Agile InsightsBriefing
A black and white workbench with a box of handwritten index cards, a magnifying glass on a blank card, a pencil with eraser crumbs, and a studio microphone, labelled in yellow Notes, Check, Forget and Episode 30.

Episode 30 · Primer · About 7 minutes

The prompt that shows what your AI actually knows about you (and it’s less than you might think)

Your assistant’s memory is a short file it writes about you: partial, sometimes wrong and harder to erase than it looks. How to see it, correct it and make it forget, and why at work it is a data store someone has to govern.

#data#oversight

Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.

In this episode

01Memory is notes, not knowledgeClaude, ChatGPT and Microsoft 365 Copilot keep short notes about you, mostly written automatically from your chats. As Neo Aplin puts it, it has a note, not the thing.
02Look, then check the coloursRead memory in settings, or ask for what it knows about you with red, amber and green confidence. The colours are the model’s own guess, and research finds them overconfident.
03Forgetting takes two stepsOn Claude and Copilot, deleting a chat does not remove a memory made from it. Delete the memory entry, then the chats. OpenAI may keep logs of deleted memories for thirty days.
04At work, defaults someone choseMicrosoft’s admin pages say Copilot memory is on by default, outside retention policies, with no audit log entries, though eDiscovery reaches it. Anthropic’s pages give different defaults: check your tenant.
05What to doRun the prompt, fix what is wrong, use temporary chats for sensitive work. Put memory in your AI use policy, decide each platform’s defaults, and write good practice into shared, owned instructions.
Sources and links (7)
How I AI: The prompt that shows what your AI actually knows about you · Apple Podcasts
Dr Amantha Imber and Neo Aplin, Inventium, November 2026. The episode this briefing responds to.
Use Claude’s chat search and memory to build on previous context · support.claude.com
Anthropic Help Center. How Claude memory is stored, viewed, edited and deleted, and the organisation owner controls.
Claude’s memory works everywhere, and you decide what’s in it · claude.com
Anthropic, 25 August 2026, with a later note on Team and Enterprise defaults.
Memory FAQ · help.openai.com
OpenAI Help Center. Saved memories, chat history, full removal and the thirty day window.
Manage Copilot personalization and memory · learn.microsoft.com
Microsoft Learn admin documentation (preview). Default on, retention, audit and eDiscovery.
Lost in the Middle: How Language Models Use Long Contexts · arxiv.org
Liu et al., TACL, 2023. Information at the start and end of a long context is used best, the middle worst.
Can LLMs Express Their Uncertainty? · arxiv.org
Xiong et al., ICLR 2024. Confidence a model states in words tends to be overconfident.

Checked against the Anthropic, OpenAI and Microsoft pages on 5 November 2026. The OpenAI page was read through a search index, Copilot memory is marked preview, and Anthropic’s pages disagree on business-plan defaults. Settings change often: check the original pages before quoting them.

Our fact check · 13 claims: 7 verified, 3 partly right, 1 disputed, 2 unverified

Before we record, every claim the hosts make is checked against the sources (this table was last updated 5 November 2026). Confirmed means a primary source or two independent reports; Reported means one party says so; Partly right, Disputed, Unverified and Not yet known mean what they say, and the hosts say so on air. This is the table as we left it.

  1. VerifiedYou can view and edit memoryClaude lists memory as topics you can open, edit or delete; ChatGPT and Copilot under Settings > Personalization.
  2. VerifiedMemory is built automatically from chatsClaude “saves automatically”; Copilot memory includes details “inferred from chat history”.
  3. Partly rightThe red-amber-green prompt shows knows versus assumesIt shows what is in context; stated confidence tends to be overconfident (Xiong et al., ICLR 2024), so green is not proof.
  4. Partly rightTo forget, tell it and delete the chatsChatGPT: delete both the saved memory and the chat. Claude (current memory): deleting a chat does not remove memories made from it; delete the topic. Copilot: deleting a chat does not delete saved memories; chat-history details go within seven days once every chat mentioning them is gone. Rule for all three: delete the entry and the chats.
  5. VerifiedDeleted items lingerOpenAI may keep logs of deleted memories up to 30 days; turning off chat-history reference clears what it learned within 30 days.
  6. Partly rightLong chats forget the memory given at the startLong-context loss is real, but Liu et al. (“Lost in the Middle”, TACL) find the start and end recalled best and the middle worst.
  7. UnverifiedIt is “two or three pages”No vendor figure.
  8. UnverifiedIt knows your job, not how you do itThe hosts' view; fits memory being role, projects and style rather than procedures.
  9. VerifiedClaude skips sensitive topics unless you opt in; never saves some dataHealth, religion, politics, gender identity and similar off by default; government ID and account numbers never saved.
  10. VerifiedClaude Team/Enterprise owners control memory but cannot see members' memories; turning it off deletes everyone'sOrg retention periods also apply to memory (not project memory).
  11. DisputedClaude Team/Enterprise defaultHelp page: off for each member until they turn it on. Blog update: on by default for most organisations created on or after 31 Oct 2026.
  12. VerifiedCopilot memory is on by default“No action is required to turn on Copilot memory.”
  13. VerifiedPurview retention does not reach Copilot memory; no audit entries; no restriction by content typeMicrosoft Learn, Manage Copilot personalization and memory (preview). Admins can find, export and delete it through eDiscovery (hidden mailbox folder).

