ScienceClaw: A Research Agent Built on OpenClaw With a Hard Citation Rule
🔬🦞 A self-evolving AI research colleague for scientists. 285 skills, zero hallucination, persistent memory.
At a glance
- What is it?
- ScienceClaw is a TypeScript research agent that layers 285 skills, a persistent memory stack and a 629-line protocol document on top of the OpenClaw engine. Its central claim is that no citation may enter an answer unless a tool returned it in the current session.
- Who is it for?
- Adopt ScienceClaw if your work is long-running literature retrieval where a fabricated PMID would be worse than a slow answer, and if you are willing to run the OpenClaw engine yourself. Do not adopt it if you need a hosted product with a support contract, or if your task is a single lookup that finishes in two minutes.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 99 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Citation Fabrication Problem ScienceClaw Targets
The project's stated problem is narrow and concrete. General assistants invent DOIs, invent author names, and cite papers that do not exist. In a manuscript draft or a grant application, a fabricated PMID is not a cosmetic error, it is a retraction risk. ScienceClaw's answer is a rule placed above every other instruction in the system: every citation must come from a tool result in the current conversation. If a database did not return it, it cannot be cited. If the agent is unsure, the protocol tells it to write "not verified" rather than guess. The target user is a researcher running literature searches, database queries and meta-analytic work who would rather receive an incomplete answer than a confident wrong one. That is a real audience, and it is not the same audience as someone who wants a chat assistant to summarize a single paper. The README frames the contrast as general-purpose versus built for researchers, and the citation rule is the sharpest expression of that distinction.
How the Skill Evolution Loop Actually Works
ScienceClaw ships with 285 skills and, according to the README, grows from there. The mechanism is file-level: the agent writes new SKILL.md files at runtime without redeployment. That is the whole trick. Skills are not compiled into the binary or registered through an API call, they are markdown documents the agent authors and then reads back on later runs. The README's worked example is an immunology researcher. In week one the system records that PubMed plus Semantic Scholar works best for that user's queries, that forest plots are preferred over tables, and that citations always need both PMID and DOI. By week four it has written subfield-specific search templates and a database priority chain. By month three the README claims it handles the domain like a trained assistant. The comparison offered is against stock OpenClaw, which the README describes as shipping roughly 54 general-purpose skills that do not change. Two caveats belong here. First, the timeline is illustrative prose, not a measured result, and the repository provides no evaluation of how quickly skill quality actually converges. Second, runtime file authoring means skill quality is only as good as the agent's own judgement about what worked, and nothing in the supplied material describes a review step before a new SKILL.md takes effect.
Four-Layer Memory and What Survives a Full Context Window
The memory design has two parts that matter. The first is cross-session retrieval: the README gives the example of asking the agent to continue a literature review started the previous Tuesday, or to reuse the search strategy from a BRCA2 project, and states that findings from one project can inform another. The second is compaction behaviour. When the context window fills, ScienceClaw is described as preserving statistical results, effect sizes and key citations while compacting intermediate steps. That ordering is the interesting design decision. Most summarization compresses uniformly and loses exactly the numbers a researcher needs to reproduce a result. Prioritizing effect sizes over the reasoning that produced them is a defensible trade for a literature tool and a poor one for a debugging tool. The README also states that ScienceClaw adds temporal decay weighting and LanceDB vector storage on top of what it describes as OpenClaw's basic memory plugin. Temporal decay means older research patterns count for less, which is sensible for a field that moves, and questionable if you are returning to a project after a long gap. Nothing in the supplied material explains how the decay rate is set or whether a user can override it.
Session Timeouts, Heartbeats and the Depth Checklist
The README publishes a comparison table that is unusually specific. Standard OpenClaw is listed with a 600 second agent timeout, described as ending when the conversation ends, single-pass in research depth, and using basic truncation for context. ScienceClaw is listed at 3600 seconds, with a heartbeat that keeps sessions alive across interruptions, a multi-phase protocol with depth thresholds, and smart compaction. The thresholds are given as tool-call minimums: Quick is 5, Survey is 30, Review is 60, Systematic is 100 or more. Before concluding, the agent must satisfy a checklist: at least three different databases or sources searched, full metadata retrieved rather than titles alone, findings cross-referenced, contradictory evidence checked, key statistics verified against primary sources, and results written to a structured output file. If any item is unchecked, the README states the agent keeps working instead of returning a partial answer. This is the most falsifiable claim in the repository, because the threshold numbers and the checklist items are concrete enough to test against a real task. It is also where the cost sits: a Systematic run at 100-plus tool calls against live literature APIs is slow and rate-limit sensitive, and the README does not discuss what happens when a database throttles the agent mid-run.
