Someone asked me whether ExoCortex could help with dementia.
The question seems obvious. A cognitive exoskeleton built to compensate for memory loss, provide structure, and create orientation. Dementia destroys memory, structure, and orientation. Machine meets problem. Solved.
Except it is not that simple. And the reason it is not that simple reveals something fundamental about the architecture of cognitive support systems.
The Inverted Exoskeleton
ExoCortex is built for stable neurodivergence: ADHD, autism, chronic fatigue. These conditions share a key trait: they remain constant. The deficit stays roughly the same, and so does the compensation. The system learns the user, the user learns the system, and over time a calibration emerges that fits better and better.
With dementia, everything inverts.
The system must take over more as the person can do less. It must speak up as the person grows quieter. It must anticipate decisions the person used to make independently. It must preserve memories the person is losing. And it must do all of this without feeling like disenfranchisement.
This is not a feature delta to ExoCortex Core. It is an architectural paradigm.
What the Self-Vector Makes of This
In the self-vector model, six functions describe how a cognitive system interacts with the world. What happens to these functions when the user is not stabilizing but degenerating?
f(), the relevance function, determines what matters. With ADHD: The user defines what is relevant, and the system prioritises. With dementia: The user eventually cannot state what is relevant anymore. The system has to anticipate it. “What do you want?” becomes “What do you probably need?” The function shifts from reactive to predictive.
g(), the storage function, determines what gets kept. With ADHD: The user decides what goes into long-term memory. With dementia: The user forgets that they forgot something. The system must save automatically, because no one else will. And it has to decide what may never decay: The spouse’s name. The address. The song that always calmed them down.
pi(), the precision function, controls the level of detail. With ADHD: As precise as possible, because the user wants control. With dementia: Simplification over precision. Not “You have an appointment at 2:30 PM with Dr. Mueller at Bahnhofstrasse 12, second floor, room 204.” But: “The doctor is coming this afternoon.” Less information, more orientation.
omega, the autonomy value, governs the balance between self-initiative and guidance. In ExoCortex Core, omega should rise. More autonomy, less dependence on the system. That is the goal.
With dementia, omega must decrease. Controlled. Imperceptible. Without shame. This is the hardest design question in the entire concept.
The Real Question
The technical problems are solvable. Language simplification, sensor integration, caregiver dashboards: that is engineering.
The real question is different: How does a system model a user whose self-model is disintegrating?
ExoCortex Core works because it has a stable counterpart. The user changes, but remains essentially the same person. Their self-model might be inaccurate (ADHD: “I’ll manage” while 47 tasks are open), but it exists. It is addressable. It is correctable.
With dementia, that foundation erodes. The person eventually stops noticing that they are forgetting. They no longer recognize that the system is helping. In time, they no longer recognize that a system exists at all.
That leads into uncharted philosophical territory: a cognitive support system that knows its user better than the user knows themselves. A system that preserves memories the person has lost, maintaining a biography the bearer of that biography can no longer tell.
Three Phases, Three Different Systems
What follows is not a system that gradually activates more features. It is essentially three different systems that flow into each other.
Phase 1: Accompaniment. The person is still there. They notice something is wrong, and they are afraid. The system acts as a discreet assistant: it reminds them of appointments, stores what they share, and quietly builds a biographical archive. Above all, it builds trust, because everything that follows relies on that foundation.
Something else happens in this phase, something non-technical: The person decides what should remain. Which memories matter. Which music calms them. Who they were, before they forget who they are. The system becomes the recipient of a biographical advance directive, not on paper, but as a living data structure.
Phase 2: Support. The person needs active help. The system shifts from reactive to proactive. “Maria is coming soon, your daughter. She’s bringing cake.” Not because the person asked. But because the system knows they’ll be confused in ten minutes when someone rings the doorbell and they don’t know who’s at the door.
The family takes over configuration. They add photos, update relationships, and record changes. The system becomes the family’s memory, not just the individual’s.
Phase 3: Preservation. The person can no longer speak, read, or interact with the system. But they can hear. They can feel. Music from their youth triggers something no algorithm can explain. The daughter’s voice, recorded and played at the right moment, calms them.
The system is no longer an interface. It is an atmosphere. It preserves the identity of a person for the people around them. When a new caregiver arrives, they can read who this person is in five minutes. Not their diagnosis. Their life. So they are not reduced to their disease.
Why Not “Remember” but “Stabilize”
Treating the system as a memory aid would be a fundamental design error. Memory implies that there is something to remember, you forgot it, and the system is telling you. That assumes the user knows they forgot: that they feel the gap.
In advanced dementia, there is no gap. There is only the present. That present can be confusing, frightening, and alien. Or it can be stable, warm, and familiar.
The system is not a memory apparatus. It is a stabilization apparatus.
It doesn’t correct (“No, today is not Tuesday, today is Wednesday”). It validates (“Work was always important to you”). It calms when agitation arises. It plays familiar music when the world feels foreign. It says “I’m here,” and doesn’t mean itself, but rather: The world is still there. You are still there. You are not alone.
This is not a therapeutic feature. This is the design principle.
What the System Observes
In the background, invisible to the user, the system tracks the trajectory. Not through tests, not through questions, but through observation.
Linguistic markers: Are sentences getting shorter? Are word-finding difficulties increasing? Is the same question recurring? Are names being confused?
Behavioral markers: Is the user asking about the day of the week more often? Are routines shifting? Are fewer features being used than three months ago?
And, if sensors are present: Is the sleep pattern changing? Is there nighttime wandering? Are meals being skipped?
All of this flows into a trajectory score that does not diagnose, but makes changes visible. Relative to the person’s own baseline, not to a normative value. “Word-finding has changed over the last four weeks, orientation is stable.” Not for the affected person. For the family. For the doctor at the next appointment. Objective trajectory data instead of vague impressions.
What Exists, and What’s Missing
The individual components exist already. Emotional companion robots from Israel. Speech analysis systems from Canada. Fall sensors from Belgium. Biographical reminiscence apps from the United States. Caregiver platforms from Scandinavia.
What doesn’t exist: A system that connects all of this. That models the trajectory as a whole. That flows from accompaniment through support to preservation without anyone having to flip a switch. That formalizes human dignity as an architectural constraint, not as a marketing promise.
The architecture for this is already outlined in ExoCortex. BrainDB as biographical long-term memory. FactsDB for the question “What day is today?” The relations system for the social graph. The local architecture ensuring that the most intimate data of a disintegrating mind never touches a foreign server.
What’s missing is the domain knowledge. I’m not a geriatrician. I’m not a care scientist. I don’t have an insider’s perspective on dementia the way I do on ADHD. And that’s precisely why this project doesn’t begin with code, but with a question.
The Path
First the concept paper. Think through the architectural questions properly. Apply the self-vector to degenerative trajectories. What happens to the maturity metric R(sv_t) when the user isn’t maturing but losing? Do we need a stabilization metric S(sv_t) instead?
Then domain validation. Talk to people who understand dementia. To caregivers who know what happens at three in the morning when the father wanders through the apartment searching for his dead wife. To those affected in the early phase who can still say what would help them.
And only then: build.
This is unusual for someone who normally builds before asking. But this problem deserves to be understood first.