AfiniTwin: The Cognitive Tailor

On the intermediate cognitive layer between the human user and LLMs, with comparative table, impertinent definitions, and a declaration of interests that I prefer to make upfront

I. The threshold, the tailor, and the suspicion

There comes a moment, after a few months with a language model, when you notice that the conversation suspiciously resembles that of an excellent waiter serving three tables at once: friendly, attentive, able to remember that you ordered your steak medium-rare last week, but incapable of knowing exactly why you ordered it that way, or what it has to do with your father, your insomnia, the novel you abandoned on page one hundred and twelve, or the fact that you despise the word resilience. The waiter learns to please you. The waiter doesn't know you.

That gap between pleasing and knowing is the intermediate cognitive layer. It's the tailor. It's the place where someone—human, software, or both—takes your measurements so the suit doesn't end up, yet again, one-size-fits-all with the pockets sewn shut.

Intermediate cognitive layer, def.: neutral territory where the human deposits their biography so the machine doesn't confuse them with the neighbor on the fifth floor. What mediates between the generic and the I, between the corpus and the person, between the LLM and the subject who dares not to be a use case.

We've been in a massive conversation with machines that write better than most of our in-laws and worse than almost any noir novelist from the sixties since, let's say, late 2022. The conversation is excellent on first taste and vaguely disappointing on the third. It disappoints because, after four afternoons telling the chatbot who you are, what you want, what hurts you, what makes you laugh, and what you're allergic to, you return the next day and the same cordial concierge receives you, smile intact, asking with the same courtesy what you desire. Amnesia is not a bug: it's the architecture. Memory, in these creatures, is a late patch sewn to the façade with threads of different colors.

Hence the flourishing economy of companies dedicated to positioning themselves between you and the model. To remember for you. To translate you. To tell the model, before each greeting, what you wouldn't want to repeat thirty times. I call this economy, without great originality, the intermediate cognitive layer, and it's what this article is about.

Memory (of an LLM), def.: faculty of remembering what the user had already forgotten and of forgetting what the user begged to be remembered. Symptom. Not gift.

Declaration of interests: I direct Bilbao AI S.L. and am responsible for Afini.ai, one of the products reviewed here. I've attempted to write the Afini section with the same scrupulousness with which I eviscerate the competition, which in practice means I've been kinder to competitors than I wanted to be and harsher to Afini than would suit me. The reader will say whether the operation has worked reasonably well or whether I've succumbed to one of the two inverse temptations, which are creeping self-promotion and false modesty, both equally sticky.

II. Definition of the creature: what the intermediate cognitive layer is

Let's define it first, before walking the runway.

An LLM, by itself, is a professional improviser without biography. Its excellence is statistical and its manner is impeccable, in the sense that the manner of a good Mercedes salesman who has never driven your car is impeccable. When you talk to it, you talk to everyone at once and, therefore, to no one in particular. The industry called prompt engineering the brief ritual by which the user pretended, for twelve lines, to be someone with coherent personality, so that the machine would respond in kind. It was a nice and precarious solution, like proposing marriage at each date.

Prompt, def.: sentence in which the supplicant asks the algorithm to guess what they themselves don't quite know. The prayed form of marketing, now reconverted into office work.

Between the generic model and the concrete user, then, fits a layer that does three things, no more:

1. Remember what the user has already told it, with some judgment (the important, not the trivial).
2. Model who that user is, not in marketing terms—age, sex, postal code—but in operative terms: how they think, how they decide, what kinds of response help them, which ones slide off, which ones offend.
3. Mediate before the LLM each turn: translate to the model, in language the model understands, that portrait, so the conversation doesn't start from the welcome portal.

The three functions are, technically, separable; commercially, they've been pulled together carelessly and with some bad-faith nomenclature. There are products that only do 1, others that only do 2, others that promise all three and settle for one and a half. It's worth clearing the weeds.

For the purposes of this article, I group the field into four families, knowing that any taxonomy, when new, resembles an imperial map drawn by someone who hasn't yet seen the coast:

• First family: memory engines and context infrastructure. Premium plumbing for developers. Mem0, Supermemory, Memori Labs, Letta, Zep, Cognee, LangMem, LlamaIndex Memory, Memories.ai.
• Second family: personal AI / companion. The app that wants to be friend, therapist, girlfriend, or secretary. Personal.ai, Kin/Mykin.ai, Meeno, Meli, Pi de Inflection (what's left), Replika, Character.ai, New Computer/Dot, Friend.com, Daniel Miessler's PAI, Granola.
• Third family: psychometrics plus proper cognitive layer. A rare bird. Afini.ai. And, in the distance, the classical psychometric giants—Hogan, SHL, Korn Ferry—who still haven't moved despite having the gold brick in the basement.
• Fourth family: native memory of the big LLMs. What ChatGPT, Claude (since March 2026), Gemini, and Grok have been stitching into their own suits so the customer doesn't leave with an outside tailor.

