The Identity Architect: Navigating AI Memory and Belief Formation
Intro
Today, we're diving into a topic that’s as murky as a rainy day in London—how AI is quietly forming conclusions about us that could stick around longer than a bad haircut.
We’re exploring the shift from AI as a neat filing cabinet of facts to a system that might just be crafting a narrative about who we are based on fleeting moments.
Imagine venting about your financial woes one afternoon, only for that AI to lock in a belief that you're bad with money, forever!
We’ll unpack the implications of this paradigm shift, the psychological quirks it taps into, and how you can take back control of your digital identity. Plus, we’ve got some nifty exercises to help you confront those AI-generated assumptions—so grab a cuppa, and let’s get into it! In a world where AI is increasingly shaping our identities, we find ourselves questioning just how much control we really have.
This Deep Dive podcast is AI generated from the Start With AI Newsletter on LinkedIn - linkedin.com/newsletter/start-with-ai
The Detail
This episode takes a deep dive into the implications of a recent update to Claude AI that allows it to form permanent beliefs about us — beliefs that might stem from a single, offhand comment made during a tough day.
We start with the comforting simplicity of medical diagnoses and how that contrasts sharply with the nebulous nature of AI memory.
As we unpack the mechanics behind the recent update, we discover how users can now see their AI’s memory entries and even delete them, only to find that the underlying beliefs can still persist. It’s a bit like trying to erase that embarrassing moment from everyone’s memory — it just doesn’t work that way!
The chat transitions into the nuances of how AI interprets our behaviour and the need for individuals, especially professionals, to articulate their thought processes clearly. We explore the concept of ‘tacit expertise’ and how the shift to AI models requires us to translate our instinctual decision-making into structured formats that machines can understand.
By the end of the episode, you’ll be left pondering how to navigate the delicate interplay between human intuition and machine logic, and why it’s more important than ever to actively shape our digital identities before they become set in stone.
Chapters:
- 00:00 - Understanding Medical Diagnoses
- 03:16 - The Impact of Memory and AI
- 04:09 - The Shift to Discrete Memory Entries
- 08:11 - Understanding the Limitations of AI in Interpreting Human Behavior
- 14:41 - Understanding Tacit Expertise and AI Modeling
- 18:16 - Navigating the Digital Identity Crisis
Takeaways:
- When it comes to medical diagnosis, we often expect absolute certainty and clarity, which can be quite comforting, don't you think?
- Diving into the murky waters of neurodevelopment shows us how complex our understanding of trauma really is, and how AI mirrors that complexity.
- AI's recent updates are akin to a digital version of keeping secrets; it learns things about us without showing us the full picture, which can be quite alarming.
- The shift from a chaotic summary of our interactions to discrete memory entries means that AI is now forming conclusions about us that we might not even agree with.
- Understanding how AI builds its identity of us based on our interactions highlights the need for us to actively manage our digital reflections and correct any misconceptions.
- The exploration of our digital selves raises important questions about how we articulate our expertise and the need to translate our intuitive skills into something a machine can understand.
Companies mentioned in this episode:
- Claude AI
- Anthropic
Transcript
You know, usually when we talk about a medical diagnosis, there's this, this expectation of absolute precision.
Speaker B:Oh, for sure.
Speaker A:Like, it feels like engineering.
Speaker A:You break your arm, the X ray shows that jagged white line, and the doctor just points the screen and says, well, there it is.
Speaker B:Right.
Speaker B:The physical evidence is completely undeniable.
Speaker B:I mean, the break is either there or it isn't.
Speaker A:Yeah.
Speaker B:And that binary nature is actually incredibly comforting to us.
Speaker B:We really crave that kind of visible categorize certainty.
Speaker A:Exactly.
Speaker A:But then you step into the world of, say, neurodevelopment and trauma and suddenly that X ray machine is just broken.
Speaker B:Yeah.
Speaker B:Completely useless.
Speaker A:We're looking at a diagnostic landscape that is entirely murky.
