# Project Trace

- **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:** [Anshul Raj](https://cerebralvalley.ai/u/Anshul48)
- **GitHub:** https://github.com/Anshul48/trace
- **Demo video:** https://www.youtube.com/watch?v=8F1kTHVWkIQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5062

(The following is primarily AI written and heavily edited by me over hours).

My coding agent missed a critical constraint because its context overflowed. Another time it got lost in all the details I sent and built only part of the spec. So I built it a memory that sorts itself.


You know the pattern. The context window fills up, and the constraint that mattered, the format, the deadline, the file you said not to touch, slips out of view. Or you hand over every detail and the agent drowns in the pile, builds some of the spec and calls it done. You paste the plan in again, and it happens again. That is the frustration I built for.


TRACE is the memory I built instead. Give it a large library of text and it sorts the passages into groups on its own, with no labels added by people. The agent does not have to carry everything. It asks a question, and only the groups most likely to hold the answer wake up. The rest stay asleep. Every answer comes with a record you can check. The design is inspired by the brain.
I built this for developers and researchers whose AI helpers forget, or drown in too much detail. Finding and trusting what was saved is the daily pain. TRACE answers with records you can check.
TRACE searches a frozen library of 100 million tokens. It scored 0.79% of the library and kept 99.85% of exact search's top ten. (50 test searches; one of four equal parts.) Choosing and building the groups is not counted in that figure. Fine print: the frozen library holds 812,545 passages and 99.9 million tokens, 0.11% short of 100 million.


Try it now. Open https://anshul48.github.io/trace/explorer/, pick a recorded question, open a result and watch the fingerprint match. The project is TRACE: github.com/Anshul48/trace.

Honest limit: on my 50 test searches, ordinary keyword search put the right source in its top ten for 82% of searches, TRACE for 52% (exact search also got 52%). TRACE is for meaning search you can check, not a keyword replacement.
But Trace supplies the self-organization capability search doesn't, and Trace hybrid will integrate the best of both.

TRACE leads an ecosystem. ANIMA is a layer on the coding agent I already use: memory, a hidden checker, and a screen to watch the work (github.com/Anshul48/anima-substrate). Reflex is a separate line of work where an AI looks at a screen and finds the right control to press (SAM 3.1 with Muse Spark).


Built with Meta's Muse Code: 670 recorded worker jobs, about 170 hours of worker time. Details in the next field.
Your next job should start with what you already know.

## More from Meta Global AI Developer Hackathon

- [DGSelf](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5059)
- [1706526367569575](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5060)
- [QuartetIQ](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5061)
- [GajaMitra](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5063)
- [Solutions For Success LLC – Digital Employee](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5064)
- [Upamelm.ai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5065)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5062. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt