AI Guides › Workbench
By Nigel Guy · 7 min read
The usual move goes like this. You paste a viral "tell me what you know about me" prompt into ChatGPT, it hands back five confident observations and a five-year forecast, and you read it like a horoscope written by someone who has seen your diary. It feels accurate because it is built from your own words, which is also why it isn't a judgement: it only sees what you typed.
The rule: treat ChatGPT's memory as a mirror of the chats it can see, not a judge of your life. Check what it is drawing on before you ask, make it show the evidence for every claim, and keep any conclusion you couldn't defend to a friend.
According to OpenAI's help centre (page updated September 2026), Memory is more than the list of facts you ask it to save. Depending on your plan, region and settings, it can draw on:
Three points matter for this kit. First, memory "does not retain every detail from every conversation". ChatGPT decides what counts as relevant, so a phrase like "every conversation we've had" overstates what it can see. Second, chats inside a temporary chat never create or update memories. Third, the controls sit under Settings → Personalization → Memory. Depending on your account, that screen shows Memory, Reference saved memories, Reference chat history, Memory summary and Manage/Saved memories.
| Item | What it does | Cost / free tier at time of writing | Best for | Catch |
|---|---|---|---|---|
| Memory summary | Gives a high-level view of what ChatGPT may use about you | Free users get "limited" memory and paid plans get more, per OpenAI's pricing page | Seeing the raw material before you ask anything | OpenAI says it doesn't include every detail Memory can use |
| Sources (below replies) | Shows which chat, saved memory, file or email shaped an answer | Same plans as Memory | Checking a claim against where it came from | "May not show every factor" |
| Unpersonalised temporary chat | Starts a chat that uses no memories or custom instructions | Available to all account types | A control run: what does it say with no history? | Must be chosen before the chat starts |
| The Mirror Chain (three prompts below) | Turns memory into observations, patterns and two scenarios, each with evidence | Works on any plan with Memory on | Self-reflection you can actually check | Output is only as wide as your chat history |
On price: OpenAI's help centre lists Plus at US$20 a month. The UK pricing page shows a local amount, but a £ figure couldn't be confirmed at time of writing, so check chatgpt.com/pricing before you subscribe. You don't need a paid plan for this kit.
Go to Settings → Personalization → Memory and read the memory summary. Correct anything that's wrong in the text field, or highlight it and type a fix. If something is out of date (an old job, a project you dropped), fix it now. Otherwise the chain will read it back to you as an insight.
If you want a comparison, run Prompt 1 in an Unpersonalised temporary chat as well. Anything that shows up in both versions is generic and could describe anyone.
You are acting as a careful, evidence-led observer. Your job is to help me see patterns in how I use this assistant, and you are not here to flatter me or to diagnose me.
Context: use only what you can actually access about me through saved memories and referenced chat history. If you can't see much, say so plainly.
Task:
1. List five observations about me that I have not stated outright but that my past chats suggest.
2. For each, name the kind of evidence it rests on (topic, repeated request, phrasing, timing), and do not quote anything sensitive.
3. Rate your confidence in each as low, medium or high, and explain the rating in one line.
4. Mark any observation that could just as easily describe most people who use ChatGPT for [MY_MAIN_USES].
Format: a numbered list, with each item set out as Observation / Evidence / Confidence / Generic? (yes or no).
Constraints: be direct and don't soften anything to please me. Don't make medical, psychological or financial diagnoses. Don't invent evidence.
If you need to know what I mainly use ChatGPT for and it isn't filled in, ask me before answering.
Before replying, check that every observation has evidence you could point to, and delete any that doesn't.
Fill in: [MY_MAIN_USES], for example "work emails and meal planning".
Run this in the same chat so it builds on Prompt 1.
Keep the same role and the same rules on evidence.
Goal: describe how I tend to think and make decisions, based on the observations above and anything else you can access.
Steps:
1. Name up to three recurring patterns in how I approach problems or choices (for example, how much information I gather, when I ask for reassurance, what I drop halfway).
2. For each pattern, describe one situation in my chats where it showed up, in general terms.
3. Suggest one thing I might be avoiding or underplaying. Present it as a hypothesis for me to test, not as a fact about me.
4. Say what information you would need to be more confident, and what in my chats points the other way.
Format: a short table with the columns Pattern | Where it showed up | Counter-evidence, then the hypothesis in two or three sentences.
Constraints: no labels from clinical psychology, no moralising, and no certainty you can't support.
If my history is too thin for this, say so rather than filling the gap.
Before replying, re-read the hypothesis and soften any wording that sounds like a verdict.
Nothing to fill in. It relies on the chat above it.
Keep the same role and rules.
Goal: help me think through what happens if the pattern you flagged stays as it is, compared with if I change it. These are scenarios, not predictions.
Inputs: my current situation is [ONE_LINE_SITUATION]. The pattern I choose to work on is [PATTERN_FROM_PROMPT_2].
Steps:
1. Scenario A: describe a plausible five years if nothing changes, in five short bullets (one per year).
2. Scenario B: the same, but I work steadily on [PATTERN_FROM_PROMPT_2].
3. List the three assumptions both scenarios depend on most.
4. Suggest one small action I could take in the next seven days, and one sign that would tell me it is working.
Format: two bulleted lists, then the assumptions, then the action and the sign.
Constraints: no money, health or legal forecasts presented as fact. Make it clear these are illustrative.
If either input is missing, ask for it before writing anything.
Before replying, check that Scenario B isn't just Scenario A with a happier tone. It must differ because of the pattern.
Fill in: [ONE_LINE_SITUATION] and [PATTERN_FROM_PROMPT_2], which you pick yourself, not ChatGPT.