Soul (agent personality)
The Soul is the per-user personality and preferences config for the agent. A short markdown document layered on top of the base agent prompt, the Soul shapes how the agent talks to this specific user: tone, what to call them, what they care about, what to avoid.
Why it matters
Every user's relationship with their AI agent should feel different. The agent that talks to a 60-year-old salon owner about their work life should sound different from the one that talks to a 22-year-old college student about social plans. The agent that talks to someone who values brevity should sound different from the one talking to someone who likes thoroughness.
Building this customization through retraining or fine-tuning is impractical at scale. The pragmatic alternative is a per-user personality layer: a short markdown document the user (or the agent on their behalf) writes, layered into the agent's system prompt. The base agent stays consistent across users; the Soul personalizes.
What the Soul typically encodes:
- Name and how to address. "Call me Sara, not Sarah."
- Tone preferences. "Keep replies short," "more conversational," "less corporate."
- Topical preferences. "I care about my fitness; ask about that occasionally," "do not bring up politics."
- Stylistic markers. "Use emojis sparingly," "no gendered language," "match my message length."
The Soul evolves over time. A new user has a thin Soul; six months in, the Soul has accumulated nuance from how the user has corrected the agent, what the agent has noticed, and what the user has explicitly told it to remember.
How Ajeris handles this
Each Ajeris user has a Soul that lives in their per-user encrypted store. The agent reads the Soul into context at the start of every turn. Updates to the Soul happen organically: the user tells the agent "stop calling me 'sir,' just call me Mike" and the change saves to the Soul; the agent's tone adjusts immediately.
The user can view and edit their Soul through the manage page if they want explicit control. Most users never look at it; the natural-language updates are sufficient.
