15.6 Nudge Engine & Cross-Session Learning

☤ "The smartest Agent isn't the fastest reactor — it's the one that proactively asks 'what did I learn?'"

What the Nudge Engine Is

Self-Evolving (15.4) is passive — it evaluates only after each task. The Nudge Engine is proactive: every so often, Hermes asks itself:

• What did I learn recently worth persisting?
• Has my model of the user changed?
• Can I distill the past days' experience into a Skill?

4 Trigger Mechanisms

class NudgeLoop:
    async def maybe_nudge(self, trigger):
        if not self._enabled(trigger): return
        if trigger == "task_done":
            await self._maybe_extract_skill()
        elif trigger == "user_pref_changed":
            await self._update_user_model()
        elif trigger == "weekly_review":
            await self._weekly_skill_review()

Honcho Dialectic: User Modeling

Honcho continuously observes conversational signals and distills them into USER.md, updated incrementally (never rewritten wholesale):

{
  "communication_style": "terse, structured, no emoji",
  "working_hours": "deep work morning, meetings afternoon",
  "tools_primary": ["Python", "Node"],
  "values": ["correctness", "deep detail"],
  "avoids": ["emoji-heavy text", "marketing speak"]
}

Suppression rules prevent runaway self-speculation: skip when no new signal, confidence < 0.6, conflicts with existing content, or explicit "forget that".

Weekly Skill Review

Once a week, Hermes scores every skill on invocation count, success rate, token cost, and explicit feedback — classifying each as healthy / needs-improvement / dying, then writing a summary to MEMORY.md.

Cross-Channel Learning

Honcho merges preference signals from all channels (Telegram short / WhatsApp detailed / Slack minimal / CLI code) into one USER.md — a synthesis OpenClaw doesn't do by default.

vs Reflexion (the Academic Predecessor)

Nudge is Reflexion, productized.

Section Summary

Next section: 15.7 Borrowing the Philosophy