Scienceclaw
🔬🦞 A self-evolving AI research colleague for scientists. 285 skills, zero hallucination, persistent memory.
Last verified:
What is Scienceclaw?
ScienceClaw is a self-evolving AI research platform designed specifically for researchers across all scientific disciplines. It functions as an AI colleague that searches literature, queries academic databases, and runs analyses — improving the more you use it. The platform features 285 research skills that grow and adapt through runtime self-evolution, persistent research memory with four-layer storage using LanceDB vector storage, and support for 1-hour research sessions with heartbeat monitoring.
Key features include zero hallucination protocol (every citation must come from verified tool results with no fabricated DOIs or invented authors), access to 25+ academic databases including PubMed, Semantic Scholar, OpenAlex, arXiv, UniProt, KEGG, PDB, ChEMBL, PubChem, NCBI Entrez, Materials Project, and World Bank, and mandatory depth thresholds that ensure real research without premature conclusions. The platform covers 20+ disciplines spanning natural sciences, social sciences, and humanities from biomedicine to economics and genomics to political science.
ScienceClaw is built for academic researchers, PhD students, postdocs, research scientists, and faculty members who need reliable literature reviews, database queries, and data analysis. Unlike general-purpose AI assistants, it maintains persistent memory across sessions so you can continue literature reviews where you left off. The skill evolution cycle executes research tasks, reflects on outcomes, evolves new skills based on results, and persists improvements for future sessions.
Scienceclaw pricing
Pricing model: Freemium
Open source and free to use. Your only costs are hosting and AI API usage. Free option: Run locally on Mac with Ollama for completely free AI (electricity ~$5/month). Cheap VPS option: Hetzner VPS €4/month + GPT-4o mini ~$5/month. Good balance: VPS $6/month + Claude Sonnet ~$15-20/month. AI model pricing per 1M tokens: Claude Sonnet 4.5 $3 input/$15 output, Claude Haiku 4.5 $0.80/$4.00, GPT-4o $2.50/$10.00, GPT-4o mini $0.15/$0.60, Ollama Llama 3.3/Mistral free forever. Average personal usage: 5-20M tokens/month (~$5-20 for Claude Sonnet). VPS options: Hetzner €3.79/mo (2 vCPU, 2GB RAM), Contabo €4.99/mo (4 vCPU, 8GB RAM), Vultr $6/mo, DigitalOcean $6/mo, AWS Lightsail $5/mo.
Scienceclaw pros
- 285 self-evolving research skills that improve over time
- Zero hallucination protocol with verified citations only
- Four-layer persistent research memory with LanceDB vector storage
- 1-hour research sessions with heartbeat monitoring
- Access to 25+ academic databases across natural and social sciences
- Covers 20+ disciplines from biomedicine to political science
- Mandatory depth thresholds prevent shallow answers
- Runtime skill creation optimized for your specific field
- Open source and free to use (only pay for hosting/API)
- Research reflection engine continuously learns from workflows
- No fabricated DOIs or invented authors
- Continue literature reviews across multiple sessions
- Structured API queries for 25+ connected databases
- Gets smarter the more you use it
- Set up from clone to research in under 5 minutes
Scienceclaw cons
- Requires technical setup via git clone and command line
- Only costs are hosting and AI API usage (not truly free)
- Runs inside Docker containers with no host system access
- Cannot access personal files or environment variables
- Requires 16GB+ RAM for local Ollama operation
- Monthly costs can reach $5-20 for typical personal usage
- No built-in GUI (uses npx openclaw gateway command)
- Self-evolution requires multiple sessions to show benefits
Frequently asked questions about Scienceclaw
What is ScienceClaw?
ScienceClaw is a self-evolving AI research platform built specifically for researchers. It acts as an AI colleague that searches literature, queries databases, and runs analyses — getting better at research the more you use it. Unlike general-purpose AI assistants, it features 285 research skills, zero hallucination protocol, persistent research memory across sessions, and access to 25+ academic databases across 20+ disciplines.
How does the self-evolving skills system work?
The skill library evolves through a continuous four-step cycle: Execute (run research tasks using existing skills), Reflect (evaluate what worked and what didn't), Evolve (create or improve skills based on outcomes), and Persist (store improvements for future sessions). Week 1 you get general tools, but by Month 3 you have a trained research assistant tuned specifically to your field as the system creates new specialized skills at runtime optimized for your research area.
What databases does ScienceClaw connect to?
ScienceClaw connects to 25+ academic databases with structured API queries including PubMed, Semantic Scholar, OpenAlex, arXiv, UniProt, KEGG, PDB, ChEMBL, PubChem, NCBI Entrez, Materials Project, and World Bank. These cover both natural sciences (biomedicine, genomics, materials science) and social sciences (economics, political science) across 20+ disciplines.
How does zero hallucination work?
Every citation must come from an actual tool result. ScienceClaw never fabricates DOIs, never invents authors, and never makes up references. If the system cannot verify information, it explicitly states that it cannot verify rather than making something up. This ensures all research outputs have verified, traceable citations from real database queries.
How long can research sessions last?
ScienceClaw supports 1-hour research sessions with heartbeat monitoring, compared to standard AI's typical 10-minute timeout. Mandatory depth thresholds ensure real research is conducted without premature conclusions or shallow answers. This allows for extended literature reviews, complex multi-step analyses, and thorough database queries that require sustained attention.
How do I get started with ScienceClaw?
Get started in 3 steps taking under 5 minutes: 1) Clone the repository with git clone https://github.com/beita6969/ScienceClaw.git, 2) Setup by running chmod +x setup.sh && ./setup.sh, 3) Start researching with npx openclaw gateway. The platform is open source and runs on your own infrastructure.
What is the research memory system?
ScienceClaw features a four-layer persistent memory system with temporal decay and LanceDB vector storage. This allows you to continue literature reviews across sessions — the system remembers where you left off. Unlike standard AI that forgets after each session, your research context, previous findings, and workflow patterns persist and inform future research tasks.
Which disciplines does ScienceClaw support?
ScienceClaw covers 20+ disciplines across natural sciences, social sciences, and humanities. Natural sciences include biomedicine, genomics, materials science, chemistry, and biology. Social sciences and humanities include economics, political science, and related fields. The platform is designed for cross-disciplinary research with database access spanning both domains.
Is ScienceClaw free to use?
ScienceClaw is open source and free to download. Your only costs are hosting infrastructure and AI API usage. You can run it completely free locally using Ollama on your Mac (only electricity cost ~$5/month), or use paid options like VPS hosting (€3.79-$6/month) plus AI API costs ($5-20/month for typical usage with Claude Sonnet or GPT-4o).
How is ScienceClaw different from standard AI assistants?
ScienceClaw is built specifically for researchers while standard AI is built for everyone. Key differences: 1-hour+ sessions vs 10-minute timeout, 285+ self-evolving skills vs ~54 fixed skills, 4-layer persistent memory vs forgetting after session, zero hallucination protocol vs possible hallucinations, 25+ academic databases vs none built-in, mandatory depth thresholds vs single-pass answers, and runtime skill evolution vs static capabilities.