Margaret Chen, the same Margaret from the opening of this series, has noticed something she has not told her daughter. Twice in the past month, she has walked into a room and lost the thread of what she came in for. Once, last Wednesday, the loss persisted for what felt like several minutes, although she suspects it was less. She felt unsteady afterward, the kind of unsteadiness that follows a moment of recognizing oneself unaccountably out of step with one’s own intention. She has not mentioned it to anyone. She is seventy-three. Her own mother had Alzheimer’s. Margaret knows what this might be.
It might be nothing. It might also be the beginning of something. She is not ready to find out. She is also not ready to begin the conversation with her daughter that finding out would entail. So she does what most people in her position do. She watches herself, quietly, with a kind of vigilance she keeps hidden from the people she loves.
The cognitive concierge is the agent that serves Margaret across this trajectory. It is the most ethically complex agent in the BlueMirror system because it serves people whose capacity to consent to being served is changing. The architecture must adapt to her cognitive state without requiring her to acknowledge the change. It must protect her dignity while providing support. It must involve her family at the right moments, in the right ways, without infantilizing her at any moment that does not warrant escalation. Every design decision in this agent has an ethical dimension that cannot be separated from the technical implementation.
The dignity constraint#
The governing design principle. Every interaction the cognitive concierge has must pass the dignity test: would a competent, caring human companion handle this the same way? If the system would embarrass a thoughtful human companion, the design has failed.
This is not a soft principle. It is a hard constraint that drives technical decisions throughout the agent. Latency is a dignity metric: the person who is disoriented and asks “What day is it?” needs an answer in under one second. Five seconds of silence is five seconds of visible confusion that undermines her sense of competence. The latency requirement drives the edge-only deployment of the memory care infrastructure agents. Network round-trips are too slow.
Language is a dignity metric. The system speaks to the person Margaret is, not the person she used to be and not the person her diagnosis might suggest she is becoming. It does not adopt the simplified, infantilizing register that healthcare environments often default to with cognitively impaired patients. It adapts language complexity continuously based on the Cognitive State Estimator’s reading of her current state, but adapts toward simplicity, not toward condescension. The threshold the design holds: every utterance the system produces should be one a respectful human friend would use in the same moment.
Memory is a dignity metric. The system remembers what Margaret has told it and does not require her to repeat herself. The person whose memory is changing finds it humiliating to be asked, repeatedly, the same questions she has answered before. The architecture refuses this failure mode. Context persistence (Series 05) is not a feature of the cognitive concierge alone; it is a foundational property. The cognitive concierge depends on it more deeply than other agents because its users are precisely the population for whom repetition is most damaging.
The dignity constraint shapes what the system does and what it refuses to do. The system does not, for example, score Margaret’s cognitive state and report the score to her, even when she asks. The score is an internal signal that drives behavioral adaptation. Reporting the score to the person whose score it is would communicate something the architecture does not believe is true: that her worth as a person is a function of a number a model produces. The system does not do this. It would embarrass a human companion to do it. The architecture refuses.
Six infrastructure agents#
The cognitive concierge composes from six infrastructure agents, more than any other concierge agent in the system. Five of the six deploy edge-only because the latency requirements are non-negotiable.
The Orientation Assistant answers the time, place, and person questions that arise as orientation slips. It is calibrated to be ready: the question “What day is it?” returns an answer in under 800ms with high probability and under 1.5 seconds in the worst case. It runs at 0.75 autonomy because the situations it addresses require immediate response. Edge only.
The Reminiscence Facilitator holds the structured life-story context that grounds Margaret in her own history. It surfaces memory prompts that connect the present moment to her past in ways that comfort and orient her: the photograph of her wedding day, the recording of her late husband’s voice from a 2018 birthday message, the song that she has loved since she was nineteen. It runs at 0.75 autonomy. Edge only.
The Routine Anchor maintains the structure of Margaret’s day. Wake routine, meals, medication, activities, sleep routine. When the routine wavers, the agent gently re-anchors. The routine is one of the most powerful cognitive scaffolds available. People with cognitive change who maintain routine maintain function longer. The agent’s contribution is keeping the routine consistent across days that would otherwise drift. It runs at 0.75 autonomy. Edge only.
