An AI research assistant that actually does the homework

What it does

ScienceClaw is a research-specialized fork of the OpenClaw agent engine. It connects to 25+ academic databases (PubMed, UniProt, arXiv, World Bank, and others), runs literature searches, performs analyses, and generates structured outputs. The system enforces minimum research depth—quick tasks require 5 tool calls, systematic reviews require 100+—and won’t stop until it has cross-referenced findings across at least three sources.

The interesting bit

The project treats hallucination as a protocol-level bug, not a user-experience annoyance. Its 629-line SCIENCE.md forbids citing anything not returned by a tool in the current conversation; unsure claims must be explicitly labeled “not verified.” Meanwhile, the agent writes new SKILL.md files at runtime to capture your preferences—search templates, statistical methods, citation formatting—without redeployment.

Key highlights

• Self-evolving skills: Starts with 285 skills, grows via runtime skill generation based on your research patterns
• Persistent memory: Four-layer system with LanceDB vector storage, temporal decay weighting, and cross-session pattern retrieval
• Long-session architecture: 1-hour timeout with heartbeat monitoring, smart context compaction that preserves statistics and citations
• Discipline coverage: Biomedicine, chemistry, genomics, physics, economics, psychology, linguistics, and others
• Anti-hallucination enforcement: Hard protocol rules, not suggestions; every claim needs tool-derived evidence

Caveats

• The “zero hallucination” claim is aspirational protocol, not empirical guarantee—no benchmark data or third-party validation is provided
• Self-evolution mechanics are described narratively (“Week 1… Month 3”) without technical detail on how skill generation actually works
• Built on OpenClaw, so much of the core agent infrastructure is inherited rather than original

Verdict

Worth exploring for researchers drowning in literature reviews who need structured, source-backed output and don’t mind babysitting an agent for an hour. Skip it if you want a drop-in chatbot; this is deliberately slow and pedantic by design.