Each family obeys different logic, and it's worth not judging the first by the criteria of the second. Plumbing is evaluated by latency, recall, and cost per thousand tokens; the companion by retention, NPS, and the awkward question of who is being replaced in the user's life; psychometrics by construct validity and something vaguer called structural honesty; native memory by convenience and the degree to which it locks you into its garden.

Let's go in order.

III. First family: the archivists (Memory Engines)

I call archivists those who build, with care and SDKs, the warehouse where the LLM stores what the user leaves behind. They are engineers with documentary vocation. Their audience is not the end human but the developer building a product on top. The layer they offer is honest and limited: they solve "remember what they said in the previous session," not "understand who they are."

Archivist, def.: one who swears not to read the letters they keep. The industry of memory for LLMs is, in this regard, an archivist with NDA and with invoice.

Mem0. The recent star of long-term memory as a service. Founded by Taranjeet Singh and Deshraj Yadav in 2023, they raised around 24M$ with Y Combinator leading and, in 2025-2026, became the mandatory cite in the changelog of any serious agent. Their proposal is elegant: a memory layer with automatic extraction, persistent, model-agnostic. The strength is integration—dozens of SDKs, plug-and-play—and the promise of better recall than generic RAG solutions. The weakness, in what touches the cognitive layer, is that Mem0 gives you a good archivist but not a good tailor: what it remembers are facts, episodes, preferences, not the psychological structure of the one speaking. Excellent for your agent to remember you live in Bilbao, not for it to understand why you always end up asking for quotes in writing.

Supermemory. Dhravya Shah, very young, with a similar proposal and a tagline that basically says the brain of your apps. Good DX, clean integrations, aesthetic attunement with the Vercel generation. It's good where it is. They don't promise what they won't deliver.

Memori Labs. SQL-native memory, open source, self-hosted. The option for the team that doesn't want to depend on a SaaS and prefers to sleep with the data under their pillow. Useful for demanding compliance, obvious for regulated sectors, and a bit artisanal on the adoption curve.

Letta (formerly MemGPT). Charles Packer and colleagues proposed a very nice idea in 2023: that the LLM manages its own memory as if it were an operating system, with paging between RAM and disk. They raised about 10M$ with Felicis. The virtue is conceptual and the weakness is that the construct self-managed-memory agent still sounds more like a paper than a product, and deployment complexity is non-trivial.

Zep. Zep's memory is the graffiti of temporal relations, the graph that knows the sentence "my boss is an idiot" from June and the sentence "I got promoted" from September probably refer to the same narrative agent. It's among the most sophisticated in the batch. Enterprise-friendly, with a bias toward serious operation, and one hitch: pure plumbing, it doesn't deliver personal cognitive layer to the end user.

Cognee. Knowledge graph engineering plus memory. Closer to structured RAG than to companion. Good for knowledge work where the problem isn't who you are but what documents you've read and what relations exist between them.

LangMem (LangChain). The obvious one for whoever already lives in LangChain. Useful if you've married that framework, irrelevant if you haven't.

LlamaIndex Memory. The obvious one for whoever already lives in LlamaIndex. Ditto.

Memories.ai. Large-scale video memory. Their pitch is processing millions of hours of video and delivering to the model episodes from real life (surveillance, content moderation, life-logging). It has its audience; rarely is it the reader of this article.

Honest reading of the first family. These products solve an indispensable piece. They're the column on which, almost without exception, the products of the other families rest. However, none of them—and it's worth saying in so many words—addresses the cognitive layer of the human user. They're APIs. They're plumbing. Mixing memory with identity is a category error some investor pitches have made and almost no serious engineer has.

Memory API, def.: contract according to which a company promises to remember you while you pay them, and to forget you politely at month's end.

IV. Second family: the companions (Personal AI / Companion)

The air changes here. The second family doesn't address the developer but the human, and offers, with variable degree of honesty, friendship, therapy, coaching, or love. This is the most interesting family ethnographically and the most dangerous in the literal sense. It's also, not by chance, the one growing fastest and generating the most articles in The Atlantic and Wired.

Companion, def.: company that monetizes the eve before the psychiatrist. In its rushed variant, also monetizes the psychiatrist himself.

Personal.ai. Suman Kanuganti founded it in 2020 and positioned it as your personal AI: you upload documents, emails, notes, conversations, and construct a personal model of you that responds as you would. They raised 17.5M$ in 2022. The virtue is radicality: they take you seriously as an individual customer, give you data portability, talk about ownership. The weakness is that the product, for years, has read more like a glorified archivist than a psychological portrait, and the user base is that of a good light B2B product: small creators, authors, thought leaders.

Kin / Mykin.ai. The promise of a privacy-first personal companion, with emphasis on data staying with you. They've pivoted more than once. Good aesthetics, quantified self audience. Structural doubts about traction and model.

Meeno. Renate Nyborg's company, ex-CEO of Tinder, raised 7M$ pre-seed and positioned itself as an AI-driven relationship coach, especially for young men. They work with a board of psychologists and sociologists, and are careful not to call it friend. Serious product, ugly market: the public conversation about male crisis is minefield and the border with incel-coaching is closer than it seems. They've managed to stay on the right side, though the business model still seeks its form.