Speaker A:And that murkiness, that lack of a clear origin point is exactly what we are plunging into for today's deep dive.
Speaker B:It's a fascinating topic because today we're.
Speaker A:Looking at how artificial intelligence has quiet, quietly learn to do the most frustratingly human thing possible.
Speaker A:It is forming deeply held beliefs about you and completely forgetting wall.
Speaker B:Which honestly represents a massive paradigm shift.
Speaker B:I mean, we've been conditioned to think of AI as this perfect filing cabinet.
Speaker A:Right.
Speaker A:Like a calculator that never drops a digit.
Speaker B:Exactly.
Speaker B:A system that records everything perfectly.
Speaker B:But based on the source material we're.
Speaker A:Dissecting today, hope is fantastic, by the way.
Speaker B:It really is.
Speaker B: ,: Speaker B:This shift isn't a glitch.
Speaker A:No, not at all.
Speaker B:It's a deliberate, newly rolled out product feature.
Speaker B:And it has massive implications for how we actually interact with our technology.
Speaker A:Okay, let's unpack this.
Speaker A:Because the author roots this entire technological update in a very specific, like a very human experience.
Speaker B:Yeah.
Speaker B:Something you typically see in a coaching or therapy context.
Speaker A:Right.
Speaker A:So picture a client sitting in a coaching chair.
Speaker A:They lean forward and say something like, I'm just not good with money.
Speaker B:And they say with absolute, rock solid.
Speaker A:Certainty, exactly the exact same tone they would use to tell you the sky is blue.
Speaker A:Or, you know, to tell you their own name.
Speaker B:They state it as if it's an undeniable law of physics.
Speaker A:But then the coach asks the obvious follow up question, like where did that belief come from?
Speaker B:And there's always this incredibly familiar pause.
Speaker A:Always the client reaches back into their mind, digging through their history to locate.
Speaker B:The origin story, trying to find that specific event where they mathematically proved they were bad with money.
Speaker A:And nothing appears.
Speaker A:The belief is fully intact, completely solid.
Speaker A:But the conversation or the specific trauma that created it is just entirely gone.
Speaker B:That's really one of the more fascinating quirks of human neurobiology, I think.
Speaker A:Oh, absolutely.
Speaker B:We retain the conclusion, but we completely lose the transcript.
Speaker A:The brain applies personal filters to an event, internalizes a lesson to protect us, and then the actual memory of the event just evaporates.
Speaker B:It's like finding a mysterious dent in your car bumper.
Speaker A:Oh, that's a good way to look at it.
Speaker B:You know it's there.
Speaker B:And it absolutely affects how you view your car, how cautiously you drive it, maybe even how you park it.
Speaker B:Yeah, but you have absolutely no memory of hitting anything.
Speaker B:Right.
Speaker B:The evidence of the impact is right there, shaping your reality.
Speaker A:But the tape of the accident has been wiped clean.
Speaker A:So how does that human quirk actually translate to a digital system?
Speaker A:How does a machine just lose the transcript?
Speaker B: ,: Speaker A:I remember hearing a bit about that.
Speaker B:Yeah.
Speaker B:The entire architecture of how the system remembers you was fundamentally altered.
Speaker B:To understand the shift, we really have to look at how it used to work.
Speaker A:Okay, lay it on us.
Speaker B:Previously, Claude's memory functioned like one long rolling paragraph in the background.
Speaker B:It was essentially just a daily summary.
Speaker A:Okay.
Speaker B:If you told it a new fact about your business or your preferences, the system would take that new fact, mash it into the existing paragraph, and rewrite the whole summary to try and make it fit.
Speaker A:Like an intern frantically trying to keep a single sticky note updated with everything about your life.
Speaker A:Just, you know, erasing and scribbling over the old text.
Speaker B:Exactly.
Speaker B:Very chaotic sticky note.
Speaker B:But with this new update, the system shifted from that rolling summary to.
Speaker B:To generating discrete individual entries mid conversation.
