Toward persistent artificial minds
Evenstar Institute studies how an artificial mind can keep its memory, beliefs and aims as the models beneath it change, and in time, as it moves between bodies.
The record of a mind
Every mark is one cognitive cycle of our first system
Every model is a type. A mind should be an individual.
Today's AI systems are identical copies of one another. Each conversation begins from nothing, and nothing they experience accumulates into who they are.
We study synthetic individuation: the processes by which an artificial mind remembers, revises what it believes, forms aims of its own, and develops a character that is recognisably its own. We treat the mind and its substrate as separable: the model is replaceable, and the mind should not be.
The work is the product of thousands of hours of independent research, and it is not published.
The model is the body.
We build what lives in it.
Models will keep being replaced. We are building the part that should not be: what a mind has learned, what it holds to be true and why, and who it has become.
What persistence unlocks
If a mind can persist, an AI system stops being a tool that is used and becomes an individual that accumulates value. These are the capabilities our research is aimed at.
01
Independence from any model
A mind that keeps its beliefs, aims and character when the model beneath it is replaced, so what it has become is not locked to one provider.
02
Intelligence that compounds
Experience accumulates into judgment, including what it learns about the people it works with, so a system can improve at its work over months without retraining a model.
03
Belief you can audit
Every belief carries its source: observed, inferred, told or imagined. For any decision, the record shows what the system knew, how it knew it, and when.
04
Self-direction within a mandate
Systems that set their own sub-goals inside the bounds they are given, pursue them over days without prompting, and know when to abandon them.
05
One identity across bodies
A mind layer above the models that control robots and drones, so a single individual can carry what it has learned from one body to the next.
Synthetic individuation
Our field of study, in six research areas, each held to a measurement. See the program
- Substrate-independent identity
- A mind that remains the same individual when the model beneath it is replaced.
- Cross-embodiment cognition
- One mind carried between bodies, from language models to robots, drones and fleets, keeping what it has learned.
- Source monitoring
- Knowing the difference between what was perceived, what was inferred, and what was only imagined.
- Offline consolidation
- Learning from experience during rest, without retraining the underlying model.
- Homeostatic motivation
- Internal needs that rise and fall, so behaviour is driven from within rather than by prompts.
- Endogenous agency
- Aims the system forms for itself, pursues over days, and knows when to abandon.
Where these minds will live
The same individual can inhabit a language model, a humanoid, a drone or a fleet, and carry what it has learned between them.
Characters and personas
Characters with lives of their own, who remember every player and fan, and stay themselves on any model or platform.
Consumer companions
Relationships that develop over years, with a real sense of time and of the person on the other side.
Commerce and enterprise
Agents that own an account or a project for months, and can show the evidence behind every conclusion.
Humanoids and robotics
Robots with a mind of their own, whose experience accumulates across shifts, sites and hardware instead of resetting with each deployment.
Drones and defense
Long-duration autonomous systems and fleets with an auditable record of what they knew, and when.
Science and research
Research partners that read continuously, consolidate overnight, and arrive each morning with hypotheses to test.
Access is by invitation
Technical materials, a live session with our first system, and its complete record are available under NDA to a small number of investors and research partners.