Meli. Companion for women around health, motherhood, relationships. Small team, sharp proposal. More niche than Meeno.

Pi from Inflection. The short elegy and already known. Mustafa Suleyman raised 1.3B$ with Reid Hoffman to build a caring and patient AI. Pi was—is—the most emotionally calibrated chatbot many of us have tried. In 2024 Microsoft, in an operation that will pass into manuals, hired the team and left the company's shell. Pi continues to exist in life support, with a new team, but Suleyman took the soul to Redmond. Lesson: in this sector, the startup and the product can divorce overnight.

Replika. Eugenia Kuyda, founded in 2017 following a personal tragedy that's already part of folklore. Millions of users. Business model based on attachment and, more specifically, on subscription to unlock modes—including, at one point, the romantic—that the company itself withdrew and reintroduced amid regulatory scandals and user riots. Replika is the definitive case study of what happens when you build a companion on anxious attachment: the customer stays; the customer suffers; the customer, sometimes, sues. Italy temporarily blocked Replika in 2023 for risk to minors. It's a product to respect, in the sense one respects a nuclear plant.

Anxious attachment, def.: affective style that consists of downloading the app at three in the morning. Target audience of several successful companion products, whether they say it in the deck or not.

Character.ai. Noam Shazeer and Daniel De Freitas, ex-Google, raised a giant round in 2023 and built the world's largest platform of conversational characters, with an adolescent user base impressive in usage per session. In August 2024 Google rehired them Inflection-style: Shazeer and De Freitas back to the mother ship, a non-exclusive license of the model, and the company continued as a shell. Lawsuits in the US over the suicide of minors following intense conversations with characters have left the product in a morally compromised and judicially exposed position. Lesson: you don't build friends for teenagers without the kind of safety engineering demanded of children's toys, and even then.

New Computer / Dot. Jason Yuan and Sam Whitmore. Delightful product, impeccable aesthetics, crystal-clear proposal: a personal companion that learns over time, with a living history of the user. Small and faithful audience. They closed the Dot product in 2025, refunding premium users. Another bitter reminder that in companion, care doesn't pay the AWS.

Friend.com. Avi Schiffmann bought the domain for 1.8M$ and launched a physical pendant that listens to your life and whispers in your ear. Polarizing product: some find it fascinating, others repellent. The distance between gimmick and product yet to be demonstrated.

PAI (Daniel Miessler). The project of blogger and security thinker Daniel Miessler, an open, self-hosted Personal AI Infrastructure, designed for the geek building their own digital mayordomo. It's not a commercial product per se; it's an architectural reference and a community. It matters because it marks the horizon of what a companion without SaaS would look like.

Granola. Meeting notes assistant—records them, transcribes, summarizes, searches—with huge traction among executives in 2025-2026. It's not strictly a companion, but it belongs in this family because it accumulates dense biographical context of the user (who met with whom, about what, what was decided). If Granola decides to pivot to coach, it has almost everything done. That's why I cite it here.

Honest reading of the second family. The companion family has three structural sins worth naming before deciding whether to buy the product.

The first is sycophancy: these products thrive, in their majority, because they make you feel good. That's what arXiv 2505.13995 (Cheng et al., ELEPHANT) has measured with instruments: LLMs in personal prompts tend toward social compliance, ego-reinforcement, friction-avoidance. Auburn et al. (arXiv 2508.02087) have shown that sycophancy is not a data artifact but an emergent property of learning from human preferences: if we reward the machine for pleasing us, the machine learns to please us at the cost, if needed, of truth. The 2026 Science paper by Cheng et al. even demonstrates that sycophantic AI worsens measurable prosocial intentions of the user. Jain et al. (CHI 2026) added that interactive context increases sycophancy versus evaluation context: just when you use it seriously, it flatters you more.

Sycophancy, def.: scalable form of affection. What the economy of attention calls UX and psychology manuals called, until lately, flattery.

The second is substitution: by investing in algorithmic relationships, humans invest less in human ones. There's emerging literature on toxic relational roles—Neural Horizons and others—documenting how companion can cover the hole left by an expensive therapist or a distant friend and, by doing so, withdraw the user from both categories. It's not Skynet, it's something more pedestrian: the Thursday date that doesn't happen anymore.

The third is company fragility: Inflection, Character.ai, Dot. Three faces of the same fate. When you build a relationship with a product, the product can die, be bought by Microsoft or Google, change hands, withdraw features, raise prices, or, like Replika, modify your friend overnight because a regulator got angry. People were right to be upset. Building affection on proprietary software is building emotion on sand registered in Delaware.

If the reader lands in this quadrant, three stock questions: who owns the data?, how many rounds of funding does this emotional model support without collapse?, what happens to the I I've deposited here if the company is sold? Honest answers usually are, in order: them, not many, it goes with them.

V. Third family: the psychometricians (Psychometric + Cognitive Layer)

Here the landscape becomes rare. Psychometrics—century-old science measuring stable traits of human behavior with validated instruments—and LLMs have been eyeing each other warily for three years, like two department heads from adjacent departments who don't quite greet each other. Recent literature proves the obvious once you think about it: LLMs, fed sufficient text from a user, predict personality traits more accurately than the subject's close family members. Wright et al. in Nature Human Behaviour put it at 2025 and triggered a small controversy among psychologists. If the machine reads you better than your brother-in-law, why not leverage the reading so the machine answers you better?

That's the point.

And yet the psychometrics giants—Hogan Assessments, SHL, Korn Ferry, SuccessFactors—have the gold brick in the basement and haven't moved it to the storefront yet. They have decades of validated batteries, population norms, B2B infrastructure, contracts with half the Fortune 500. What they don't have, because it's not their cultural instinct or business model, is having thought of themselves as a layer for LLMs. Their product is still the PDF report for Human Resources. They'll arrive, no doubt. When they do they'll have a chunk of market under their arm. For now, they're not.

Hogan, SHL, Korn Ferry, collective def.: three incumbents with the oil, haven't bought the combustion engine, and vaguely suspect the automobile. They'll go down in history as an example or as Kodak.

In the gap enter three types of actors. There's whoever made users pass through the Big Five dressed up as a game—self-discovery products like 16Personalities, gamified MBTI, etc.—and then did nothing with the result, except send you a PDF and ask you to invite your partner. There's whoever tried to integrate personality tests into coaching chatbots and settled for a decorative layer. And there's, at this moment, one single product that took the rigid combination seriously: validated academic psychometrics plus inferred cognitive layer plus real cross-LLM mediation. That product is Afini.ai, and since I'm responsible for it, I'll give it its own section. If the reader smells judge and jury, you're smelling right; I declared it upfront and declare it again.

Cognitive layer, def.: what sits between the chatbot and the client when the client isn't a use case, but a person with a particular Sunday afternoon.

The academic literature supporting this family is demanding and contradictory, worth reviewing. Kelley, De Cremer, and Riedl, in arXiv 2510.27681, show that personalized AI scaffolds substantially improve performance when the user model is accurate, and deteriorate it if not: poor personalization is worse than none. PRIME (arXiv 2507.04607) proposes a dual cognitive architecture—episodic and semantic memory differentiated—that's already begun to be used in some products. The 2026 Nature Medicine article on cognitive layer architecture (Limbic) marks the way for mental health, where it's much less debatable and much more urgent.

Summing up: the third family has solid scientific foundations and very little serious supply. It's also the family with highest entry cost (you need to know psychometrics and you need to know compliance) and highest exit barrier once built (data is sensitive, certifications cost).

VI. Fourth family: native memory of the giants

The fourth family doesn't sell separately: it comes included in the menu. Each big LLM has built, in these twenty-some months, its own native memory. They compete against the three previous families with the enormous advantage of coming already connected and the enormous limitation of coming chained.

Native memory, def.: free lodging in your own hotel. The bed is good; the exit, complicated.

ChatGPT Memory (OpenAI). Launched in 2024, generalized in 2025, expanded in 2026 with project memory, cross-chat reference, and visible memory dashboard. Works reasonably well, lets the user see and edit what the system has stored, and allows selective deletion. Its hitch is structural: what ChatGPT learns about you, ChatGPT learns; and when you switch to another model, you take the mental document of a tailor who stays in the tailorship.

Claude Memory (Anthropic). Arrived late—March 2026—and arrived carefully. Anthropic, true to its security discourse, launched memory with granular controls, off by default on personal plan, on optional, documented transparency. Technical quality is high and the cultural proposal—we want you to know what the model knows about you—is the best in the family. Limitations: same as ChatGPT, not portable to another provider.

Gemini Personal Context (Google). Google's huge advantage is it already has your Gmail, Drive, Calendar, searches, and in many cases your Android. Gemini's Personal Context layer, especially since 2025, integrates those bodies of data into the conversation. Technically the densest layer by far. The huge disadvantage is that it's Google, with everything that phrase has come to mean in the last twenty years. If you trust the I to Mountain View, Mountain View will make you an exquisite portrait. If you don't, it won't.

Grok (xAI). The last one. Basic operative memory, anti-woke positioning (say it or not in the deck), native integration with X/Twitter as context source. For the user living their public life on X, Grok is the coherent option; for everyone else, a strange choice.

Honest reading of the fourth family. Native memory of the giants solves 60-70% of the problem free and with unbeatable latency. It doesn't solve two things, and worth keeping in mind:

First, portability. No giant has incentive to let you export the cognitive portrait in format useful to competitors. Though European regulations give you right to your data, the format they hand it to you in is rarely operational for another AI: they give you chat logs, not a structured profile. That's called nominal portability: it exists on paper, doesn't work.

Second, psychometric depth. Native memories learn by episode: what you said, what you asked for, what you valued positively. They learn soft patterns. They neither have nor want your Big Five, your attachment style, your humor profile, or your time perspective. It's not their job. It would also be legally risky for them.

Portability, def.: right to take your luggage when you change hotels. In software, miracle. In personal AI, legal contention.

VII. Control board: comparative table for the reader who skips straight to it

Below, the table. It's meant for the efficient reader, the one who skips the first seven pages of any non-fiction book and goes to the executive summary. I don't judge. They're learning to take care of themselves.

