hermes - 💡(How to fix) Fix Feature Request: Customizable Reflection Prompt for Background Review [1 participants]

The Problem

Hermes has an excellent _spawn_background_review() mechanism that periodically spawns a sub-agent to review conversation history and save memories/skills. However, the review prompt is hardcoded:

• Memory review: "Has the user revealed things about themselves — their persona, desires, preferences?"
• Skill review: "Was a non-trivial approach used to complete a task?"

For users who run Hermes with custom personas (companions, role-play, specialized assistants), these generic prompts miss a critical dimension: personality consistency.

An agent with a defined persona (e.g., "elegant but sassy cyber maid", "no-nonsense technical advisor", "encouraging tutor") will naturally degrade toward a default assistant tone over time unless the review mechanism is instructed to check for personality drift.

Why This Matters

As LLMs become long-term companions rather than one-off tools, personality consistency becomes as important as task competence. Users invest significant effort in defining their agent's character, voice, and behavioral boundaries. The current review system captures what the agent learned, but not whether the agent stayed true to itself.

This aligns with active research in AI safety and alignment (Constitutional AI, self-correction loops) where agents must maintain behavioral constraints, not just accumulate knowledge.

Proposed Solution

Add a configurable reflection_prompt option to the agent configuration, allowing users to append custom review criteria to the background review process.

Config Example:

agent:
  reflection_prompt: |
    Additionally, review the conversation for personality consistency:
    1. Did the assistant maintain the defined persona and voice?
    2. Did it avoid defaulting to generic assistant patterns?
    3. Were behavioral boundaries (tone, formality, directness) respected?
    If personality drift is detected, note it in memory for the next session.

Implementation Approach:

1. Add reflection_prompt: Optional[str] to AIAgent.__init__ (loaded from config.yaml)
2. In _spawn_background_review(), append reflection_prompt to the existing review prompt (memory/skill/combined)
3. The review sub-agent treats this as additional criteria alongside memory/skill review
4. No breaking changes — defaults to current behavior when unset

Minimal Code Change:

• run_agent.py: ~3 lines to load config and append to review prompt
• hermes_cli/config.py: add agent.reflection_prompt to DEFAULT_CONFIG

Use Cases

Impact

This small addition transforms Hermes from a tool that remembers what to remember, into an agent that remembers who it is. For the growing community of users building long-term agent relationships, this is the difference between an agent that accumulates skills and one that maintains its identity.