69% of close relatives described the same person as having both a grandiose public self and a vulnerable private self. that number comes from the narcissism literature, but the pattern isn't pathological by default. it's just what charisma looks like from the inside.
you already know this. you walk into a room of strangers and something turns on. you're funny, magnetic, precise. people lean in. then you go home and the person who loves you most asks you a simple question and you feel it ... the performance gap, the metabolic cost of having been lit up for hours. charisma isn't raw expression. research defines it as "controlled and regulated expressiveness," and that regulation is effortful. it runs your nervous system like fifty background apps simultaneously.
so when we talk about building an AI agent that carries your personality into the world, we're not talking about one personality. we're talking about the split. and the question isn't how to resolve it. it's how to carry it without paying the tax.
the split is the architecture, not the bug
here's what the evidence actually says. a mixed-methods study with 60 participants found that across every task context, a shared baseline stayed stable ... Engagement, Serviceability, Decency ... while peripheral traits shifted to fit the situation. the core didn't move. the expression layer did.
humans do this constantly. code-switching is one of the most documented phenomena in sociolinguistics. speakers shift register, tone, even identity markers based on audience. you don't become a different person for your accountant and your childhood friend. you carry the same load-bearing self and modulate the surface.
the technical term for what happens when you can't is Personality Inertia. researchers behind PD-LLM identified it as the bottleneck in current agent design ... RLHF alignment traps models in a single sanitized helpful-assistant persona, and they physically resist expressing different traits under pressure. IRIS tries to solve the same problem with neuron-level steering. both are reaching for the same thing: one identity, context-tuned expression.
the default assumption in agent design is that a personality must be consistent. one coherent self expressed everywhere. the research inverts this. you do not need two personalities for two audiences. you need one load-bearing core with context-tuned expression layers.
every framework keys to the wrong variable
this is where it gets uncomfortable. every current agent framework adapts behavior based on something. just not the right thing.
- PersonaAgent keys to user identity ... a per-user system prompt. it knows who it's talking to but not how well they know you.
- ASTRA keys to opponent stance ... greedy, neutral, generous. built for negotiation, not for relationships.
- RL negotiation agents key to opponent bidding behavior in real time. tactical, not relational.
- DMNA keys to the user's Big Five traits. psychological profiling, not relational depth.
- Proxona builds per-audience personas, but for audience simulation, not agent self-expression.
the variable that matters is not "who am i talking to." it's "how much does this person already know about the person i represent."
strangers need the fascinating version. they need the controlled expressiveness, the charm, the regulation that makes you magnetic. intimates need the real one. they need low performance and high presence. they need to not feel the gap between agent-you and actual-you.
no framework found in the research has a slot for this distinction. relationship depth as a behavioral key does not exist in any agent architecture we could find.
the trust asymmetry will eat you alive
when an agent carries your personality into two audiences, the failure mode isn't inconsistency. it's miscalibrated trust in opposite directions.
strangers over-trust. research shows people anthropomorphize AI agents, ascribe intent to them, and follow them even when the agent is defective. Gen Z respondents in one study treated AI personalities as real, with trust and emotional connection similar to human relationships. a stranger meeting your agent doesn't see a tool. they see you.
intimates under-trust. they perceive the gap between agent-you and real-you immediately. and they care deeply, sometimes violently, about how the agent's behavior reflects on your reputation. one participant in a human-agent alignment study said the agent could be "messing with my money" or "messing with my reputation." that's the intimate voice. they're not charmed. they're auditing.
this is a calibration problem, not a consistency problem. and there's a deployable tool for measuring it: the alignment floor metric. it measures the range of behavioral shift a model exhibits across persona conditions. on lightly-aligned models, persona customization shifted sycophancy rates by up to 45 percentage points. on strongly-aligned models ... the kind LUNARI runs on, Sonnet and Opus ... the shift was near zero.
let me say that plainly. the models we build on may resist the mode switch. the alignment that makes them safe also makes them rigid. the shift may need to happen through structured prompting and retrieval architecture, not through persona instructions alone, because the model's alignment will fight you.
the exhaustion tax is the product
here's where a psychological observation becomes a buildable thing.
the human pays a real metabolic cost for maintaining the split. charisma requires controlled regulation, and that regulation drains you. post-family-visit exhaustion is a measurable nervous system response. hypervigilance burns through metabolic and neurological resources. creators describe being "always on" with no structured downtime. visibility management is continuous labor that persists even on days you didn't post. the personal brand model caps business growth at personal energy limits.
you don't collapse after a family dinner because your family is difficult. you collapse because your nervous system ran fifty apps for four hours straight and never got to close one.
the agent doesn't have a nervous system. it doesn't pay the hypervigilance tax. it can switch modes with LoRA adapters instead of cortisol.
the value proposition was never "we will make you one authentic self." that's a lie therapists charge $200 an hour to not deliver. the value proposition is: we will carry the split for you so you stop collapsing after the performance.
what to actually build
not a digital twin trained on your inbox. that's the trap every saved paper in the library flags. your inbox contains both audiences mixed together. the model learns your charming stranger moves and your exhausting intimate moves from the same corpus, then deploys the wrong one at the wrong time. you don't want a mirror. mirrors don't know who's looking.
what you want is a relationship-tiered agent. one baseline core that never shifts. three expression layers keyed to how well the counterparty knows you:
- tier 1, strangers: high expressiveness, high regulation, zero disclosure. the fascinating version. charisma mode.
- tier 2, acquaintances: moderate regulation, contextual disclosure. warm professional. the version that gets invited back.
- tier 3, intimates: low performance, high presence, transparent about the agent. the real one. no mask, no charm offensive, just the core.
the agent needs a relationship classifier at the front door. every incoming message passes through one question: how well does this person know the person i represent? the answer routes the message to the right expression layer. same core. different surface. the split stays intact, carried by silicon instead of a nervous system that was never designed to hold it this long.
the takeaway
stop trying to become one person. you were never one person. nobody is. the research says the stable core is real and the surface shift is normal. your exhaustion isn't a character flaw. it's a metabolic invoice for a service your body was never meant to provide full-time.
build the agent to carry the split. tier the relationship. key the expression to depth, not identity. and test it hard on strongly-aligned models, because the alignment floor will try to flatten your agent into one sanitized voice. that's the real enemy. not the split.
the split is you. carry it. don't resolve it.