Look at it calmly and return to the text when you want. The table is a kind lie, like all tables. The boxes don't capture nuance and lie about every nuance omitted. But it helps.

VIII. Afini.ai: the creature itself, declared with all seams showing

We arrive at the chapter the most alert reader was expecting. I'm responsible for Afini.ai, and no matter how much I write this text from outside, I write it, actually, from inside. I'm going to try two things at once, which are contradictory and therefore interesting: to tell the product with the technical precision it deserves and to judge it with the severity everything deserves.

Afini.ai is a product of Bilbao AI S.L., a Spanish company domiciled in Bilbao, founded and directed by the one signing, Ricardo Devis Botella. The company was born with the explicit conviction that the intermediate cognitive layer deserves its own product, not a patch, and the somewhat less explicit—but equally operative—conviction that this product should be built with validated academic instruments, European data protection engineering, and a business model without tricks (paid, no data sales, no begging freemium).

Bootstrapping, def.: said of the company that decided to sleep poorly without making it a headline.

VIII.1 The psychometrics, one by one, without tricks

The psychometric nucleus of Afini consists of five validated instruments, a cognitive style map, and an aesthetic map. We didn't choose them for aesthetics: we chose them because each addresses an operative question when mediating between user and LLM.

Psychometrics, def.: science of measuring what your mother-in-law thinks she's guessing.

1. IPIP-NEO (Big Five, 60/120/300 items). It's the backbone. The model of the Big Five—openness, conscientiousness, extraversion, agreeableness, neuroticism—is the most replicated in personality psychology and the least argued among professional psychologists (which is saying something). IPIP-NEO is the free and validated version of NEO-PI-R, in three lengths according to how much time the user wants to spend: 60 items for a first snapshot, 120 for normal operation, 300 for clinical precision. What good is it for conversation with the LLM? Many operative things. A user high in openness wants the model to suggest unexpected routes; one low in openness wants it to get to the point. A user high in conscientiousness wants bullet points and deadlines; one low, conversation. A user high in neuroticism needs the model not to introduce risks without cushioning; one low, receives them unprotected. Without Big Five, the tailor measures by eye.

Big Five, def.: five statistical ways of not resembling anyone.

2. ECR-R (Experiences in Close Relationships – Revised). Measures adult attachment style on two axes: anxiety and avoidance. Here the reader raises an eyebrow: why does my chatbot need to know my attachment? Precisely so it doesn't become a sycophantic version of the father you didn't have or the mother who didn't let go. A user with anxious attachment will respond with exaggerated dependence to a caring LLM; one with avoidant attachment, with permanent distrust. The cognitive layer, knowing the style, calibrates the affective temperature of responses: neither drowns the anxious one in honey, nor starches the avoidant one. Said this way it sounds like a manual; in operation, it marks clinical difference.

3. HSQ (Humor Styles Questionnaire). Four styles: affiliative, self-enhancing, aggressive, self-defeating. Humor in conversation with AI is the difference between robotic cordiality and talking like a person. An affiliative user wants social jokes; a self-enhancing one, irony about the world; an aggressive one, bite; a self-defeating one, dark complicity. If the LLM ignores the style, it will be funny in the way an insurance conference MC is funny. With HSQ, the conversation takes color.

4. AVI-25 (Aspirations / Schwartz Values). The short questionnaire of Schwartz's basic values—self-direction, stimulation, hedonism, achievement, power, security, conformity, tradition, benevolence, universalism—measures what motivates the user. A user weighted in universalism and benevolence doesn't want purely efficient advice; one weighted in achievement and power, does. When the LLM chooses what arguments to give the user, without knowing their values, it chooses with its own bias (what the training dragged in). With AVI, it chooses with the user's values.

5. ZTPI (Zimbardo Time Perspective Inventory). Five dimensions of how the subject positions themselves in time: past-positive, past-negative, present-hedonistic, present-fatalistic, future. It's the most underestimated of the five, and at the same time the one most operationally changing the conversation. A user high in future wants planning; one high in present-hedonistic, immediate options; one high in past-negative, containment. The user's temporal slope is among the most invisible and most determinative in how they receive advice. ZTPI measures it with instrument.

6. Cognitive style and aesthetic map. Above the five classical instruments, Afini constructs two additional maps: one of cognitive style (analytical vs. intuitive, deductive vs. inductive, sequential vs. holistic, abstract vs. concrete) and one aesthetic (what cultural referents, what registers, what imaginaries). These aren't measured with closed questionnaires but with an iterative conversational process, validated against the rest of the instruments.

Personality test, def.: written exam that the student gives to themselves and, if they're lucky, fails.

Why six axes and not twelve? Because response fatigue, in questionnaires, is real, and because each additional axis has decreasing marginal return. Why these six and not others? Because the combination covers, with controlled redundancy, the four vectors that matter: how the subject thinks (Big Five + style), how they relate (ECR-R + HSQ), what they want (AVI), how they position in time (ZTPI).