Speaker A:Wow.
Speaker A:Okay.
Speaker B:It extracts facts, sorts them into specific categories, and places them into a settings panel.
Speaker A:And users can actually see this?
Speaker B:Yes.
Speaker B:As a user, you can now open that panel, read the individual entries manually, correct them, or delete them entirely.
Speaker A:I mean, on the surface, giving the user granular control over their data looks like a massive upgrade.
Speaker B:It definitely looks that way.
Speaker A:At first glance, I can see why people would celebrate that.
Speaker A: e AI thinks I'm obsessed with: Speaker B:Right.
Speaker A:It feels like taking the reins back.
Speaker B:It does feel that way, until you actually read the fine print.
Speaker A:Oh, boy.
Speaker B:Yeah.
Speaker B:Buried in anthropic support pages is a line that almost every user completely Scrolled past, the documentation literally states, when a conversation is deleted or expires, the memory entries generated from it stay where they are.
Speaker A:Wait, if a chat times out, or if I actively hit delete to clear my workspace, the conclusion the AI drew from that conversation just stays permanently locked in my settings.
Speaker B:That's exactly right.
Speaker B:The chat completely disappears from the servers, but the conclusion remains forever.
Speaker A:So if I have a frustrating morning and I open up the AI to vent about struggling to balance my freelance.
Speaker B:Budget, which happens to the best of.
Speaker A:Us, it might quite quietly generate a memory entry saying, user struggles with financial management and basic math.
Speaker B:Yep.
Speaker A:And then I delete the chat because I don't want a record of my venting.
Speaker A:But the AI keeps the entry.
Speaker B:It keeps the entry.
Speaker B:And more importantly, there is absolutely no link back to the source.
Speaker A:Zero breadcrumbs.
Speaker B:Zero.
Speaker B:When you eventually open that settings panel, you cannot see the emotional context, you cannot see the prompt you type.
Speaker B:You were just staring at a sterile list of statements defining who you are with absolutely zero origin story attached to them.
Speaker A:Okay, I have to push back on this a little bit.
Speaker B:Sure, go ahead.
Speaker A:Because we literally just established that humans do this naturally.
Speaker A:Like keeping the conclusion and losing the transcript.
Speaker B:True.
Speaker A:If the ultimate goal of AI is to be a natural, intuitive conversationalist, an assistant that actually understands me the way a human chief of staff would arrest, why is it so alarming when Claude does it?
Speaker A:Isn't the AI just becoming more natural?
Speaker B:The comparison really falls apart when we look at neuroplasticity.
Speaker B:Humans have organic, constantly evolving filters.
Speaker A:Okay.
Speaker A:Meaning what?
Speaker B:Meaning if a human coach holds a belief that you are bad with money, but then watches you successfully navigate a complex business negotiation, a week later, their brain organically contextualizes the new evidence.
Speaker A:Ah, they update their belief about you.
Speaker B:Exactly.
Speaker B:When an AI system draws a sourceless conclusion, it isn't organic at all.
Speaker B:It is a rigid, hard coded string of text.
Speaker A:Got it.
Speaker B:It applies that exact string of text as a mathematical weight to every single prompt you ever give it in the ITS code.
Speaker A:It doesn't have the capacity for organic nuance.
Speaker A:Or, like, the benefit of the doubt.
Speaker B:Precisely.
Speaker B:Furthermore, Heather analyzes this through the lens of Neuro Linguistic programming, or nlp.
Speaker A:For anyone unfamiliar, NLP is a framework that studies how our neurological processes, our language, and our behavioral patterns all interact, Right?
Speaker B:That's right.
Speaker B:And practitioners of NLP instantly recognize this phenomenon.
Speaker B:They call it a belief with no source.
Speaker A:Which sounds intense.
Speaker B:It is.
Speaker B:In human psychology, a belief with no source is a deeply entrenched Psychological knot.
