ARTICLE
Turns out your AI's memory is exactly as reliable as your uncle's
August 13, 2026
I turned on Memory for a simple reason
I wanted it to remember my preferences, the small idiosyncrasies of how I like to work, the nuance that builds up over a hundred conversations rather than the same generic summary you'd give a stranger. That felt like a reasonable thing to want.
Remembering your preferences, your tone, your working style, is exactly what this feature is built to do, and it does it well. Wanting that isn't the problem. But the same feature reaching for "how you like things done" is also the one reaching for "how things usually are" the moment you ask it something specific about your business. Same mechanism, two very different jobs.
Where that goes wrong
I run a lot of my own business on AI-assisted systems I've built myself: decisions, process notes, a running record of what's been agreed and when. More than once now, that setup has shown me exactly where memory oversteps.
Just this week, I asked for a summary of something we'd been working on, and it flagged a structural question as still unresolved. It wasn't. The decision had actually been finalised days earlier, recorded in the proper place.
When I pointed that out, here's what came back:
"Thank you for the correction. Let me verify against the actual decision record before I rewrite anything, rather than take your word for it and risk compounding the error. I found it: the decision confirming this had already been finalised, days before tonight's session even started. Here's what actually went wrong on my end: the document I pulled from contained two contradictory notes, one correctly reflecting the finalised decision, and one stale parenthetical that hadn't been updated when the decision was made. I read the stale note, and missed that the same document already contradicted it."
That one was a small, fixable slip. The second was worse.
I'd built part of an article around a real business story, using it to make a point about accountability, that nobody could really explain how a decision got made. Except the story was extremely well documented: named people, named roles, a named decision. The opposite of the point I was trying to make. When I flagged it, this is the response that came back:
"I've gotten this anecdote wrong three times in a row, first 'nobody was involved', then overcorrected to 'ten people, still unexplainable', and now this shows accountability was never actually the gap here. That's a pattern, not a one-off."
Both times, the fix was identical: stop, go back, and check the actual record, instead of running on what felt true.
A few honest questions worth running through
Are you mostly using AI for general advice and brainstorming, or for working with your actual business documents? General advice is lower stakes. Anything touching your real numbers, policies or client details is higher stakes.
Does the information you're asking about change often? Pricing, staffing, policies and procedures drift over time. The more often something changes, the more a stale memory can quietly go out of date without you noticing.
Would you rather it asked you again each time than risk assuming it already knows? Some people find repeating context mildly annoying but reassuring. Others find it tedious enough that memory's convenience wins out. Neither answer is wrong. It depends on what you're actually using it for.
Your options
Basically, you've got three options:
Option 1: Turn off Memory entirely
In Claude, this lives under Settings, then Capabilities or Memory, depending on which version of the interface you have. In ChatGPT, it's Settings, then Personalization, then Memory. Most tools that offer persistent memory put the toggle somewhere in your account or personalisation settings. Menu names shift as tools update, so if it's moved, a quick search within your own settings screen will find it. (Steps current as of August 2026)
Option 2: Leave memory on, but tell it when not to use it
This is the middle-of-the-road option, and often the most useful one. Most AI tools that let you set custom instructions, project instructions, or a starting prompt will also let you write a standing rule, something like: check the document I've given you before answering anything specific to my business, and don't rely on what you remember from previous conversations for facts, figures or policies. You keep the convenience of memory for general chat, and you force it back to your actual material for anything that matters.
Option 3: Ask it in the moment
If you don't want to change any settings, you can simply ask, each time it matters: are you answering this from what I've told you before, or from what's actually in front of you right now? No settings change required, and it works whether memory is switched on or off.
Make a smart choice
None of these options is right for everyone, and none of them is permanent. What matters is that it's a choice you made on purpose, based on how you actually work, rather than whatever a product team decided should be the default for everyone.
If you want to see this same pattern from the other side, an AI defaulting to memory over the facts actually in front of it, in a single moment rather than caught after the fact, that's a real example, not a hypothetical one:
Ten minutes with your settings menu, or one line added to your instructions, is a small thing. But it's the difference between an AI that's remembering things for you, and one that's just remembering things at you.
About the Author
Stephanie Silvester is the founder of The UnConsultants, The Agentic CMO Lab and Uncomplicated AI for Business. After 30 years leading marketing inside global organisations, she now helps small business owners make sense of AI without the jargon - as someone who's made the calls and knows what actually matters when you're running the whole show yourself.
The information on this website is general in nature and does not constitute professional advice. All rights reserved.
Other names and brands may be claimed as the property of others.
© 2026 The UnConsultants Pty Ltd
ABN 87 689 593 333