Getting It Running: What the README Does and Does Not Give You
This is the weakest part of the supplied material, and it should be said plainly. The README describes the architecture, the protocol file and the comparison table, but it does not include an install command, a configuration file, or a set of environment variables. What can be confirmed is the shape of the thing: it is a TypeScript project, it is built on the OpenClaw engine linked from the README, and its behaviour is governed by a file named SCIENCE.md at the repository root, described as 629 lines long. Skills live as SKILL.md files that the agent writes at runtime. Memory uses LanceDB for vector storage. A homepage is listed at scienceclaw.science and the repository carries the mcp topic, which suggests a Model Context Protocol integration, though the README excerpt does not show the server configuration. An engineer evaluating this should treat the absence of a quickstart as the first thing to resolve: clone the repository, read SCIENCE.md in full, and inspect the skill directory layout before assuming the 285 figure maps to connectors you actually need. The discipline and database list is collapsed behind a details block in the README and is not reproduced here, so coverage claims cannot be checked from this material alone.
Where the Design Breaks Down
Three failure modes follow from the architecture as described. The first is the zero-hallucination rule itself. A hard constraint that every claim needs a tool result in the current conversation will, in practice, produce "not verified" where a general assistant would offer a plausible synthesis. For exploratory work, where you want a hypothesis to chase, that is a worse tool. The second is the one-hour session with heartbeat. Long sessions against PubMed and Semantic Scholar are exposed to rate limits, transient failures and API changes, and the README does not describe retry or backoff behaviour. The third is runtime skill authoring. Skills the agent writes for itself are unreviewed artifacts; if a search template is subtly wrong, it persists and gets reused. Nothing in the supplied material mentions a validation gate, a diff review, or a way to roll back a bad SKILL.md. The self-evolving loop is the headline feature and the least governed part of the system.
ScienceClaw Versus Stock OpenClaw
The honest alternative is the engine underneath. The README positions ScienceClaw as built on OpenClaw but redesigned for academic research, and the differences it names are specific: roughly 54 fixed general-purpose skills versus 285 that grow; a 600 second timeout versus 3600; a basic memory plugin versus temporal decay plus LanceDB plus cross-session pattern retrieval; no hallucination controls versus the SCIENCE.md protocol. The approaches differ in kind, not degree. OpenClaw is a general agent runtime where you supply the discipline. ScienceClaw bakes the discipline into a protocol file that applies before any other instruction and enforces tool-call minimums before an answer is allowed out. If your work is sending messages, scheduling and short lookups, the 600 second default is sufficient and the extra protocol is overhead you will pay for on every run. If your work is a systematic review where the citation trail has to survive scrutiny, the protocol is the point. The choice is whether you want the constraints enforced by the runtime or by your own prompting.
Licence, Maintenance and What to Verify First
The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is permissive and low friction, and it is worth noting that MIT offers no patent grant, which matters more in biotech than in most software contexts. This is a description of the licence text, not legal advice; if you are embedding the agent in a product, have counsel read it. On maintenance, the observable facts are limited: the default branch is main, the repository is not archived, and the last push recorded is 2026-06-08. No releases were retrieved, so there is no versioned artifact to pin against, which means tracking main is the only upgrade path visible from here. For a tool that writes its own skill files at runtime, that is a real cost: an upstream change to SCIENCE.md or to the skill loader can alter agent behaviour without a version number to point at. Verify three things before adopting. Read SCIENCE.md in full and decide whether you accept its constraints. Confirm that the specific databases your field depends on are represented in the skill set rather than inferred from the discipline list. And locate the directory where runtime-written SKILL.md files are stored, so you can put it under version control and see what the agent has taught itself.
Editorial conclusion
Adopt ScienceClaw if your work is long-running literature retrieval where a fabricated PMID would be worse than a slow answer, and if you are willing to run the OpenClaw engine yourself. Do not adopt it if you need a hosted product with a support contract, or if your task is a single lookup that finishes in two minutes. Before committing, read SCIENCE.md end to end, confirm that the database connectors you need are actually present in the skill set rather than assumed from the discipline list, and check whether the runtime-written SKILL.md files land in a directory you control and can version.
Community notes