# Nevermind

- **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:** [Harsh Mishra](https://cerebralvalley.ai/u/Harshvijay)
- **GitHub:** https://github.com/HarshM0210/ChronoFact
- **Demo video:** https://youtu.be/Qrsvf5Yh5EU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2682

ChronoFact: memory that knows when something was true, and why it believes it

AI assistants that remember their users fail in two ways. They store things the user never said, and they keep giving answers that changed months ago. The usual fix, retrieval (RAG) over raw chat transcripts, returns old and current statements side by side. It can't tell which one is still true, and it often answers "I don't know" when the answer is there.

ChronoFact is a time-aware, verified memory layer built entirely on Meta models. Llama-3.1-8B extracts facts from each message, each with an exact quote. A fact is stored only if that quote appears in the user's words and BART-large-MNLI confirms the message supports it. When a value changes, the old fact keeps its end date instead of being overwritten. Each question is scoped to now, a past date or the full history. Matching facts and their source messages are then packed into a small, fixed context. Every fact links back to the user's own words and can be deleted on request.

On LongMemEval-S (500 users with ~115k-token histories), ChronoFact beats RAG by 13.6 points: 58.2% vs 44.6%. It is 2.5 times as accurate on time-based questions and twice as accurate across sessions. It also gives 3.9 times fewer false "not in memory" answers. It runs on one GPU at zero API cost.

## More from Meta Global AI Developer Hackathon

- [JSama](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2679)
- [Souldbound](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2680)
- [GridTwin-HIL](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2681)
- [XOF](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2683)
- [TrueScene](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2684)
- [Chalk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2685)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2682. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt