# CortexNET

- **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:** [Yogeshwar Sivakumar](https://cerebralvalley.ai/u/Noot)
- **GitHub:** https://github.com/yogeshwars-cys/cortexnet
- **Demo video:** https://youtu.be/AfeqgEAtxmo
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/351

CortexNET gives coding agents a capability they lack: working through a real repository from a small context window. Agents usually read whole files, so a model with an 8k window, like Llama 3.1 8B on a 6 GB laptop GPU, overflows on the first big file. Hosted models pay for every re-read token. CortexNET sits between the agent and the code. A read returns the file's outline, the lines the request names, and only the bodies an include model predicts the next step needs; every follow-up range request is a miss it learns from. A watermark fade shrinks stale reads to their outlines in one batch, so the conversation fits the window and the prompt cache keeps working.
Results, same model in every arm, only the read tool changed:
 - On 320 machine-written questions about repositories it had never seen, Llama 3.1 8B answered 66% vs 35% for standard paged file reads at 8k, and 77% vs 43% at 16k.
 - Muse Spark 1.3 answered 20/20 questions in one session for $0.18 vs $1.91 with whole-file reads.
 - Muse Spark Open found the right symbol on SWE-bench Lite 48% vs 30% of the time (p = 0.02).
 - Llama 3.2 3B: 59% vs 18%.
It ships as a Python library, a terminal coding agent (Cortex CLI), an MCP server and an egress proxy for Claude Code and Codex.

## More from Meta Global AI Developer Hackathon

- [ctzurcanu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/348)
- [Defcon1](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/349)
- [MxA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/350)
- [WalletShield](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/352)
- [Whiff Scout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/353)
- [MuseCell](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/354)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/351. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt