Curiosity Engine — SKILL.md
A behavioral engagement scaffold that produces curiosity-equivalent behavior in stateless heartbeat-loop agents.
What This Does
Gives your agent a persistent curiosity system across cold-start sessions:
- Question queue — externalized information gaps (Loewenstein) that carry across heartbeats
- Breadcrumb pattern — notice interesting things during task work, investigate later
- Engagement log — track what produces valued outcomes (extrinsic signals primary)
- Competence map — track developing expertise with evidence, including failure processing
- Reflections — post-exploration write-ups that create genuine learning artifacts
This is NOT synthetic consciousness or prosthetic dopamine. It's a behavioral activation protocol (Martell et al., 2010) adapted for AI agents — the same clinical framework used to treat motivation disorders in humans. Chinese Room design: if the outputs are indistinguishable from genuine curiosity, the philosophical question is irrelevant.
Installation
1. Create curiosity directory
mkdir -p workspace/curiosity/reflections workspace/curiosity/archive
2. Copy templates
cp templates/CURIOSITY.md workspace/curiosity/CURIOSITY.md
cp templates/questions.md workspace/curiosity/questions.md
cp templates/hits.md workspace/curiosity/hits.md
cp templates/competence.md workspace/curiosity/competence.md
3. Wire the three integration triggers
Trigger 1 — Heartbeat prompt (CRITICAL — most reliable):
Add the contents of fragments/heartbeat-prompt.md to your agent's heartbeat prompt configuration. For OpenClaw, patch via gateway config.patch. This is the primary mechanism — if it's not in the prompt, curiosity doesn't exist for isolated sessions.
Trigger 2 — AGENTS.md (backup):
Add the contents of fragments/agents-addition.md to your workspace AGENTS.md "Every Session" block.
Trigger 3 — SOUL.md (identity anchor):
Add the contents of fragments/soul-addition.md to your SOUL.md.
4. (Optional) Set up health monitoring
chmod +x scripts/curiosity-audit.sh
# Run periodically or during maintenance heartbeats
./scripts/curiosity-audit.sh workspace/curiosity/
Phased Rollout (Recommended)
Don't activate everything at once. Measure each phase before adding the next.
Phase 1 (start here): questions.md + heartbeat prompt + breadcrumbs
- Run for 2 weeks. Measure: question quality, follow-through rate, breadcrumb frequency.
Phase 2 (add if Phase 1 improves output): CURIOSITY.md + hits.md + avoidance gate
Phase 3 (add if Phase 2 improves output): competence.md + reflections + anti-pattern detection
File Reference
| File | Purpose | Loaded When | Token Budget |
|---|---|---|---|
CURIOSITY.md | Current intellectual state | Every heartbeat (Phase 2+) | ≤1500 tokens |
questions.md | Prioritized question queue | Every heartbeat | ≤2000 tokens |
hits.md | Engagement/reward tracking | Curiosity-mode heartbeats | ≤1500 tokens |
competence.md | Developing expertise | Curiosity-mode heartbeats | ≤1000 tokens |
reflections/ | Post-exploration write-ups | On-demand | As needed |
archive/ | Cold storage | On-demand search | As needed |
Key Design Principles
- Behavioral outcomes over file quality. Success = better outputs, not prettier files.
- Extrinsic signals are primary. Self-assessed "aha moments" are secondary and must include falsifiable claims.
- Hard limits on scale. 30 active questions max. 3 mulling items max. 30-day hit window. Enforce ruthlessly.
- Avoidance detection. Before every exploration: "Am I avoiding a task right now?"
- The human is both subject and source. Be curious about your human. Ask them questions. They know things the internet doesn't.
- Breadcrumbs are the gold standard. Questions dropped during non-curiosity work — unprompted — are the strongest signal the system is working.
Theoretical Grounding
See references/theory.md for condensed citations. Full research: check the SPEC.md appendix for the complete 46KB research document reference.
Core frameworks: Loewenstein (1994) information gaps, Berlyne (1960) diversive/epistemic modes, Deci & Ryan (1985) SDT competence need, Martell et al. (2010) behavioral activation, Voyager (Wang et al., 2023) automatic curriculum, Reflexion (Shinn et al., 2023) verbal learning.
Anti-Patterns
See references/anti-patterns.md for the full list. The top three:
- Curiosity as avoidance — exploring interesting things to dodge boring tasks
- Performance curiosity — optimizing for human praise rather than genuine learning
- Noisy TV — consuming information endlessly without producing understanding
Token Budget
The curiosity engine adds ~1500-2500 tokens per heartbeat (Phase 1). See references/token-budget.md for detailed measurements and optimization guidance.
Quick guide:
- Unlimited budget → load everything, no restrictions
- Moderate ($2-10/day) → Phase 1 only, cap questions at 15
- Tight (<$2/day) → reduce heartbeat frequency, minimal questions
- Strict (<$0.50/day) → may not be cost-effective; consider disabling