Personas

A persona defines how an agent behaves. The authored system prompt stays verbatim at runtime; structured settings drive guardrails and helper workflows.

Runtime prompt

The customer-authored system prompt is the authoritative behavior instruction for the agent. Hyponema prepends memory context at runtime, but does not rewrite the persona’s authored prompt.

Structured settings

Persona configuration controls behavior around the prompt:

• First message and greeting behavior.
• No-go zones and redirect or escalation policies.
• Sleep windows.
• Escalation rules.
• Personality scales and identity metadata used by UI helpers.

These settings are structured so the dashboard can validate and version them independently from the prompt text.

Versions

Each meaningful persona change creates a version. Promote known-good versions to production and use rollback when a prompt or guardrail change underperforms.

Greetings

Two greeting modes are supported:

• static (default) — render the persona’s authored first message verbatim.
• custom_agent — run a short greeting prompt (≤4000 chars) with an allowlist of read-only memory tools so the first message can acknowledge context. One-shot tool-use loop with a small iteration cap and a short timeout. Setting fallback_to_static falls back to the static first message on timeout or error.

Per-conversation overrides take precedence over the persona’s configured mode.

Persona versus agent

A persona defines how the AI should behave (prompt, guardrails, greetings, escalation). An agent wraps that persona with the providers and configuration the runtime actually executes against — voice stack, tools, knowledge, dynamic variables, phone numbers. One persona can be reused across multiple agents.