Skip to main content

02. Human Model And Memory

Goal​

Seraph needs a layered model of the human, not just a conversation transcript.

Layers​

Identity layer​

  • values
  • long-horizon goals
  • preferences
  • recurring constraints
  • self-story and direction

Working state layer​

  • current focus
  • calendar pressure
  • current commitments
  • recent conversations
  • screen and environment context

Learned pattern layer​

  • attention patterns
  • procrastination and avoidance patterns
  • intervention receptivity
  • timing preferences
  • decision strengths and blind spots

Current Design Direction​

The present system already has the right raw surfaces:

  • soul file for durable identity
  • vector memory for long-term recall
  • goals for explicit commitments
  • observer refresh for live state
  • strategist and scheduler outputs for periodic reflection

The open research problem is not whether Seraph stores enough. It is whether those layers become one coherent world model with measurable intervention quality.

Research Questions​

  • how should Seraph distinguish identity from transient mood or noise?
  • what should become long-term memory versus expire?
  • how should intervention outcomes rewrite the user model?
  • what should be explicit in the soul versus learned implicitly?