AIMemLinkBlog
ChatGPT Memory vs Custom Instructions: The Difference
Custom Instructions are what you tell ChatGPT. Memory is what it decides to keep. Here is how the two layers work in 2026 — and when neither is enough.
What ChatGPT Custom Instructions actually are
Custom Instructions are a standing block of text you write yourself. OpenAI applies them across chats when personalization is enabled. On the web, open Settings → Personalization, turn on Enable customization, and edit the Custom Instructions field.
Free and Go plans allow up to 1,500 characters; Plus, Pro, and business tiers allow up to 5,000. Changes apply to future conversations only. OpenAI's own guidance: put explicit rules and preferences here — tone, role, format, standing constraints.
What ChatGPT Memory actually is in 2026
Memory is a separate personalization layer. ChatGPT synthesizes facts from your conversations, files, and connected apps into a rolling Memory Summary. OpenAI's June 2026 "Dreaming" update runs a background process that revises what it considers important — not only facts you explicitly asked it to save.
View it under Settings → Personalization → Memory. The summary shows highlights and a last-updated timestamp, but OpenAI notes it will not include everything ChatGPT remembers. You can still access the legacy saved memories list from the same settings page.
Side-by-side: who controls what
| Dimension | Custom Instructions | Memory |
|---|---|---|
| Who writes it | You, manually | You + the model (synthesis) |
| Stability | Static until you edit | Evolving; Dreaming auto-updates |
| Best for | Tone, format, role, standing rules | Facts learned across conversations |
| Portability | ChatGPT only | ChatGPT only |
Practical rule: durable rules belong in Custom Instructions; evolving personal facts may land in Memory — with the caveat that Dreaming can revise them without a direct prompt from you.
When to use each — and when neither is enough
Custom Instructions excel at "always respond in bullet points" or "I am a product manager working on climate analytics." Memory excels at "you mentioned you are vegetarian last month." Neither travels to Claude or Gemini, and neither gives you a complete, line-by-line record you authored yourself.
For sensitive topics, OpenAI recommends Temporary Chats, which do not read or write memory for that session.
The agency gap
Research on real ChatGPT users (arXiv 2602.01450, accepted at ACM Web Conference 2026) found that in a dataset of 2,050 memory entries from 80 users, 96% were created by the system without an explicit user command to remember. More than half contained sensitive psychological inferences. The memories were often faithful to conversations — but the authorship was not yours.
A third option: user-authored context you own
AIMemLink is a plain-text context record you write in your own words, served at a private link any AI with web access can read. Nothing is saved automatically. You inspect, edit, and rotate the link on your schedule — and the same record works across ChatGPT, Claude, and Gemini.
A common setup: paste your AIMemLink instruction snippet into Custom Instructions so ChatGPT fetches your authored record at the start of every chat, while Memory handles conversational details you are comfortable letting the model infer.
Ready to create your AI memory?
Write your context, use your private link for personal memory, and publish optional public basics when you want them discoverable by username.