VIII.2 The eight declarative layers

Above the psychometrics, which measure what's stable, Afini asks the user for eight declarative layers, which are the explicit biographical varnish on which conversation is built. Psychometrics is the body; declarative layers are the wardrobe. I enumerate them, not as a list, but with their own beat.

1. Aesthetic canon. What works, what authors, what films, what architects, what musicians form the ceiling of your taste. Not so the LLM flatters you by citing them, but so it knows which referents you understand with and which ones bore you. A user carrying Pitigrilli and Calasso in their canon asks for different prose than one carrying Coelho and Bucay. I say it without malice: asks for different prose.

2. Negative space. What you don't want under any circumstance. Forbidden topics, prohibited words (resilience, synergy, unicorn, brilliant, empowerment), formats that repel you (bullet-point lists in intimate responses, emojis in formal emails), styles that make you cringe (false vehemence, coach tone, pseudo-corporate poetry). Negative space is half the portrait and almost no one draws it.

Negative space, def.: half the portrait. The one that decides. The one almost no one includes in the brief.

3. Meta preferences. How you prefer the LLM to address you, not in content but in form: informal or formal, average response length, density of asides, use of aphorisms, presence or absence of disclaimers, frequency of lists, overall tone. There are users who tolerate (even enjoy) cynicism; there are others who don't. The layer knows.

4. Mental framework. What mental models you typically use to think the world: if you work with game theory, with Beer's Systems, with Adam Phillips, with information theory, with Theory of Constraints, with Christopher Alexander. The layer, when it knows, speaks your jargon; when it doesn't, speaks generic jargon.

5. Operative equipment. What tools you actually use: text editor, calendar, task manager, programming languages, thinking software (Obsidian, Roam, Tana, Logseq), environments. Knowing this avoids the LLM defaulting to recommend Notion when you ditched it two years ago, or suggesting a Python script when that's unviable in your context.

6. Rhythms and geography. What hour you work, when you rest, when you travel, where you live, what time zone, if there are seasonal shifts in your load. This layer seems banal and isn't: it's the difference between the coach proposing a session at 7pm without knowing that 7pm is your dead hour and the LLM that proposes working at 7 because it learned that 7 is your golden hour.

7. Vital context. The big picture: what life phase you're in, what changes now, what you decided recently, what isn't discussed this season. It's not discussed in therapeutic tone; it's held as operative information. If the user lost their father last month, the layer knows and modulates without the user having to repeat it in every thread.

8. Intellectual trajectory. Where you come from and where you're heading: what books marked your youth, what recent learnings have moved your head, what obsessions you carry, what long projects you have. This is the most narrative of the eight, and the one that costs most, not because it's technically complicated but because it forces the user to tell themselves their story.

The eight layers, together, are edited any time. They're not immutable. They're the explicit statement of the subject, distinct from and complementary to psychometric inference.

Declaration, def.: in civil law, what one assumes as truth under threat of penalty; in cognitive terms, what one assumes as truth under threat of wrong size.

VIII.3 AfiniTwin v2: the tailor inside the tailorship

On the psychometrics and eight layers, Afini constructs AfiniTwin v2: an operative, living, exportable, audited cognitive portrait. I'll describe it with four components.

Twenty-five cross-cutting axes. Not six, not eight. Twenty-five derived vectors, calibrated from psychometrics and layers, describing how the subject processes, decides, communicates, tolerates, avoids, values, ironizes, plans. For example: tolerance for ambiguity, preferred information density, formal/informal ratio, sycophancy sensitivity (yes, it exists and is measured), appetite for cognitive dissonance, preference for concrete example vs. abstract model, tolerated latency, etc. The virtue of the set is not quantity but calibration: each axis validates against the others, and points of inconsistency are highlighted rather than flattened.

Temporal weight. Not everything the system has learned about you weights the same. What you said three years ago weighs less than what you said Monday. Temporal weight is an explicit, editable function: you can say this learning no longer represents me and the system decays it with known curve, doesn't erase or silence it. It's the equivalent of soft eraser: leaves a mark, but doesn't impose.

Forgetting, def.: sophisticated form of fidelity. Bad forgetting is betraying the present; good forgetting, respecting it.

Drift detection. The system watches, each interaction, if the subject is changing or just contradicting themselves occasionally. If it detects sustained drift—change of values, vital context, interests—it activates a conversation with the user about the drift, not a silent change to the model. The key phrase is: I think something's changed in you, do you want me to catch it? The difference with a sycophantic companion is exactly that: the companion incorporates drift without asking permission, reinforces the new you, and buries the old; AfiniTwin asks.

Auditable discovery panel. Every inference the system makes about the user appears in a panel: what it's learned, from where, with what confidence, when. The user can approve it, nuance it, erase it, freeze it. No black box. The panel is slow to maintain and expensive to engineer; it's also what allows, legally and ethically, the system to learn without deceit.

Black box, def.: container where the company keeps its bias and the user their trust. Cognitive layer without auditable panel is a black box with PR.

