# Remembery

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Ibrahim Memon](https://cerebralvalley.ai/u/IAMibrahim)
- **GitHub:** https://github.com/IAMIbrahimmemon/remembery-model
- **Demo video:** https://youtu.be/P_QlLeQzux4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1569

Meta needs a way for customers to feel that their data is safe when wearing Meta's VR Glasses, while still receiving a personalized AI experience. Sending that stream to a cloud LLM all day is slow, costly and a privacy problem.

So, I built a tiny AI model meant to store memory from a person's life. It's a non-autoregressive 360M model. This allows memories to stay on device and only send relevant memories to the cloud. 

Remembery is a new AI model that turns conversation into structured memories entirely on device. A fine-tuned Meta MobileLLM-R1 (360M) with a bidirectional adapter and non-autoregressive heads reads the live transcript and, in a single forward pass, labels every entity, event frame and role. A calibrated validator keeps the lines worth remembering as searchable records. 

On my held out dataset it scores an 80% on recall while a cloud LLM that is 100x the size only scores 81%. LLM makes a wrong memory on 5 in 500, while Remembery only does on 1 in 500.

This is a massive and successful development and project because of how much it does with how little. A model small enough to run in airplane mode on a phone matches a cloud LLM 100× its size at remembering what matters, and makes five times fewer wrong memories. It needs no token generation and no server, and nothing about your day leaves the device unless you ask. That turns always-on memory for Meta’s glasses from a privacy risk into something people can trust. I built it solo in 9 days on Meta’s open MobileLLM-R1 and ExecuTorch.

## More from Meta Global AI Developer Hackathon

- [Dev-Senpai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1566)
- [Project CAD](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1567)
- [Flamers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1568)
- [Confam](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1570)
- [ReARC Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1571)
- [TailorApps](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1572)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1569. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt