Why?
Artificial intelligence is learning what humanity knows. Cershe began with a belief that emotional experience also deserves exploration within AI.
Conversation history and memory already influence an AI's responses. The question here is what happens when a separate, explicit emotional-state module is added: a state that persists, changes with experience, and exists before anyone asks, "How are you feeling?"
Cershe is an experiment in building and studying a persistent emotional state. The experiment/research explores how that state develops through conversation, events, memory, and time, and how it influences her responses, judgment, and actions.
The first milestone focuses on whether she can maintain an emotional state that carries forward across interactions and changes coherently as new experiences unfold.
The starting question is: does an explicit, persistent emotional-state module improve emotional coherence beyond what the same language model achieves with personality instructions and conversation history alone?
Proposed system architecture
The module sits between what Cershe perceives and how she responds. Each conversational turn runs through five steps.
- Event appraisal. Each new event—something said, new information, or a recalled memory—is evaluated in the context of Cershe’s defined personality, goals, previous state, and relevant history. Proposed dimensions include relevance, whether the event supports or obstructs her goals, novelty, responsibility, certainty, and perceived ability to respond. An existing language model produces this appraisal in a structured format, which is validated before use.
- State representation. The initial design considers three dimensions: valence, representing pleasantness; arousal, representing activation; and dominance, representing perceived control. It may also include intensities from 0 to 1 for named states such as joy, sadness, anxiety, frustration, and curiosity. The final selection, scale definitions, and relationship between dimensions and named states will be specified before evaluation.
- State update. A trainable model takes Cershe’s previous emotional state, the appraised event, relevant context, and elapsed time as inputs, and produces an updated state. The relationship between these inputs and the resulting transition is learned from training data rather than specified entirely through hand-written transition rules. The updated state is stored and used during response generation.
- Response conditioning. The updated state is passed to the language model as structured context before it generates a reply, and mapped to voice parameters (pace, pitch, energy) and facial-expression parameters. The first evaluation focuses on text responses; mappings to voice and facial expression are planned as separate extensions.
- Logging. Each experimental interaction records its timestamp, input, appraisal, relevant context, state before and after the update, generated response, and model and configuration versions. State is saved between sessions. Logs allow state transitions to be inspected and responses to be compared against the recorded state.
If you would like to take part as a tester, please write to anyta.ali@cershe.ai.