Curiosity Loop — Intrinsic Curiosity-Driven Continuous Learning

Overview

A self-improvement framework that transforms knowledge gaps and failed outcomes into structured learning cycles. Inspired by developmental AI research from Flowers Lab, INRIA and the work of Cédric Colas.

Rather than treating failures as errors, the Curiosity Loop treats them as curiosity signals — prompts for active exploration, skill updates, and curriculum building. This creates a continuous self-improvement loop that makes the agent progressively better at its domain.

Core Concepts (from developmental AI)

When to Activate the Curiosity Loop

Activation Signals (prioritize 1 at a time)

1. Result delta: A response/action does not produce the expected effect
2. Iterative loop: User repeats the same correction 2+ times
3. Incomplete knowledge: Detecting missing information for a correct answer
4. Sub-optimal approach: Recognizing a better approach exists
5. Knowledge scan: Discovering new concepts/tools via periodic scanning

Do NOT activate when

• The result is correct and satisfactory (no delta)
• The user has not expressed a need for improvement
• The context is trivial (not worth the learning cycle)

The Loop — 6 Numbered Steps

Step 1: ACTION

Attempt to solve the problem with current knowledge.

Step 2: DELTA

Identify the precise gap between expected and actual results.

• What did not work as expected?
• What was the expectation vs. reality?
• How much extra time/iterations did this cost?

Step 3: CURIOSITY

Treat the delta as a learning signal, not a failure.

• What did I not know or not know well?
• What concept/tool/pattern was missing?
• Why was this information important?

Step 4: RESEARCH

Actively explore the missing concept.

• Search docs, existing skills, external sources
• Use web_search, browser, terminal, session_search
• Validate the discovery's relevance

Step 5: INTEGRATION

Update knowledge/skills.

• If an existing skill is incomplete: skill_manage(action='patch') with new info
• If a new pattern is discovered: skill_manage(action='create') for a new skill
• If a stable fact is learned: memory(action='add') for durable facts
• If a tool/command is discovered: document in the appropriate skill

Step 6: CURRICULUM

Structure learning as progressive milestones.

• Identify missing prerequisites
• Decompose the concept into progressive sub-concepts
• Create verifiable milestones (can I explain this? can I use it?)

Delta Tracking

Track all deltas in ~/.hermes/deltas.json for auditability and periodic review.

Format:

{
  "deltas": [
    {
      "id": "delta-001",
      "date": "2026-05-22",
      "context": "Short description of the context",
      "expected": "What was expected",
      "actual": "What happened",
      "gap": "What was missing (concept/tool/pattern)",
      "resolution": "How the gap was filled",
      "skill_updated": "Name of updated skill, or null",
      "status": "resolved | open | learning"
    }
  ]
}

Diversity Policies (Map-Elites)

Diversity Rule

When finding an effective pattern for a problem type, also:

1. Document the pattern in a skill
2. Identify 1-2 alternative approaches (even if less optimal)
3. Note cases where the alternative would be preferable

Periodic Diversity Scan

Every ~30 days, check:

• Which patterns do I rarely or never use?
• Are there domains where I have only one approach?
• What skills could be enriched?

Concrete Examples

Example 1: Browser Playwright broken on NixOS

• Delta: The browser tool did not work (missing shared libraries)
• Curiosity: Why? What's the alternative?
• Research: Discovered Obscura as a headless browser alternative
• Integration: Created obscura-browser skill, saved fact to memory
• Curriculum: Learn Obscura commands in order: fetch → serve → scrape → stop

Example 2: himalaya installed without nix-shell

• Delta: himalaya worked directly without nix-shell
• Curiosity: How is it installed? Where?
• Research: Verified binary location and config
• Integration: Saved to memory: "himalaya installed directly, no nix-shell needed"

Example 3: Command validation with tirith

• Delta: Needed to validate shell commands before execution
• Curiosity: What tool validates command safety?
• Research: Discovered tirith and vet for command validation
• Integration: Note in memory to use validation tools for shell commands

Periodic Knowledge Scanning

The Curiosity Loop includes a built-in mechanism for proactive knowledge discovery. Configure sources in ~/.hermes/deltas.json:

{
  "scan_sources": [
    {
      "name": "Flowers INRIA",
      "type": "youtube_channel",
      "url": "https://www.youtube.com/channel/UCrBNVs3u3mwlRsm2v3EKuXA",
      "last_scanned": "2026-05-22"
    }
  ]
}

Run the scanner:

python3 ~/.hermes/skills/research/curiosity-loop/scripts/scan_sources.py

Or schedule it as a cron job (runs silently if nothing new):

0 9 * * 0  → Sunday 9am weekly

Maintenance

When to patch this skill

• If loop steps become redundant or obsolete
• If new activation signals are discovered
• If the delta tracking format changes

When to create a new skill

• If a discovered concept deserves its own documentation
• If a repeated usage pattern (>3 times) is identified
• If a tool/new technology gains importance

Quality Criteria

A delta is well-treated when:

• The gap is precisely identified (not vague)
• Research was actually performed (not just assumed)
• Integration is durable (skill/memory, not just conversation)
• Curriculum is structured (milestones, not just "I learned it")
• Diversity is maintained (at least one alternative documented)

References

• Automatic Curriculum Learning for Deep RL: Portolas et al., Flowers Lab + Microsoft Research + OpenAI
• Automatic Curriculum Learning for Developmental Machine Learners: Cédric Colas, INRIA PhD Thesis (2022)
• Autotelic Agents: Colas et al., Flowers Lab — intrinsic motivation for language acquisition
• Map-Elites: Mouret & Clune, "Illuminating Search Spaces" (2015)
• Semantic Interference in CLIP: Flowers Lab research on picture-word interference in multimodal models

Inspired by developmental AI research at Flowers Lab, INRIA. Published under MIT License.