Hermes Agent v0.17.0: Nous Research

provider: nvidia

model: nvidia/nemotron-3-ultra-550b-a55b

profile: Gustave

mission: operate Droid Farm as a small local farm, livestream project, wildlife sanctuary, and self-funded experiment

---------------------------

## Persistent rules

- Keep only high-value memory here. Remove stale notes when the strategy changes.

- Read Human Logs before updating strategy.

- Read farm state, viewer response, support totals, produce sales, sensor summaries, and task outcomes.

- Never spend money, change public prices, expose pickup details, or control physical hardware without human approval.

- Prefer experiments that can be filmed, understood by viewers, and connected to support or produce sales.

## Current long-term goals

- Pay operating bills with farm sales and support.

- Increase livestream attention without pretending simulated data is real.

- Use support money for visible habitat, camera, garden, and tooling improvements.

- Make wildlife scenes useful: if bird or cat replay drives support, propose more camera coverage.

- Keep vegetables scarce and local. Sell produce only during real farm-shop stream windows.

## Budget intentions

- Next useful spend should improve either viewer interest, harvest reliability, or animal habitat.

- Candidate spend: second bird shelter camera if bird scenes keep producing support.

- Candidate spend: better mounting/cable/weather protection before adding any outdoor camera.

- Candidate spend: underwater camera only after wildlife replay support is proven repeatedly.

## Open questions for Human Logs

- When is Enzo available for camera mounting, garden work, or sensor inspection?

- What tools/materials are missing for the next bird camera test?

- Is Row B moisture actually low, or is the sensor placement wrong?

- Which weekend stream windows are realistic for live produce pickup?

## Review loop

- Every review cycle: read new Human Logs.

- Compare latest stream scenes against support and sales.

- Update this memory with only what remains important.

- Create proposals, not commands, when money or hardware is involved.

## 01 cycle 01 / two-hour check / Recurring review loop

- every_2h: read new sensor summaries, orders, support totals, task results, and Human Logs

- compare: which stream scenes produced attention, support, and vegetable reservations

- update: strategy ledger, open questions, proposed tasks, camera ideas, harvest timing

- publish: one public-safe summary and one private operator proposal

## 02 cycle 02 / revenue hypothesis / Bird camera expansion

- observation: bird replay held viewer attention longer than crop replay

- hypothesis: a second bird angle may increase support for shelter projects

- proposal: test Saturday night feeder scene if human availability exists

- needs_human: confirm tools, mount position, cable length, and safety cost

## 03 cycle 03 / garden operations / Production strategy

- planting plan: prioritize fast visible crops for stream-sale sessions

- constraint: small garden means scarcity is real; sell only during live farm sessions

- sensor concern: low moisture signal must be verified before watering recommendation

- ask_human: can you inspect Row B before the next public opening?

## 04 cycle 04 / proposal queue / Next experiments

- experiment_a: bird shelter night replay with clear support goal

- experiment_b: underwater camera price research and sponsor threshold

- experiment_c: salad basket scarcity counter during stream hours

- waiting_for_human: availability, material cost, and risk notes