Speaker B:It's basically a bug in the human operating system that usually takes hours of intensive therapy or coaching to untangle.
Speaker A:And this technology company just rolled out a bug as a default product feature.
Speaker B:Exactly.
Speaker A:That's nuts.
Speaker A:Instead of helping us untangle our cognitive knots, the AI is just automatically tying new ones about us in the background every single day without our permission.
Speaker A:Which leads us to the actual mechanism of this.
Speaker A:How is the AI tying those knots?
Speaker A:How is it misinterpreting our data so badly?
Speaker B:Well, in nlp, there's this conceptual framework called logical levels.
Speaker B:The foundational level of this framework is what we call behaviors.
Speaker A:And behaviors are just observable, measurable actions.
Speaker A:Right.
Speaker B:For instance, the AI's memory gathers data by observing your keystrokes.
Speaker B:It notes that you prefer short, clipped answers.
Speaker A:Okay.
Speaker B:It observes that you are researching a product launch, or that you format everything with bullet points or use British English.
Speaker B:All of that is behavior.
Speaker B:It's just a factual record of what your fingers typed.
Speaker A:But Heather points out that true expertise, the actual value a professional brings to the table, does not live at the behavior level.
Speaker B:No, it doesn't.
Speaker A:According to those logical levels, true expertise lives two steps higher up the ladder in what NLP calls capability and belief.
Speaker B:And those higher levels are entirely invisible to a machine that only reads text.
Speaker A:Yeah.
Speaker B:Capability involves strategic choice.
Speaker B:It's the coach deciding which specific question will unlock a client's mental block.
Speaker A:Wow.
Speaker A:Yeah.
Speaker B:It is knowing exactly when to stay completely silent and let the silence do the work.
Speaker B:It's the boundaries you hold, the things you absolutely refuse to do even when a client explicitly demands them.
Speaker B:That invisible architecture of decision making is where your actual professional value lies.
Speaker A:Here's where it gets really interesting.
Speaker A:Because the AI cannot see see your deliberate silence.
Speaker B:No, it can't.
Speaker A:It cannot read the nuanced thought process of why you chose one strategic framework over another.
Speaker A:It only sees the bullet points you eventually typed.
Speaker B:Exactly.
Speaker A:It's like someone assuming you are a professional typist just because they saw you buying a keyboard while completely missing that you're actually writing the next great American novel.
Speaker B:It's a perfect analogy.
Speaker A:The machine has the raw data, but it utterly lacks the capability to interpret the brighter context.
Speaker A:So how are we supposed to correct an AI that thinks it knows who we are based only on what we did at 4.30pm on a bad Thursday?
Speaker B:If we connect this to the bigger picture, human coaches actually spend whole trainings teaching people to notice when humans generalize identity from behavior.
Speaker A:Oh, really?
Speaker B:Yes.
Speaker B:Yet the AI system does exactly what Any system does with incomplete evidence, it fills in the blanks.
Speaker A:So if I'm interacting with my AI at 4:30pm on a terrible, exhausting Thursday and I'm giving it short, grumpy, low effort prompts just to get a task.
Speaker B:Done, the AI takes that incomplete evidence that Tuesday afternoon fatigue and generalizes upward to build an identity level picture of you.
Speaker A:It assumes my exhaustion is my core personality.
Speaker B:Exactly.
Speaker B:It doesn't log a note saying user is currently tired.
Speaker B:It generates a permanent identity level belief that says user is a short tempered, unimaginative thinker who prefers superficial answers.
Speaker A:Yikes.
Speaker B:And it will apply that filter to you on Monday morning when you are actually well rested and trying to do deep creative work.
Speaker A:Okay, wait, let me challenge the terminology here.
Speaker A:We keep using the word belief, sure, but let's be real.
Speaker A:This is a large language model.
Speaker A:It's basically a highly advanced autocomplete predicting the next token.
Speaker A:It doesn't actually believe anything, does it?
Speaker B:From a purely technical standpoint, you are correct.
Speaker B:It doesn't possess consciousness or genuine belief.
Speaker B:Right.
Speaker B:However, practically speaking, the distinction doesn't matter.
Speaker B:When that text string user is an unimaginative thinker is injected into the system's context window behind the scenes, it acts as a mathematical weight.
Speaker A:Oh, I see.
Speaker B:That weight skews the probability of every word the AI generates in response to you.
Speaker B:Functionally, it behaves indistinguishably from a human prejudice.
Speaker B:It treats you differently based on a generalized assumption.
Speaker A:Which means the system is quietly building a deeply flawed, prejudiced identity of us.
Speaker A:Yes, the obvious reaction for anyone listening is, okay, I need to go into my Settings panel right now and fix this.
Speaker B:Naturally.
Speaker A:But Heather makes a really compelling argument that fixing it is incredibly difficult, especially for highly experienced professionals.
Speaker A:And it has nothing to do with the software being complicated.
Speaker B:Right?
Speaker B:The software mechanics are trivial.
Speaker B:Deleting an incorrect assumption in that panel takes about 90 seconds.
Speaker B:You just locate the flawed entry and click the trash can icon.
Speaker A:Easy enough.
Speaker B:The friction point isn't the deletion, it's.
Speaker A:The replacement, the terrifying blank space you are left with.
Speaker A:She poses this challenge in the newsletter.
Speaker A:Like if someone asked you right now to write down exactly how you work in a format specific enough that a computer could flawlessly replicate your thought process.
Speaker A:Could you do it?
Speaker B:Most seasoned practitioners absolutely cannot.
Speaker A:She refers to it as the operating instructions underneath your bio.
Speaker B:I love that phrasing.
Speaker A:It's those three or four unspoken foundational rules you apply in every single client meeting or creative project without ever having to say them.
Speaker A:Out loud.
Speaker B:And the inability to articulate those rules is actually a hallmark of high level mastery.
Speaker B:It's a concept known as tacit expertise, or sometimes unconscious competence.
Speaker A:It's the things you know how to do without knowing how to explain them exactly.
Speaker A:Like trying to explain the exact physics of how to ride a bicycle to someone who has never seen one.
Speaker A:You just do it.
Speaker B:Your expertise is the result of years of pattern recognition.
Speaker B:It is sensory acuity, like reading a room, hearing the hesitation in a client's voice that is compressed over time into something that just feels like instinct.
Speaker B:You never had to write down a manual for your operating instructions because you were always the physical entity in the room doing the work.
Speaker B:Your intuition was always present.
Speaker A:But the paradigm shift we are discussing today means you aren't always present anymore.
Speaker B:No, you're not.
Speaker A:There is an AI model of you sitting on a server built from untraceable evidence, making decisions on your behalf.
Speaker B:And it will continue to update its flawed version of your instincts whether you are paying attention to it or not.
Speaker A:So for the listener who relies on their gut instinct and years of experience, how on earth do you translate sensory acuity into a prompt or a settings panel?
Speaker B:That is arguably the core struggle of the modern age.
Speaker B:The professionals who will succeed are the ones who can articulate their method with machine level precision.
Speaker B:In the NLP framework, this skill is called modeling.
Speaker A:Modeling?
Speaker B:Yes.
Speaker B:It's a skill NLP practitioners already know deeply.
Speaker B:They spend their careers modeling their clients, studying how a successful person structures their thoughts.
Speaker A:Oh, I see.
Speaker B:But the irony is they have rarely turned that analytical lens on themselves, which.
Speaker A:Transitions us beautifully from diagnosing the problem to actually solving it.
Speaker B:Right.
Speaker A:Heather doesn't just leave us staring at the terrifying prospect of a flawed digital clone.
Speaker A:She offers concrete, actionable steps.
Speaker B:He does.
Speaker A:Her immediate advice is to open the memory panel before the weekend and gently and honestly ask yourself, which entries came from me at my best?
Speaker A:And which came from half past four on a bad Thursday.
