ChatGPT Memory & Custom Instructions: Settings That Stick
Why Custom Instructions Beat Repeating Yourself
Every time you open a new ChatGPT chat and re-type “I’m a marketer, keep it short, use US English,” you burn tokens and patience. Custom Instructions and Memory are the settings that make those preferences stick across conversations. Used well, they turn a generic bot into one that already knows your role, your tone, and your no-go zones. Used badly, they pile on so many rules the model gets confused and ignores half of them.
This is a settings-selection guide. The table below maps what you want ChatGPT to remember to the right control — Custom Instructions, Memory, or neither. Then each row has a copy-paste snippet. For the difference between system and user prompts that also shapes behavior, see system vs user prompt, and for dozens of reusable role frames check 40 system prompt templates.
Memory Settings Decision Table
| What you want to stick | Control | Stays for | Use when |
|---|---|---|---|
| Your role & audience | Custom Instructions (About you) | All chats | You always want answers framed for your job |
| Tone & format rules | Custom Instructions (How to reply) | All chats | You always want the same voice and length |
| Facts from this project | Memory (saved) | Until deleted | A detail matters across many future chats |
| One-off context | Neither (in-chat only) | This chat | It will not recur; don’t clutter memory |
| Hard guardrails | Custom Instructions (How to reply) | All chats | You never want certain outputs (e.g., no emojis) |
When to Use Each Control
Custom Instructions — About you. Put your role, industry, and who you write for here. It is the cheapest win: every answer starts already knowing you are a B2B SaaS founder or a biology grad student. Set it once.
Custom Instructions — How to reply. Put durable style rules here: length, language, formatting, and guardrails. This is where “never use emojis” or “always give a table” lives. Keep it to a handful of rules or the model starts dropping them.
Memory (saved facts). Use for project-specific truths ChatGPT should recall — your product name, a client’s brand colors, a recurring deadline. Memory is editable, so treat it like a notebook, not a contract. Review it monthly.
Neither (in-chat only). Throwaway context — a one-time document, a temporary task — belongs in the chat, not in memory. Saving it bloats future answers with irrelevant detail.
Hard guardrails. If there is something you never want (no markdown, no unrequested code, no medical advice), state it in “How to reply.” Guardrails in Custom Instructions beat hoping the model remembers mid-conversation.
Prompt Snippets for Each Setting
Custom Instructions — About you
I am a [ROLE] at a [COMPANY TYPE] working on [MAIN GOAL].
My audience is [WHO I WRITE/SPEAK FOR].
I often need help with [TASK 1], [TASK 2], and [TASK 3].
Assume this context in every chat unless I say otherwise.Custom Instructions — How to reply
Reply rules I want every time:
- Keep answers under [N] words unless I ask for detail.
- Use [US / UK] English and a [CASUAL / PROFESSIONAL] tone.
- Prefer tables or short lists over long paragraphs.
- Do not use emojis.
- If a request is unsafe or unclear, say so instead of guessing.Memory — save a project fact
Remember this for future chats: Our product is called [NAME], it helps [AUDIENCE] do [OUTCOME], and our brand voice is [VOICE]. Use this whenever I ask for copy or ideas about our product.In-chat only (do not save)
For this chat only: summarize the attached document and list 5 open questions. Do not save any of this to memory; it is a one-time task.Hard guardrail
Hard rule for all replies: never invent statistics or citations. If you do not know a number, say "I don't have that data" instead of estimating. Also avoid starting answers with "Sure" or "Of course."Common Mistakes With Sticky Settings
1. Overloading Custom Instructions. Twenty rules means the model follows none reliably. Keep “About you” to three facts and “How to reply” to five rules maximum.
2. Saving throwaway context to Memory. A one-time document is not a permanent fact. Use in-chat context for those so Memory stays clean and future chats are not weighed down by irrelevant detail.
3. Forgetting to review Memory. Project facts change — a client rebrands, a deadline passes. Set a monthly reminder to open Memory and delete what is stale, or old answers will quietly contradict new work.
4. Putting guardrails in the chat instead of Instructions. “No emojis” typed once vanishes in the next chat. Hard rules belong in “How to reply” so they apply everywhere without repeating.
Quick Setup Checklist
| Step | What to do | Where |
|---|---|---|
| 1 | Write role, audience, recurring tasks | About you |
| 2 | Set length, language, format, guardrails | How to reply |
| 3 | Save durable project facts | Memory |
| 4 | Monthly: prune Memory & Instructions | Settings |
The Trade-Offs
Custom Instructions are global and silent — great for consistency, risky if you switch contexts (the same “short and casual” rule may hurt a formal report). Memory is flexible but can drift; old project facts linger and contradict new work. The fix is a quarterly cleanup: open Memory, delete stale facts, and re-read your Custom Instructions to cut rules you no longer need. Less is more — five sharp instructions beat twenty vague ones. The role-framing technique in role prompting is a good companion when you want to override the default voice for a single task.
Sources & Notes
Settings described reflect ChatGPT Custom Instructions and Memory as available on free and Plus tiers in October 2026 (GPT-4o / GPT-4.1). The “About you” and “How to reply” fields and the saved-memory feature may differ by account region and rollout. Snippets are plain text the reader pastes into the Custom Instructions boxes or chat; no API calls were made. Slug and category mapping follow the guunn.com AI Prompts plan (ChatGPT = 961).