Is an AI character 'staying in character' one problem, or two: matching its own profile, and holding steady over time?

This explores whether 'staying in character' is one problem or two: consistency (does the persona's output match its profile and its own earlier lines) versus stability (does the persona hold its shape over time, across repeated runs, and under pressure). The corpus never draws this line in so many words. The distinction emerges from how the notes measure each one, and the two fail in different ways.

Consistency is a match property, checked against a reference. The user-simulator work scores it three ways: does a line fit the persona prompt, does it fit the previous lines, and does it survive a factual Q&A about the persona. Those checks catch three separate failures: local slips, drift across a whole conversation, and contradicted facts Can training user simulators reduce persona drift in dialogue?. Consistency is also easy to game. High scores often come from parroting the character description while ignoring what the other person just said, so consistency and dialogue coherence can pull against each other Do persona consistency metrics actually measure dialogue quality?. It doesn't come free with scale either. Claude 3.5 Sonnet beat GPT 3.5 by under 3% on persona consistency, apparently because training rewards per-turn quality rather than cross-turn coherence Does model capability translate to better persona consistency?.

Stability is a resistance property: what happens to the persona when something pushes on it. The notes show three kinds of push. The first is conversational pressure over time. The trained Assistant persona sits on one dominant axis of 'persona space', and emotional or meta-reflective conversations move the model off it in predictable ways. Capping activations along that axis limits harmful shifts without hurting capabilities How stable is the trained Assistant personality in language models?. The second is repetition. Run the same persona prompt many times, and the variation between runs matches or exceeds the variation between different personas. That means the model's own uncertainty, not stable knowledge of the persona, is steering the output Why do LLM persona prompts produce inconsistent outputs across runs?. The third is adversarial attack. Trained Assistant personas persist under jailbreak pressure where prompt-induced role-play collapses, which one philosophical account reads as evidence they are realized dispositions rather than performances Are RLHF personas performed characters or realized dispositions?. That sits awkwardly beside the drift finding above. A persona can be sturdy against attacks and still slide under gentle emotional conversation, and the notes don't reconcile the two.

The two properties can come apart. A layered, scripted persona that only reacts within a structured script produces human-like dialogue consistency, but at the cost of emotional expression Can layered persona architecture sustain coherent character behavior?. That is consistency bought by constraining the model. Going the other way, the sticky trained Assistant looks very stable, but stable at being itself. That is my inference from the notes, not something they state. Its stability doesn't guarantee it stays consistent with the character you asked it to play. The two also feed each other. The 55% drop in persona drift comes from using consistency scores as the training reward, so consistency is a lever for fixing one kind of stability failure.

Stability shouldn't be confused with being frozen. PersonaAgent rewrites its persona at test time to fit what a user wants, and the learned personas cluster distinctly by user Can personas evolve in real time to match what users actually want?. So there are really three things to tell apart. Consistency is not contradicting the intended persona. Stability is not changing when you didn't mean it to. Intended evolution is changing on purpose. The word 'drift' gets used for all three, and separating them shows which fix applies: better training signals, activation-level guardrails, or a different persona design.