Working on more believable computer opponents in Chessfolk

[space64]

Hi everyone,
I’m working on a chess game called Chessfolk, focused on offline play against computer opponents. It has fictional players, tournaments and an internal rating system, with the aim of giving individual games a bit of continuity.
The part I’m finding most interesting—and difficult—is making weaker opponents believable. Our setup combines Maia and Stockfish with some custom move-selection rules. Sometimes the result is reasonably consistent; other times it feels like two different players are taking turns.
One player described a bot moving its queen away from an attacking rook, then putting it back in danger on the next move. Their point wasn’t that humans never make that mistake, but that the bot seemed to have forgotten its previous decision.
That has me thinking about how to preserve a little more intent between moves. A human might miss a defender because they’re concentrating on an attack, or stick with a plan after it stops working. Those mistakes have some connection to what came before.
Has anyone here experimented with modelling that kind of behaviour? I’d be interested in approaches that produce more coherent weaker play without making the opponent’s mistakes too predictable.