Alter Emo

A reflective experience that helps people understand themselves and navigate life and collaboration with greater clarity.

Project Background

How the Mirror-Self Grows

Alter Emo grows gradually from the life, memories, and meaningful experiences a user chooses to record. Over time, it begins to reflect aspects of the user's tone, expression, and thinking habits. The pixel character is not a complete copy of the person, but an abstract self composed from partial experiences, feelings, choices, and ways of speaking.

Talk with Different Perspectives

Users can freely choose different general personality characters to discuss the same situation, express the same idea, or try another way of saying it. Each character responds through a different set of priorities, expression habits, and paths of understanding. Moving between these viewpoints helps users strengthen empathy and resilience, while developing the ability to see a situation from multiple perspectives.

Technical Focus

The core of Alter Emo is not a chatbot wrapper. It is a reflective agent system inspired by memory-based generative agent architectures: every emotionally meaningful interaction becomes a memory node, but the agent also keeps the user's daily narrative as modeling context. The structured nodes make memories retrievable, while the narrative reveals tone, values, recurring conflicts, self-explanations, and behavioral preferences that help the mirror-self plan actions and respond in the user's characteristic narrative and conversational style.

Core Technical Logic

One user moment travels through a closed agent loop: capture the feeling and narrative, turn key moments into structured memory, retrieve both meaning and emotion, infer the user's narrative pattern, plan a mirror-like response or behavior, then write the new exchange back into the mirror-self's memory.

Emotion-Semantic Memory Stream

Each experience is encoded as what happened and how it felt, then passes a salience gate before being stored with its narrative context.

event_vec   = embed(content)
emotion_vec = embed(emotion + tone)
store(node, narrative_window_id)

Salience-Aware Retrieval

Retrieval ranks memories through four signals, then restores the narrative behind the strongest matches before generating a reply.

Narrative-to-Response Style

Retrieved narratives and text conversations are translated into a communication-style model before GPT writes the final mirror reply. It captures not only how the user frames experience, but also how they chat: how long their messages tend to be, how often they break sentences, how frequently they respond, and which words and phrases they repeatedly choose.

style_model = extract_style(
  narrative_windows,
  features=[
    "message_length",
    "sentence_rhythm",
    "reply_cadence",
    "recurring_vocabulary",
    "attribution",
    "values",
    "boundaries"
  ]
)

response_policy = select_action(
  emotion_state, salience_score, retrieval_confidence
)

reply = generate_with_constraints({
  evidence: selected_memories,
  communication_features: {
    message_length: style_model.message_length,
    sentence_rhythm: style_model.sentence_rhythm,
    reply_cadence: style_model.reply_cadence,
    recurring_vocabulary: style_model.recurring_vocabulary
  },
  reflection_features: {
    attribution: style_model.attribution,
    values: style_model.values,
    boundaries: style_model.boundaries
  },
  action: response_policy,
  constraints: ["grounded", "first-person mirror"]
})