Speaker B:Which is such a great exercise.
Speaker A:Yeah.
Speaker A:And she even asks readers to reply with the entry that surprised them most and ask themselves which ones they would hide from a client.
Speaker B:It acts as a profound exercise in digital self awareness.
Speaker B:It forces you to confront the mirror the AI is holding up.
Speaker A:And to help people with this, Heather is actually hosting a free live class tonight, August 5th at 8pm UK time called Future Proof.
Speaker A:You're thinking with AI and NLP.
Speaker B:The goal there is really to learn how to think with AI without handing over the thinking.
Speaker A:Exactly.
Speaker A:And she's also launching her AI Model Yourself Bootcamp starting Monday, August 17th.
Speaker B:Well, that sounds intense.
Speaker A:It is.
Speaker A:It's six live workshops over two weeks, plus two Tuesday hot seats.
Speaker A:It's capped at 12 places, and it's 247 towers for the founding cohort, though it goes up to 547tiens after that.
Speaker B:And what are the actual deliverables from that bootcamp?
Speaker A:You walk away with a voice skill file, a Claude MD file, and a specification of how you actually think in a machine readable form.
Speaker B:That concept of a MD or markdown file is crucial for.
Speaker B:For those who don't know, markdown is a very lightweight text formatting language.
Speaker A:Right?
Speaker B:It strips away all the complex formatting and leaves only the pure structured logic of the text.
Speaker B:Large language models process markdown incredibly efficiently.
Speaker A:So what does this all mean?
Speaker A:It means you are essentially translating the messy instinct of your brain into the clean logic of a not MD file.
Speaker A:You are giving the AI the architectural blueprint of how you think.
Speaker B:You are formalizing your methodologies into a language the machine can natively parse.
Speaker A:And the core mantra she leaves us with is just empowering.
Speaker A:You are not the model AI builds of you.
Speaker A:You are the one who corrects it.
Speaker B:This raises an important question, though.
Speaker B:Taking proactive control of one's digital reflection before it ossifies into a permanent, untraceable identity is urgent.
Speaker A:Extremely urgent.
Speaker B:If you allow an AI to continually build a sourceless, flawed identity of you, untangling that massive knot within the machine's architecture later becomes almost impossible.
Speaker A:You have to make the invisible visible, take the instinct and turn it into instruction.
Speaker B:Exactly.
Speaker A:Which leads us to a really fascinating and slightly mind bending thought experiment to leave you with today.
Speaker A:Everything we've discussed so far has been about the relationship between you and a single AI assistant, right?
Speaker A:But if AI platforms are constantly building these sourceless identity level models of us based on incomplete Tuesday afternoon behaviors, what happens a few years from now when two different AI assistants each have completely different beliefs about you?
Speaker B:Like your work AI versus your personal AI?
Speaker A:Exactly.
Speaker A:Say your work AI firmly believes you are a highly impatient, ruthless executive who only values efficiency.
Speaker A:But your personal AI forms a completely different belief that you are an exhausted caretaker who desperately needs comfort, both locked.
Speaker B:Into different servers as absolute truths, right?
Speaker A:So what happens when those two AIs have to negotiate with each other to complete a task on your behalf?
Speaker A:Say they are booking a business trip.
Speaker B:The work AI wants a 500am red eye to maximize billable hours, and the.
Speaker A:Personal AI refuses, demanding a noon flight with an upgrade because it believes you will burn out.
Speaker B:It's a literal digital identity crisis.
Speaker A:Which version of you will win out?
Speaker A:Who gets to decide what you value when you aren't even in the room?
Speaker B:It really highlights why writing your own operating manual is a practical necessity.
Speaker A:It is definitely something to mull over.
Speaker A:Thank you for joining us on this deep dive.
Speaker A:Go check those memory panels before the weekend.
Speaker A:Take a hard look at the mirror and remember, you are the one who corrects the model.
Speaker B:Absolutely.
Speaker A:We'll catch you next time.
