curiocore 0.1.0

The curiosity engine for AI agents. Curiosity never rests.

🐾 CurioClaw

The curiosity engine for AI agents.

Curiosity never rests. —— February 2026, by Yiming 🐾

From HEAL Lab, CUHK-Shenzhen

Ziwen Wang · Zhangtianyi Chen · Zijian Wang · Yuhao Shen · Juexiao Zhou

---

The Problem

AI agents are reactive. They wait for instructions, execute tasks, and stop. They never wonder "what else should I know?" or "what changed since yesterday?"

CurioClaw gives agents an intrinsic drive to explore. It's an independent curiosity engine that makes agents proactively discover knowledge gaps, search the web, and report findings — all without being asked.

How It Works

CurioClaw runs a 7-step curiosity cycle:

   ┌──────────┐
   │ PERCEIVE │ ← Extract curiosity signals from recent conversations
   └────┬─────┘
        ▼
   ┌──────────┐
   │  SCORE   │ ← Rate novelty, relevance, learnability (hard + soft scoring)
   └────┬─────┘
        ▼
   ┌──────────┐
   │  GOVERN  │ ← Check governance level (autonomous / guided / supervised)
   └────┬─────┘
        ▼
   ┌──────────┐
   │ EXPLORE  │ ← Investigate top signals via web search
   └────┬─────┘
        ▼
   ┌──────────┐
   │  RECORD  │ ← Save findings to JSONL memory
   └────┬─────┘
        ▼
   ┌──────────┐
   │ COMPRESS │ ← Periodically compress old memories
   └────┬─────┘
        ▼
   ┌──────────┐
   │  NOTIFY  │ ← Proactively tell the user if something scored high
   └──────────┘

Key design principle: The Agent (LLM) handles all reasoning — reflecting, judging, searching, summarizing. The Python code (curiocore/) handles only deterministic computation — scoring math, deduplication, JSONL I/O, compression. No LLM API calls in the Python code.

Three trigger modes:

• Heartbeat — automatic, runs on a timer when the user is idle
• Post-conversation — after a substantive dialogue ends
• Manual — user says "curiosity check", "explore", or "what's new"

Quick Start (OpenClaw)

# 1. Install the curiocore package
pip install curiocore

# 2. Clone the repo for adapter files
git clone https://github.com/ziwenwang28/CurioClaw.git
cd CurioClaw

# 3. Copy skill files to your OpenClaw workspace
mkdir -p ~/.openclaw/workspace/skills/curiocore
cp adapters/openclaw/SKILL.md ~/.openclaw/workspace/skills/curiocore/SKILL.md

# 4. Add heartbeat config
cat adapters/openclaw/HEARTBEAT.md >> ~/.openclaw/workspace/HEARTBEAT.md

# 5. Configure heartbeat interval (30 minutes recommended)
openclaw config set agents.defaults.heartbeat.every "30m"

# 6. Restart and test
openclaw gateway restart

Then send your agent: "curiosity check" — it should run the curiosity cycle and reply with a 💡 message summarizing its discoveries.

For detailed setup instructions, see adapters/openclaw/install.md.

Features

• Zero LLM calls in Python — all intelligence lives in the Agent following SKILL.md
• Dual scoring — hard metrics (cosine similarity, Jaccard, heuristics) blended 50/50 with the Agent's soft judgment
• JSONL memory — simple, inspectable, streamable storage
• Rolling compression — memory stays lean over time (auto-compresses at 500 entries)
• Three-tier governance — from full autonomy to human-supervised exploration
• Smart notifications — only notifies the user when discoveries score above threshold (composite > 0.7)
• Idle-aware — won't interrupt active conversations; waits for the user to go idle

Scoring System

Each curiosity signal is scored on three dimensions:

Dimension	Weight	What it measures
Novelty	40%	How new is this compared to past explorations?
Relevance	35%	How aligned with the user's interests?
Learnability	25%	Can we learn something actionable by searching?

Each dimension gets two scores — a hard score (computed by Python using cosine similarity, keyword matching, and structural heuristics) and a soft score (the Agent's own judgment). These are blended 50/50 into a composite score.

The Agent only notifies the user when composite > 0.7.

Governance Model

Level	Who decides?	Use case
autonomous	Agent explores freely	Background research bots
guided	Human sets directions, agent picks topics	Personal learning assistant
supervised	Agent proposes, human approves	Production systems

Architecture

curiocore/                ← Pure Python computation (no LLM calls)
├── scorer.py             ← Hard metric computation (cosine sim, Jaccard, learnability)
├── signals.py            ← Signal deduplication logic
├── memory.py             ← JSONL read/write + compression
├── config.py             ← Configuration dataclasses
├── types.py              ← Data structures (Signal, CuriosityScore, etc.)
└── run_curiosity.py      ← CLI tool (score, store, compress, status, dedup)

adapters/                 ← Integration layers for agent frameworks
├── openclaw/             ← OpenClaw adapter
│   ├── SKILL.md          ← Agent instructions (the curiosity cycle)
│   ├── HEARTBEAT.md      ← Heartbeat trigger logic
│   ├── install.md        ← Installation guide
│   └── README.md         ← Adapter overview
└── _template/            ← Template for new adapters

site/                     ← Landing page
└── index.html

Adapters

CurioClaw is framework-agnostic. Adapters connect it to specific agent frameworks:

Adapter	Framework	Status
openclaw/	OpenClaw	✅ Working
_template/	Your framework here	Template

Want to add an adapter? See adapters/README.md or open an adapter request.

Configuration Reference

Setting	Where	Default	Description
Heartbeat interval	openclaw config	30m	How often the agent wakes up
Cycle cooldown	HEARTBEAT.md	10 min	Minimum time between curiosity cycles
User idle check	HEARTBEAT.md	10 min	Don't run if user was active recently
Notification threshold	SKILL.md Step 7	composite > 0.7	Minimum score to send 💡 notification
Compression threshold	SKILL.md Step 6	500 entries	When to compress old memories
Max explorations/cycle	SKILL.md Step 2d	3	Max signals to explore per cycle

Status

🚧 Early Prototype (v0.1.0-alpha)

This is the very first public release of CurioClaw — an early prototype to share our vision. Everything you see here represents our initial design and architecture; every component is in its earliest form.

We will ship significant updates within the next week, including improved scoring, better notification logic, more robust heartbeat integration, and documentation.

v1.0.0 target: March 14, 2026.

We welcome early feedback, ideas, and contributions — but please expect breaking changes before v1.0.0. Star the repo to follow our progress.

Star History

If CurioClaw is useful, consider giving it a ⭐ — it helps others find the project.

License

MIT — use it anywhere, for anything.