The Wandering Prevention infrastructure agent is the most operationally consequential. It maintains location awareness within the home, recognizes when Margaret moves toward exits at unusual times or in unusual states, and offers gentle redirection: “Margaret, it’s three in the morning, would you like me to make you some tea?” rather than alarm or restraint. The agent is also integrated with the home environment concierge for the cases when redirection fails. It runs at 0.75 autonomy. Edge only.
The Sundowning Support agent addresses the late-afternoon and early-evening anxiety that affects many people with cognitive change. The pattern is well-documented: as light fades, agitation rises. The agent recognizes the pattern in Margaret’s behavior, adjusts the home environment (lighting, temperature, music), surfaces calming activities, and signals her caregiver if the pattern intensifies beyond the threshold the family has set. It runs at 0.75 autonomy. Edge only.
The Communication Adapter runs underneath the others, continuously adjusting language complexity based on the Cognitive State Estimator’s reading. It does not announce its work. The person speaks to the cognitive concierge. The Communication Adapter ensures the response is calibrated to Margaret’s current capacity to process it. It runs at 0.75 autonomy. Edge only.
All six on the edge. The edge requirement is not an architectural preference. It is the dignity constraint expressed in deployment terms.
The SLM stack for memory care#
Five dedicated models power the memory care work. The total parameter count for memory care alone is 700 million, more than is dedicated to any other concierge function. The investment reflects the complexity of the work.
The Orientation Assessor runs at 150M parameters and targets under 50ms inference. Its job is the time-place-person grounding check: does Margaret know what day it is, where she is, who she is with? The model is sized at 150M because the assessment has to be both accurate and fast, and faster than the response itself. The assessment runs in the background of every interaction.
The Cognitive State Estimator runs at 200M parameters and targets under 75ms. It is the same model that serves the health concierge. Lucidity level detection from conversational signals: vocabulary range, response latency, semantic coherence, repetition patterns. The output drives the behavioral adaptation across the system. Edge only.
The Confusion Detector runs at 100M parameters and targets under 50ms. Its job is the binary detection of acute confusion as opposed to baseline cognitive state. The distinction matters: a sustained period of confusion that exceeds the person’s typical fluctuation may signal a urinary tract infection, a medication side effect, a sleep deficit, or a transient ischemic attack. The agent escalates to the health concierge, which routes the question to the clinician. Confusion detection is a different model from cognitive state estimation because the response surfaces are different and the latency budget is tighter.
The Reminiscence Prompter runs at 150M parameters and targets under 75ms. Its job is to surface memory prompts from Margaret’s life-story context that connect to the present moment in ways that orient and comfort. The model is sized to handle the structured generation of relevant prompts from a deep context that may include decades of personal history.
The Simplification Engine runs at 100M parameters and targets under 50ms. Its job is to adjust the language complexity of any output before delivery. It is downstream of the response generators throughout the system. Every response intended for Margaret routes through the Simplification Engine for final calibration before being spoken or displayed.
The five models are kept separate rather than consolidated into one larger general model because specialization enables the sub-50ms latency on the time-critical functions. Confusion detection running inside a 700M general model would be slower than confusion detection running on a dedicated 100M specialist. The total compute is comparable. The latency profile is incomparable.
Adaptation without acknowledgment#
The architecture’s most important behavioral property: the system adjusts as Margaret’s capacity changes without asking her to acknowledge the change.
Language gets simpler. Sentences shorten. Vocabulary narrows toward the words she still uses with confidence. Visual cues replace textual ones where displays are available. Reminders increase in frequency, decrease in complexity. Choice presentations narrow from four options to three to two. The transition is continuous rather than stepped. The person never receives a notification that says “your cognitive score has declined.” The system just becomes gentler, more patient, more repetitive in its reassurance.
This is harder than it sounds. The naïve approach would be to define modes (“normal mode,” “moderate impairment mode,” “advanced impairment mode”) and switch between them at thresholds. This approach fails because the thresholds are visible to the person. The day the system switches modes is the day she notices that something has changed about how it speaks to her. The notice is itself the harm. She did not need to be told. The system told her by changing.