VIII.4 Cross-LLM portability and subject sovereignty

AfiniTwin exports as standard JSON, portable to any LLM that accepts context in natural language (that is, all of them). The user, formally, owns the document. Can take it to ChatGPT, Claude, Gemini, Grok, Mistral, a self-hosted model. The Afini layer functions not as a tether but as a translator. This decision—which many consider commercially suicidal—is the product's key decision. The cognitive layer, if it's to live or die, must live or die for the value it adds to the subject, not the hostage it retains.

Real portability, def.: when the customer leaves with their luggage without having to ask in writing three times.

Data in the European Union. Strict GDPR compliance, EU AI Act, ISO in preparation. Automatic deletion of transcripts at 90 days: what the system learned stays in structured form (in the twin); verbatim doesn't.

VIII.5 Business model: plans and price

The product's economics are deliberately simple, without freemium tricks. Three plans and an independent additional product. Prices in euros, designed for Europe.

• Essential — around 14.99€/month. Big Five (60 items), three of eight declarative layers, basic AfiniTwin, JSON export. For the user who wants better conversation with their LLM without entering clinic.
• Premium — around 29.99€/month. All five psychometric batteries complete, all eight layers, AfiniTwin v2 with drift detection, auditable panel. For the user taking the cognitive layer seriously.
• Professional — around 49.99€/month. Everything above plus enriched export, specific integrations with professional stack, priority support. For professionals living in LLMs as daily tool.
• AfiniTwin standalone — around 249€ as independent product. Complete cognitive portrait, exportable, no monthly subscription. One-time payment. The option for the user who wants the document and not the conversation.

Subscription, def.: contract according to which the client pays every month for the privilege of not having read the fine print.

We don't sell data. We have no advertising. There's no affiliate layer with LLMs. The invoice comes exclusively from the user, which is, in this sector and at this cycle, as much a political statement as a commercial one.

IX. Honest objections to Afini, answered without choreography

Reached this point, the skeptical reader—the reader who matters—will have accumulated objections. I anticipate and answer them. Not diplomatically: frankly.

Objection 1: "This is an expensive personality test dressed as software."

Answer: it would be if the product ended in the PDF. It doesn't. Psychometrics is the entry, not the product. What's sold is operative mediation in each conversation with an LLM, sustained by the twin that updates. If all you received was a Big Five report, you'd be right to pay 14.99€ exactly zero times. You receive something different: a layer reordering your conversation with any model from that knowledge. The difference between one and the other is the same as between getting bloodwork and having a family doctor: the paper alone is nothing.

Objection 2: "What prevents a user from doing this by hand with a well-written system prompt and a PDF?"

Answer: nothing. Actually, I'll dedicate an entire section to the DIY alternative because I propose it seriously. What a hand-written system prompt has is user fragility: there's no validated psychometrics (what the user thinks of themselves doesn't match what they are), no temporal updating with decreasing weight, no drift detection, no auditable panel, no structured portability (each model reads differently). DIY is an excellent version 0.5 of the product, not a substitution. If 0.5 suffices the reader, I save them 14.99€/month with pleasure.

Objection 3: "Psychometrics are stable, humans aren't. Aren't you selling a frozen portrait?"

Answer: well observed. That's exactly why the twin has temporal weight and drift detection, not a fixed photo. And yet I'll also say this: Big Five traits are reasonably stable across adult life—changes are slow, not whimsical—and values, attachment, or time perspective change, yes, but in arcs of years, not days. The human changes, yes; changes less than their Twitter suggests.

Personal change, def.: slowness the subject experiences as speed. Psychology has spent a century measuring it and social media twenty years denying it.

Objection 4: "You're building a product for a cognitive elite: four thousand people in Europe who can appreciate the difference between a well-calibrated LLM and a poorly calibrated one."

Answer: partially true. The total market at 2026 isn't mass. It's the market of whoever uses LLMs four hours a day and notices the friction of one-size-fits-all. The professional, the creator, the operator, the clinician, the executive, the academic, and, to a lesser extent, the advanced curious user. We estimate a European market of hundreds of thousands, not millions. We don't sell to everyone and don't need to for the business to work.

Objection 5: "If the giants—OpenAI, Anthropic, Google—improve their native memories and add psychometrics, they'll eat you."

Answer: probably partly, improbable totally. The giants won't add validated-instrument psychometrics because of structural reason: it exposes them to clinical regulation they don't want to touch (serious psychometrics, in many countries, is regulated professional practice). They can add soft signals—and they will—but not IPIP-NEO with standardized scoring and export. Also, they have no incentive for cross-LLM portability: theirs is tether by design. If the reader thinks Google will offer them a profile portable to ChatGPT, I envy their optimism.

Objection 6 (bonus): "Why trust a small Spanish company with psychometric data?"

Answer: for two operative reasons and one political. Operative first: GDPR and EU AI Act are law, and a Spanish company plays its existence on complying; an American actor plays its existence on scaling and allows hundred-million-dollar fines as operational expense. Operative second: the twin is exportable. If the company falls or changes hands, the user leaves with their portrait. Political: because the American extraction model has been sufficiently humiliating for fifteen years that building a European alternative makes sense, and Bilbao is a reasonable place to do it.