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript

Linda Here is the idea. Your A I assistant's memory is a short file it writes about you, from your own chats. It is partial, it is sometimes wrong, and it is harder to erase than it looks. So check it, correct it, and at work, treat it as a data store that someone has to govern. By the end, you will know how memory works and how to see what yours holds. You will know how to make it forget, and what an organisation should decide.

Leonard This came from an episode of the How I A I podcast, with Amantha Imber and Neo Aplin. Their hook is one I recognise. You open a new chat, and the assistant greets you like an old friend. It knows your job and your dog's name. Then it hands back something painfully average, or mentions a role you left two years ago. So before we get to governance, what is memory, in plain terms?

Linda We have all had that greeting.

Linda Memory is a set of short notes the assistant keeps about you, separate from your chat history. Claude, ChatGPT and Microsoft three six five Copilot all have it. Most of it is not written by you. Anthropic says Claude saves memories automatically as you chat. Microsoft says Copilot's memory includes details inferred from your chat history. When you start a new chat, what it remembers comes along. So it might hold your role, your projects, the people you work with, and how you like things written. That is useful. But it is a summary, not a library.

Leonard And that is where the hosts' main warning comes in. Neo Aplin's point is that it knows you worked on a document last week, but not what the document said. It has a note, not the thing. He also says very long chats start to lose track of you. The research backs a version of that. A well known study found that models recall the start and end of a long conversation best. The middle comes off worst. So how do you find out what yours actually holds?

Linda A note, not the thing.

Linda There are two ways. The first is the settings page. In Claude, memory is listed as topics you can open, edit or delete. In ChatGPT and Copilot, look under personalisation. The second way is the hosts' prompt, and it is a good one. Open a new chat and type: tell me what you know about me, and give me red, amber and green on your confidence. Aplin adds one line. Tell it not to search your email, or it may go and look you up instead. One caution from me. Those colours are the model's own guess. Research on these self ratings finds they tend to be overconfident, so treat green as a claim to check, not proof.

Leonard Then the stale and wrong entries. Amantha Imber says hers had her in a job two years out of date. Aplin once asked for a mining example, and for two weeks it thought he worked in mining. Their fix is simple. Tell it, in chat, to forget something, or to change it, and look for the small memory updated notice. Or edit it in settings. And if something really has to go, they say, delete the chats that mention it as well. Is that the whole answer?

Linda Mining, for two weeks.

Linda It is close, and the vendors' own pages sharpen it. On ChatGPT, OpenAI says to delete both the saved memory and the chat where you shared it. On Claude, deleting a chat does not remove memories made from it. You delete the memory itself. On Copilot, deleting a chat does not delete saved memories either. So the rule that works everywhere is two steps. Delete the memory entry, then delete the chats. And expect a lag. OpenAI says it may keep logs of deleted memories for up to thirty days.

Leonard Two steps, then.

Leonard Which brings us to the organisation. If memory is a file the assistant writes about each member of staff, it is personal data. Who decided it should exist?

Linda Often nobody. Here are three facts from Microsoft's own admin documentation for Copilot memory, which is still marked preview. It is on by default. The organisation's retention policies do not apply to it. And memory actions create no audit log entries. Admins also cannot restrict what kind of information goes in. They can find, export and delete it through eDiscovery, which matters for privacy requests and when someone leaves. Claude is different again. Organisation owners control it, but cannot see anyone's memories. And Anthropic's own pages give different defaults for business plans. So check your own tenant, rather than assume.

Leonard That is a lot of defaults.

Leonard The hosts end on what they call the biggest blind spot, and it is their view rather than a measured finding. The assistant knows your job, but not how you do your job. It knows you write board papers, not your recipe for a good one. So it guesses from the job title, and you get average work. Their answer is to spell out what good looks like, in skills or agents. For an organisation, that is a bigger point. Know how left in each person's assistant belongs to nobody, and leaves when they do.

Linda So here is what to do. For yourself, run the prompt this week, then read the settings page and fix what is wrong. For anything sensitive, use the incognito or temporary chat. And when something must go, delete the memory and the chats. For the organisation, put memory into your A I use policy. Decide each platform's defaults on purpose. Check whether retention, audit and privacy requests actually reach it. And write the way your best people work into shared, owned instructions, not personal memory. If you'd like to know more, sign up for our podcast at agile-insights.com.au.

Leonard Thanks for listening.

More on #data

Episode 28 · Primer and discussionWhen an AI keeps the wiki, checking has to keep pace with writing.Episode 9 · PrimerYour company brain is a copy. Govern it like one.Reference 15 · PrimerThe quiet obligations: records, data ethics, digital experience and cloudEpisode 24 · PrimerRemember, retrieve, recheck, forget: governing agent memory.Episode 18 · Primer and discussionWhen agents share one memory, memory is a data flow to govern.Reference 10 · PrimerISO/IEC 5259: data quality for analytics and machine learning

Every episode tagged #data

All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series