🐾

CurioClaw

Curiosity never rests.

The missing intrinsic drive for AI agents.
Explore. Learn. Evolve — autonomously.

View on GitHub How it works

Your agent only thinks when spoken to.

Today's AI agents are reactive. They wait for instructions, execute tasks, and go dormant. They never wonder. They never explore on their own. They never come back and say "I found something you need to know."

🪫

Zero Curiosity

Agents complete tasks then stop. No reflection, no follow-up, no initiative. Knowledge gaps go unnoticed.

🔇

Silent Between Tasks

Between conversations, your agent learns nothing. Idle time is wasted time. No background exploration happens.

🔁

Same Mistakes Repeated

Without persistent memory of what it has explored, the agent can't build cumulative knowledge over time.

A 5-step curiosity loop, running in the background.

CurioClaw gives your agent an autonomous curiosity engine. It perceives knowledge gaps, scores them, explores via web search, and records findings — all without being asked.

01

Perceive

After conversations, the agent reflects and detects curiosity signals — questions left unanswered, surprising claims, knowledge gaps.

02

Score

Each signal is scored on novelty, relevance, and learnability using a hybrid of deterministic metrics and LLM judgment.

03

Explore

The top-scoring signals are explored via web search. The agent generates focused queries and synthesizes findings.

04

Record

Findings are saved to persistent JSONL memory. The agent builds cumulative knowledge across sessions.

05

Notify

If a finding is highly relevant to the user's recent work, the agent proactively reaches out — no prompting required.

Your agent finds things before you ask.

Here's what it looks like when the curiosity engine discovers something relevant during a background cycle.

feishu — ai-assistant
👤
You · 2 hours ago
Can you look into how OpenAI's Codex OAuth integration works? We need it for our Slack Agent.
🐾
Agent · just now
💡 I just explored the AI Agent OAuth authentication space in the background and found something you might find interesting:

OpenAI updated their Codex OAuth scope documentation last week, adding a new tools.exec permission scope that allows Agents to request code execution privileges during the OAuth flow. This is directly relevant to the Slack integration scenario you mentioned earlier.

Want me to go into more detail?

Proactive, not reactive.

How CurioClaw compares to existing agent capability evolution approaches.

Capability Standard Agent Skills CurioClaw
Trigger mode User must ask Auto heartbeat + manual
Knowledge gaps Ignored Detected & explored
Background learning None Continuous via heartbeat
Proactive notification Never When relevance > 0.7
Persistent memory Session only JSONL with compression
Scoring method Hybrid hard metrics + LLM
Governance 3 tiers: autonomous / guided / supervised
LLM dependency in code Embedded API calls Zero — Agent handles LLM natively

Give your agent a mind that wanders.

CurioClaw is open source, MIT licensed, and designed to work with OpenClaw out of the box. Clone it, install it, and watch your agent start exploring on its own.