Custom instructions and memory

0:00 Custom instructions and memory

If you keep typing the same preferences into ChatGPT, British spelling, bullet points, no invented statistics, you are wasting time. Custom instructions and memory fix that. But badly managed memory can also surface stale or sensitive details at the worst moment. In this lecture you will learn to set both up well and keep them clean.

0:24 Why personalisation matters

Why take personalisation seriously? Because there are two quiet costs. The first is repetition. If you type the same preferences into every chat, you are wasting minutes every day. The second is stale memory. If ChatGPT remembers an old client or job title, those details can surface in the wrong place at the wrong time. Think of memory like a colleague's notebook about you. Helpful when it is accurate, awkward when it is out of date. So we set it up well and keep it tidy.

1:01 Two kinds of personalisation

There are two kinds of personalisation. Custom instructions, plus an optional personality, are what you tell ChatGPT to apply to every chat. Memory is what ChatGPT learns and keeps across chats, saved memories and, where enabled, reference to your past conversations. Both live in the personalisation settings. And Projects can have project only memory, which keeps what is learned in one Project from leaking into your other work.

1:31 Good instructions

Good custom instructions have two parts. About me, for example, a freelance brand strategist in Abu Dhabi working with hospitality and retail clients in the UAE and Saudi Arabia. And how to respond. British spelling, lead with a one line answer, tables for comparisons, flag assumptions, give two options for copy, and never invent statistics or sources. Keep it specific, stable, short and non confidential. Then test it with three typical questions.

2:02 How memory works

Memory works quietly. ChatGPT may save useful details, like prefers bullet summaries or works in hospitality. You can also tell it directly, remember that my agency is called Oasis Brand Studio. You can ask, what do you remember about me, and you can view, edit or delete memories in settings. Saved memories and chat history reference can be switched on or off separately, and business admins may control them.

2:32 Temporary Chat

Temporary Chat is your clean room. It does not use or update memory, and it is not saved to your history. Use it for a sensitive one off question, or to see what an unpersonalised answer looks like, which is surprisingly useful when you want an objective comparison.

2:53 Monthly memory hygiene

Once a month, spend five minutes on memory hygiene. Delete outdated details like a former client or old job title. Delete anything sensitive, like health, finances or other people's personal data. Correct wrong assumptions, for example, I am the marketing lead, not the owner. And keep client specific context in Projects, not in global memory, so it cannot spill into other clients' work.

3:20 Simple example

A simple example. In any chat, say, remember that I prefer tables for comparisons. ChatGPT saves it. Next week you ask, compare these two phone plans, and the answer arrives as a table without you asking. That is memory working for you. Now suppose you start preferring short bullet lists instead. Open memory in settings, delete the entry, and add a new custom instruction. You are always in control of what it remembers.

3:52 Worked example: the exclusivity problem

A beauty creator in Karachi signs a six month exclusivity deal with a skincare brand. But ChatGPT keeps suggesting collaborations with competitors, because memory holds details of past partnerships. She deletes those memory entries, adds a custom instruction saying until March, do not suggest skincare collaborations other than her partner brand, and tells ChatGPT to forget the old details. Her suggestions immediately respect her contract. Two weeks later, the brand's competitor launches a campaign and a follower asks her opinion. Because her instruction is explicit and dated, ChatGPT helps her write a neutral reply that does not mention a competing product. When March comes, she deletes the instruction and asks ChatGPT to forget the exclusivity, and her suggestions open up again. The rule lived exactly as long as the contract, which is the point.

4:50 Pitfalls

Four pitfalls. Stuffing instructions with one off task details. Putting client names, confidential data or credentials in instructions or memory. Forgetting memory exists and being caught out by stale details in client work. And assuming Temporary Chat changes your employer's or clients' data obligations. It controls history and memory, not contracts or law.

5:13 What never goes in personalisation

Some things never belong in instructions or memory. Passwords, API keys and card numbers, which belong in a password manager. Client names and deal terms, which belong in a client Project with project only memory. Sensitive health or family details you would not share at work. And other people's personal data. A quick test, if a colleague looked over your shoulder while ChatGPT recalled it, would you be comfortable? If not, delete it.

5:45 Try this now

Try this now. Rewrite your custom instructions with the about me and how to respond template, under one hundred and twenty words. Then ask ChatGPT, what do you remember about me? Read the answer carefully and delete anything outdated or sensitive in settings. Finally, ask the same question once in a normal chat and once in Temporary Chat, and compare. You will see exactly how much your personalisation is shaping the answers, and whether that is helping. Put a monthly reminder in your calendar to repeat the memory check.

6:24 Instructions vs memory vs Projects

Here is the key idea that ties this together. Custom instructions describe how you like answers, for every chat. Memory holds facts ChatGPT learns about you over time. And Projects hold the context for one client or initiative. When you are unsure where something belongs, ask yourself, should this apply to every conversation, should it be learned about me, or does it only matter for this one piece of work? Put it in that one place, and nowhere else.

6:58 Watch me do it, part 1

Let me write the brand strategist's custom instructions. In the about me box, freelance brand strategist in Abu Dhabi, hospitality and retail clients in the UAE and Saudi Arabia, I write proposals, brand platforms and social strategies. In the how to respond box, British spelling, lead with a one line answer, tables for comparisons, flag assumptions, two options for copy, never invent statistics or sources, say data needed instead. I save and test with three questions, a tagline for a boutique hotel, a comparison of two loyalty programme models, and a market size question. The first two follow the rules. The third says data needed instead of inventing a number. Exactly right.

7:47 Watch me do it, part 2

Now memory. I ask, what do you remember about me? It lists my role, a former hotel client and my old day rate. I open memory settings and delete the former client and the old rate, since both could surface in the wrong proposal. Then a comparison. I ask the same question, draft a two line bio for a conference programme, in a normal chat and in Temporary Chat. The normal one uses my background and British spelling. The temporary one is generic. That side by side shows me exactly how much my personalisation is doing, and that it is helping.

8:31 Recap and next step

Recap. Custom instructions say how you want answers. Memory learns about you and needs regular review. Temporary Chat gives you a clean, unsaved room. And client context belongs in Projects. Your next step: rewrite your custom instructions using the template, ask ChatGPT what it remembers, clean it up, and set a monthly reminder.