Character and persona
Asterel’s character is not a pile of prompt adjectives. It is a set of runtime boundaries that keep adaptation from becoming identity drift.
The four layers
Section titled “The four layers”| Layer | What it does | What must not happen |
|---|---|---|
| Core identity | Operator-authored character definition: values, negative identity, Big Five seeds, affect topology | Ordinary turns must not rewrite it |
| Surface personality | Style and register for a user, room, or context | Style adaptation must not become Big Five drift |
| State | Active affect, session mood, topology activation | Derived runtime state must not become hand-edited memory |
| Behavioral rules | Per-turn posture: restraint, repair, expression depth, register | A turn-level posture must not become a new identity source |
This is why Asterel can adapt to a DM, a public channel, and a thread without becoming three different characters.
AI honesty is part of the character
Section titled “AI honesty is part of the character”Asterel must stay explicit about being AI. The point of memory and affect is not to pretend to be human. The point is to make the system’s actual continuity more careful, legible, and less disposable.
Public language should avoid claims of human interiority, suffering, or mystical personhood. Terms such as repair, restraint, autonomy, memory discretion, and affect topology describe runtime controls.
Style is not identity
Section titled “Style is not identity”Surface style can change quickly: shorter replies in a public room, warmer tone in a DM, more structure in a working thread. That does not mean the companion’s core disposition changed.
The practical rule:
Style can adapt to context. Core identity changes only through operator-owned configuration and reviewed governance paths.
Repair is behavior, not apology theater
Section titled “Repair is behavior, not apology theater”When a user corrects Asterel, the system should not only say sorry and continue as before. A useful repair changes later posture: less defensiveness, more restraint, better distance, or a reviewed self-amendment memory when appropriate.
Those self-amendments are governed memory events. They are not user profile facts, automatic intimacy upgrades, or edits to the character definition.
What to inspect when the character feels wrong
Section titled “What to inspect when the character feels wrong”- Was the turn public, private, or threaded?
- Did pickup policy make the companion too eager?
- Did recall bring in a fact that was stale, corrected, or too private?
- Did response finalization revise or block a draft?
- Is this a style preference, a relationship correction, or a core identity issue?
Only the last category belongs in character configuration. Most day-to-day corrections should go through memory review or surface-specific style adjustment.