# Moss

- **Event:** [The Harness Engineering & Model Wrangling Hackathon](https://cerebralvalley.ai/e/mongodb-nyc-hackathon)
- **When:** Sat, Sep 26 at 9:00 AM – 10:00 PM (EDT)
- **Where:** New York, NY
- **Team:** [Shaishav Pidadi](https://cerebralvalley.ai/u/shaivpidadi), [Khushi Parmar](https://cerebralvalley.ai/u/Khushiparmar), [Vishnu Daka](https://cerebralvalley.ai/u/dakavishnu), [Gayatri Biram](https://cerebralvalley.ai/u/Gayatrib)
- **GitHub:** https://github.com/khushi491/moss
- **Demo video:** https://www.veed.io/view/fb3b666d-35ef-4ddd-a0bb-cebf8381a318?source=Dashboard&panel=share
- **Gallery:** https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/22

Moss: long-running AI teammates whose runs and memory live in one MongoDB Atlas cluster, and that forget on evidence.

You message a coordinator (HQ). It briefs named Bots, which work in their own browser and shell, sleep for months, survive a kill -9, and come back only when they need your approval. 

Everything durable sits in Atlas behind one connection string:
- Runs: a Workflow SDK "World" built on MongoDB. The event log, steps, waits, queue and streams are collections, and a job due next year is one document.
- Memory: each memory is one document, found by meaning and by words at once. Atlas Automated Embeddings (Voyage) and Atlas Search are fused with $rankFusion.
- Lessons that are measured: a lesson a Bot learns is briefed into later jobs, checked by a decision model, and retired with the evidence when a job proves it wrong. A retired lesson never comes back, even while the search indexes catch up.
- A hard metric: npm run memory:eval scores recall (hit@1, hit@3, MRR) for words, vector and hybrid search.

Built at the event (26 Sep): hybrid memory search on the Atlas Sandbox (Automated Embeddings, Atlas Search, $rankFusion, rate-limit fallback, retirement that holds while indexes lag), the recall evaluation, MongoDB's MCP server as a connector for every agent, and latency tooling that traced step time to network distance rather than the database. The durable runtime, memory and guardrails were built before the event.

Moss as Product: 
https://drive.google.com/file/d/1CvTMY9OJpboq3wp7ipZ2FfRq0ApDUztE/view?usp=sharing

## More from The Harness Engineering & Model Wrangling Hackathon

- [MongoHive](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/19)
- [Ada](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/20)
- [PolicyPilot](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/21)
- [Second_Loop_Engineering](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/23)
- [devrandom.co](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/24)
- [Antibody](https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/25)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-nyc-hackathon/hackathon/gallery/22. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt