hallucinated version of you, and every response gets a little less useful.

This is the same mistake Spotify made: listen to one random song and your Discover Weekly shifts for months. The fix is the same one Spotify users figured out years ago — clean up what the algorithm hoards, and stop feeding it junk.

Step 1: Audit your memory (find the junk)

Open your AI's memory file. Half of it is probably random garbage from months ago.

ChatGPT

Claude

Gemini

Step 2: Delete everything that doesn't belong

Be ruthless. If a memory isn't actively helping you get better responses, it's actively making them worse. Delete anything that's:

• Outdated (old projects, former preferences)
• Random (one-off questions that got saved)
• Wrong (things it misinterpreted)
• Irrelevant (it doesn't need to know you asked about sourdough bread once)

How to delete:

• ChatGPT: click the trash icon next to each memory in Manage Memories, or tell it "Forget that I mentioned [X]".
• Claude: click a memory item → delete, or tell it in chat "Forget what I said about [X]".
• Gemini: three-dot menu next to each entry → Delete, or tell it "Forget that I mentioned [X]".

⚠️ On ChatGPT, deleting a chat does NOT delete the memories it created. You have to delete memories separately.

Step 3: Put what matters in custom instructions

The memories you actually DO want your AI to know? Don't let it save them automatically. Put them in Custom Instructions instead. Custom Instructions are a separate system — they don't get polluted by random conversations and they stay exactly as you wrote them, every single time.

• ChatGPT: Settings → Personalization → Custom Instructions. Two fields — "What would you like ChatGPT to know about you?" (your role, expertise, preferences) and "How would you like ChatGPT to respond?" (tone, format, length, style).
• Claude: create a Project → add Project Instructions. Every conversation in that project uses your instructions — no memory pollution.
• Gemini: Settings → Personal context → "Your instructions for Gemini" → Add. Write standing instructions that apply across all chats. Direct link: gemini.google.com/personal-context.

Here's a template — fill in the blanks and paste it into your custom instructions:

Step 4: Use temporary chat for one-off questions

This is the one nobody talks about. Every random question — a recipe, a quick fact check, a one-off coding problem — feeds into your memory profile and dilutes it. The fix: Temporary Chat. Think of it like incognito mode for your AI — it doesn't read your memories, and it doesn't create new ones.

• ChatGPT: click the model dropdown at the top → select "Temporary Chat", or use the temporary chat toggle when starting a new conversation.
• Claude: click the ghost icon 👻 in the upper-right corner when starting a new chat.
• Gemini: start a Temporary Chat from the new chat menu — it won't reference your memory or save to your Activity.

Use it for cooking recipes, quick searches or fact-checks, one-off coding problems, travel research for a trip you'll take once — anything you wouldn't want influencing your AI permanently.

Step 5: Repeat monthly

None of these companies have built memory decay yet. Until they do, your AI will keep hoarding every random thing you've ever said. Once a month, run the checklist:

• Open memory settings on each platform you use.
• Delete outdated or irrelevant memories.
• Check for contradictory entries (it happens more than you think).
• Update your Custom Instructions if your work or preferences changed.
• Set a reminder for next month.

Pro tip: set a recurring calendar reminder — first of every month, 10 minutes. That's all it takes.

Why this works

AI memory pollution is a system-level problem — you can't fix it with a better prompt. But you can work around it:

Quick reference

Original guide by Artem Novitckii. More from him on Instagram, in the Skool community, and across his resource library.