The architecture refuses the mode switch. Adaptation is continuous, parameterized by the Cognitive State Estimator’s continuous output rather than driven by discrete thresholds. A small drop in capacity produces a small adjustment in language complexity, not a categorical shift. The person who is paying attention to changes in herself notices nothing because there is nothing to notice. The system’s adaptation tracks her own variability inside the noise floor of normal day-to-day fluctuation.
The honest limitation: the technique works smoothly in the middle of the cognitive trajectory and works less smoothly at the extremes. When the change is large enough to require fundamentally different interaction patterns (when, for example, voice interaction becomes more reliable than text, or when family caregiver involvement needs to grow substantially), the architecture’s continuous adaptation alone is not sufficient. The escalation framework (Series 04) governs those transitions, with consultation between the agent, the person, and the family. The transitions exist. The architecture does not pretend they do not. It defers them as long as continuous adaptation is sufficient.
The cross-concierge effect#
When the Cognitive State Estimator detects a low-capacity day, every other concierge agent adjusts. The earning concierge reschedules the four o’clock cooking class with the student in Brisbane to next week and frames the rescheduling to the student as a routine timezone reshuffle, not a health-related change. The buying agent defers the substitution review to tomorrow rather than asking Margaret to make decisions during a difficult cognitive moment. The financial concierge raises approval thresholds for any non-routine action: today is not a day to authorize a Roth conversion. The home environment adjusts lighting and temperature to settings that have historically supported Margaret on her better days within the same general baseline.
One assessment, thirteen adjustments. This is the integration argument made concrete. No standalone app can do this because no standalone app holds the cognitive context. The cognitive concierge holds the context, and every other concierge reads from it.
The cross-concierge cognitive state propagation is one of the architecture’s most potent properties and one of its most sensitive. The propagation is what makes the system responsive to the person across her actual day. The propagation is also what could become a surveillance pattern if the architecture were not careful about what it shares and with whom. The privacy framework (Series 04 and Series 05) is what governs the boundary. Every concierge agent sees the cognitive state estimate. No external party sees it. The family does not see the state. The clinician does not see the state without the person’s explicit consent and a specific clinical purpose. The state is internal.
Family notification boundaries#
What the family sees, when, and how. The daughter receives a weekly summary, not real-time cognitive scores. The daily fluctuations are normal; sharing them creates anxiety without actionable insight. The agent notifies family when a trend persists for three or more weeks (a sustained change is meaningful where day-to-day variation is not), when a safety event occurs (a fall, a wandering episode, a medication error with consequences), or when the escalation hierarchy requires family involvement (a clinical concern that requires a decision Margaret has consented her family to participate in).
The person’s privacy is preserved even within the family. Margaret has not told her daughter about the moments she has noticed. The system does not tell the daughter either, until Margaret tells the system to or the threshold for clinical concern is crossed and the family is part of the consented escalation.
This is not the only way to design the system. A different design would surface every signal to the family on the assumption that more information is better. The architecture rejects this design because the assumption is wrong. More information about Margaret’s interior life, distributed to family members who cannot act on most of it, produces anxiety rather than help. The family that lives at a distance becomes consumed with worry. The family that lives nearby becomes overprotective. Neither response improves Margaret’s life. The architecture’s contribution is to surface to the family what the family can act on, when the family can act on it, in a structure that respects Margaret’s continuing authority over her own information.
The next article addresses the caregiver concierge: the agent that serves the person caring for the person, and recognizes that the caregiver is often aging too.
Cross-References#
The Memory Exoskeleton (BML-04 series). The editorial framing of cognitive support from the user’s perspective, including the human texture of what continuous adaptation feels like to the person being adapted to.
The Home Environment Concierge (BMT-01.12). The agent that the cognitive concierge directs to adjust environmental factors (lighting, temperature, sound) in support of cognitive function and safety.
Cognitive Capacity and Consent (BMT-04.05). The ethical framework that governs how authority shifts as capacity changes, including the consented escalation that the family notification protocol depends on.
Irrationality Protection (BMT-11.03). The IVQ framework applied to cognitive vulnerability, which extends the protections covered here into the broader equity engineering of the system.
Technical Appendix BMT-01.07-A is available to partners and investors at partners.bluemirror.tech.