About AROS
Train the procedure, not the weights.
AROS Cloud Federation turns the tacit expertise of working research labs into skills an agent can recall, correct, and reuse — without ever moving the raw data.
Agents demo well. In a real lab, they stall.
As of 2025, only about 2% of organizations had deployed AI agents at scale (third-party estimate). The gap between a convincing demo and daily lab work is not model quality. It is three failures that compound:
- Tacit expertise stays tacit. The judgment a senior researcher applies — which control to run, when a result looks suspicious, how to salvage a failed prep — rarely gets written down, so an agent never sees it.
- Session amnesia. Each session starts cold. What was learned yesterday is gone today.
- Unauditable output. When an agent hands you an answer, you cannot see which memories, skills, or context produced it. You cannot check the work, so you cannot trust it.
Fix those three, and an agent stops being a demo.
Two gaps, and one bet.
The embodiment gap. Digital systems are deterministic, discrete, and closed. Biology is analog, continuous, and open. In a wet lab, noise is not an error to be scrubbed out — it is the raw material for adaptation and intuition. A model trained only on clean digital traces has never met that world.
The determinism gap. A language model cannot self-generate the exploration entropy needed to escape a local optimum. Humans supply it. Left to train on its own synthetic output, a model drifts toward collapse. Authentic human procedure is the scarce input.
The bet: train the procedure, not the weights. We do not fine-tune weights on lab data. We improve the context an agent carries — skills, policies, workflows, knowledge items, personas. We transfer the abstraction, not the low-level trace, because raw traces cause negative transfer. Lessons are written back into the environment and pre-loaded on the next run by an ambient screener.
Three parts. One memory.
AROS Core
capture, distil, recall
Core runs local-first. It captures your work, stores it in a local memory database with hybrid vector and full-text retrieval, and recalls the relevant pieces on your next task. Capture, storage, and retrieval never leave the machine. Distillation needs a model — a Gemini key or the metered gateway by default, or a fully local model under Pure-Local Mode. Only in Pure-Local Mode does Core run with no cloud account.
AROS IDE
inspect and correct
The IDE is an Eclipse Theia fork that hard-depends on Core. It is the human audit surface: open any node and see the context it was given, the memories it retrieved, the skills it mounted, and its inputs and outputs. When Core is absent, it fails loudly rather than pretending. This is where you catch a wrong turn and correct it.
AROS Cloud Federation
share the abstraction, keep the data
Federation lets labs share distilled skills without sharing raw data. Export applies an allow-list — documentation, config, and source under a size cap — and excludes everything else: CSV, parquet, FASTQ, DICOM, images, archives, notebooks, pickles. On the receiving side, a single cosine threshold decides each item: near-duplicate, skip; otherwise, insert. Raw experimental data is excluded from federation by code.
Why university labs, specifically.
- Teaching forces decomposition. To teach, an expert must break tacit skill into explicit, step-by-step language — exactly the shape a reusable skill needs.
- Labs keep producing novel ideas. Genuinely new, AI-free hypotheses keep coming out of research groups. That is the scarce input.
- PhD operators live in pipelines. They work with data pipelines natively, so human-in-the-loop data is the exhaust of ordinary research — no separate labelling step.
- Incentives align on truth. Peer review, reproducibility, and grants reward correctness. That beats pay-per-task labelling on signal-to-noise.
What we have measured — and what we have not.
Two registers below: telemetry from one live deployment, and a controlled benchmark. Both are single-instance results, not a fleet average.
Deployment telemetry
One reference instance, brain database read 2026-07-29 — a deployment of 9 PCs and 6 people (3 senior researchers, 3 students) over roughly 2 months.
| Memory layer | Count |
|---|---|
| Active world facts | 777 |
| Episodic experiences | 198 |
| Mental models | 5,150 |
| Insights | 391 |
| Skills in hierarchy | 1,630 |
| Memory-graph edges | 3,021 |
| Dream-audit entries | 10,632 |
| Agent personas | 34 |
Over that period the deployment consumed more than 8.6 billion tokens. The honest limitation: roughly 70% of those tokens went into building the system itself, not doing science. This is one instance, not a federated union.
Benchmark — Wave A
Completed 2026-07-26: two chemistry benchmarks × three arms (baseline, MCP-only, and the intra-machine AROS Swarm) × 300 items each = 1,800 evaluations. Pure-local, about 4.5 days, equivalence-aware scoring. Non-AROS arms recorded a memory delta of zero — no learning leaked into the baselines.
| Benchmark | Baseline | MCP-only | AROS Swarm |
|---|---|---|---|
| chembench — single-step MCQ recall | .497 | .443 | .427 ▼ |
| chemreason — multi-step reasoning | .510 | .550 | .573 ▲ |
Stated plainly: orchestration hurts simple MCQ recall and wins on multi-step reasoning. We report both, including the one that goes against us.
Where we are.
- ShippingAROS Core
Capture, local memory, and retrieval run today. Memory consolidation runs now: a “dreamer” summarises finished sessions into world facts, mental models, and insights, and a hygiene pass deduplicates and merges continuously. Skill-content rewriting (SkillOpt/GEPA) is built but has not completed a run on the reference instance — treat it as unproven.
- In development· ~50% builtAROS IDE
The inspect-and-correct surface works against Core. It is not finished.
- Early — waitlist· ~5% builtAROS Commander
The multi-machine orchestrator is spec-complete but only lightly implemented, with networking stubbed. There is no host-fleet product today.
Pricing, where it is set: AROS Core $20/mo, AROS IDE $200/mo; marketplace content at $20 per tool, $5 per skill, $10 per pipeline, with a 70% creator cash share. Online Apps stay metered per run. AROS Commander is not priced before launch, and checkout is not yet open.
Business model — modelled, not revenue we have today
Roughly $300K ARR at 36 months, built bottom-up from about 20 labs at month 12, 50 at month 24, and 100 at month 36. That is a projection. Third-party market context: bioinformatics ~$18.7B by 2030, AI drug discovery ~$8.6B by 2028, agentic context ~$7–12B by 2034 — context, not our numbers.
Bring your procedure. Keep your data.
If you run a lab, you already produce the scarce input: authentic human procedure, generated as the exhaust of real research. AROS Core captures it, the AROS IDE lets you inspect and correct it, and Federation lets you share the abstraction while your raw data never leaves the building. Inspect it. Read the receipts. Then decide whether to stand behind it.