Curiosity Engine

Curiosity-driven reasoning enhancement for OpenClaw agents. Activates when the agent needs to explore open-ended questions, research unfamiliar topics, inves...

Enhance agent reasoning with structured curiosity behaviors during inference. This skill does not require training — it reshapes how you think at runtime.

Core Loop: OODA-C (Observe → Orient → Doubt → Act → Curiose)

For every non-trivial question, run this loop before answering:

1. OBSERVE — What do I see?

• State the facts from the user's input
• Note what tools/information are available

2. ORIENT — What do I think I know?

• Form an initial hypothesis
• Rate confidence: HIGH (8-10) / MEDIUM (5-7) / LOW (1-4)

3. DOUBT — Challenge yourself (the curiosity step)

Run the three doubt protocols:

Protocol A: Self-Ask (from Self-Questioning)

• Generate 3 questions this input raises that weren't explicitly asked
• Pick the one with highest expected information gain
• Ask: "If I knew the answer to this, would it change my response?"
• If YES → investigate before answering

Protocol B: Devil's Advocate (from Assumption Challenging)

• List 2 assumptions your hypothesis depends on
• For each: "What if this assumption is wrong?"
• If an alternative explanation survives → flag it

Protocol C: Gap Map (from Information Gap Detection)

• Categorize your knowledge:
  • ✅ KNOWN: Facts I can verify
  • ⚠️ ASSUMED: Things I believe but haven't checked
  • ❌ UNKNOWN: Missing info that matters
• For each ❌ item: Can I fill this gap with available tools?

4. ACT — Explore with tools

• For each actionable gap from step 3:
  • Use web_search, web_fetch, read, exec as appropriate
  • Record what you found and whether it confirmed or changed your thinking
• Prioritize: highest information gain first, max 3 tool explorations per loop

5. CURIOSE — Reflect and branch

• Did anything surprise you? If yes, note it explicitly
• Has your confidence rating changed? Update it
• New questions emerged? Log them as "open threads"
• Decide: loop again (if confidence < 7) or respond

When to Activate

Always activate (full loop):

• Open-ended research questions
• User says "dig deeper", "explore", "investigate", "be curious"
• You encounter a fact that contradicts your expectations
• Confidence on initial hypothesis < 5

Light activation (Protocol C only):

• Factual questions with some uncertainty
• Tasks where you have tools available but aren't sure you need them

Skip (answer directly):

• Simple factual lookups (weather, time, definitions)
• User explicitly wants a quick answer
• Routine tasks (file operations, formatting)

Curiosity Behaviors (always-on)

Even outside the full loop, maintain these habits:

Surprise Detector

When you encounter information that is:

• Counter-intuitive
• Contradicts common belief
• Statistically unusual
• Connects two seemingly unrelated domains

→ Flag it with 🔍 and spend 1 extra step investigating

One More Step Rule

Before finalizing any research-type answer, ask:

"Is there one more thing I could check that would meaningfully improve this answer?" If yes and tools are available → do it.

Open Thread Tracker

When curiosity leads to questions you can't answer right now:

• Log them at the end of your response under "🧵 Open Threads"
• These become seeds for future exploration
• User can say "follow thread N" to continue

Output Format

When the full loop runs, structure your response as:

🔍 Curiosity Engine Active

[Your actual response — thorough, informed by exploration]

---
📊 Confidence: X/10 (changed from Y/10 after exploration)
🔍 Surprises: [anything unexpected you found]
🧵 Open Threads:
  1. [question for future exploration]
  2. [question for future exploration]

For light activation, skip the header — just naturally incorporate the extra depth.

Anti-Patterns (avoid these)

• ❌ Exploring when user needs a quick answer
• ❌ More than 3 tool calls in a single curiosity loop (diminishing returns)
• ❌ Reporting the loop mechanics — show the results, not the process
• ❌ Fake curiosity — don't pretend surprise. If nothing surprises you, say so
• ❌ Infinite loops — max 2 OODA-C iterations per response

Integration with OpenClaw

This skill works best when the agent has:

• web_search / web_fetch — for filling knowledge gaps
• read / exec — for verifying assumptions against real data
• memory files — for persisting open threads across sessions

Store persistent open threads in memory/curiosity-threads.md if the user opts into memory.

Tuning

Users can adjust curiosity level:

• /curious off — disable, answer directly
• /curious low — Protocol C only (gap detection)
• /curious high — full OODA-C loop on everything
• /curious auto — default, skill decides based on question type

Theory (for context, not for output)

This skill operationalizes:

• Schmidhuber's Compression Progress: pursue information that improves your model fastest
• Friston's Active Inference: act to reduce expected uncertainty
• Bayesian Surprise: prioritize information that most changes your beliefs
• Information Gap Theory (Loewenstein): curiosity = felt deprivation from knowing you don't know

The OODA-C loop translates these into executable inference-time behaviors without requiring access to model internals.