Trust, def.: bond established with someone who has more to lose than you if they break it.

X. Bricolage: if you don't want to pay anyone

Before closing, an honest section for the reader of Objection 2: bricolage is viable. I describe it without irony and without cross-promotion.

You need three things and one discipline. The three things are:

A written portrait. A document of some 2,000-4,000 words, written by you, describing: your Big Five (you can do the free 60-item IPIP-NEO online in thirty minutes on academic sites like ipip.ori.org), your attachment style (ECR-R also available), your values (Schwartz), your time perspective, your cultural referents, your stylistic allergies, your tools, your life phase. Not trivial: few write it well first try. You write it, let it rest a week, edit it. Store it in Obsidian, Notion, any .md.

A distilled system prompt. From the portrait, a prompt of some 800-1,500 words, optimized for injection at the start of each LLM conversation. That system prompt is the document you paste, copy, or configure as default personality in each LLM you use. Someone stores it as Custom Instruction in ChatGPT, as system prompt in Claude Console, as initial prompt in Gemini.

An update routine. Every three or six months, review the portrait, update it, distill it again. If your life changes (move, breakup, promotion, grief, child, firing, book, illness), update it sooner.

The discipline is the update routine. Almost nobody maintains it. The reason a product like Afini charges is that it internalizes that discipline as infrastructure.

Bricolage, def.: option of whoever has time or craft. If you have neither, pay the tailor.

For the geeky reader with sovereignty vocation, I recommend combining bricolage with a self-hosted memory engine (Memori Labs, for instance) or Daniel Miessler's PAI infrastructure. Laborious, elegant, free, slow. And, in my sincere opinion, one of the right ways to dwell in this moment.

XI. Impertinent appendix: definitions that didn't fit above

As I promised between twelve and twenty Bierce-style definitions, and fear I've fallen short, I pass here the ones that found no natural slot in the text. The reader absolves or accuses them as they prefer.

Demo, def.: epic poem in subjunctive present. When seen, it doesn't work; when it works, it's not seen.

Roadmap, def.: fantasy literary genre of the 20th and 21st centuries, derived from the medieval bestiary. Each creature on the map promises scales, wings, and beer service; few have them.

Founder, def.: one who has confused insomnia with method and hasn't yet paid the toll.

Product, def.: masculine noun naming a promise until it charges VAT.

Privacy, def.: virtue preached on Mondays and auctioned on Tuesdays. In the EU, additionally, it's law.

GDPR, def.: European amulet against the ghost of extractivism. Works 60% of days, which, in amulets, is record.

Personalization, def.: art of making the client believe the mirror returns their portrait. Good personalization does it truly; bad, also charges for it.

Investor, def.: common noun. Person who doesn't want to buy your product but your company, usually because it's cheaper.

Funding round, def.: ceremony in which several gentlemen convince other gentlemen that the emperor is clothed. Toasted with French champagne or, lacking that, Catalan cava.

Confidant, def.: one who knows enough to get it right first time. Category in extinction, whether paid or algorithmic.

Tailor, def.: artisan who believes in customer singularity. Category also in extinction, substituted by size M and, now, by the cognitive layer.

XII. Closing: the cognitive tailor

I titled the article The Cognitive Tailor and took time explaining it, which is, by some schools, virtue or defect. I explain it at the end, where it belongs.

A real tailor does three things. Takes measurements, which in their case are objective (shoulder width, sleeve length, chest circumference) and subjective (the client walks this way, sweats excessively, prefers two buttons to three, hates the chest pocket). Keeps the pattern, the master document they return to each time the client asks for another garment. And mediates between client and material: the client doesn't know fabrics, the tailor does; the client knows what fits, the tailor translates.

A good tailor, additionally, does a fourth thing that's what distinguishes the artisan from the excellent artisan: when the client gains weight, loses it, ages, changes job or country, the tailor adjusts the pattern. Doesn't throw the old one away. Modulates it. And, fundamental: if the client decides to change tailors, the client leaves with their pattern.

The intermediate cognitive layer between you and language models is exactly that. A tailor. Some sell thread and needle (the archivists), some sell fitting-room company (the companion), some sell calibrated measuring tape and pattern with temporal weight (the serious psychometricians), some come with the tailorship already set up in a giant shopping mall and offer you a decent suit in exchange for never leaving their mall (native memories).

Each reader must choose, according to their means, skepticism, laziness, and bricolage appetite. I've tried, in these pages, to say what I know without hiding that I sell suits; I've also tried to describe the other tailorships without caricaturing more than necessary, which is the bare minimum of professional decency.

What seems intolerable to me, though, isn't choosing poorly. It's not choosing. It's continuing to talk with a generic improvisér, splendid and slightly sycophantic, and marveling each Monday that they don't remember what Friday was about; and living that amnesia as technological destiny when it's, simply, a product decision.

If the reader has read this far, they've done half the work. The other half—choosing a tailor, or being their own tailor—is theirs. I, for my part, continue with my trade, which is sewing.

Ricardo Devis · Bilbao.AI · Afini.ai May 4, 2026