# PMO (Permanent Memory Optimizer)

- **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:** [Adam Barnett](https://cerebralvalley.ai/u/Aphiticus)
- **GitHub:** https://github.com/adamspencerbarnett-afk/Muse_Dashboard_PMO
- **Demo video:** https://adamspencerbarnett-afk.github.io/Muse_Dashboard_PMO/
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4123

Muse Dashboard + PMO gives Muse Code a permanent memory of a project outside the chat.

Today an AI agent rereads its whole history at every step. By prompt 20 most of that history is stale, yet it is paid for again in tokens, time, GPU cache and power.

PMO recalls only what each message needs, a card of about 3k tokens instead of about 179k tokens of old transcript. It captures rules, files, commands, results and open work from the session log, with no extra model calls. At 50k tokens a fresh session starts from the card. The chat continues and nothing is lost.

The Dashboard makes this visible with live cost and token meters, handover markers, file previews, a 3D viewer, snapshots with undo and scheduled jobs, all running locally.

Memory is the platform and skills are the apps. PMO shows Muse the one skill that fits the message, for chat, coding, research and debugging, or 3D modelling from one picture.

In paired tests with PMO on and off, input tokens fell 73% with 46 of 46 quality checks passed. Peak context fell from 179k to 56k tokens. One 30 prompt project fell from $12.08 to $1.40 with the same test results. The 3D skill passed 9 of 9 checks where plain Muse passed 0 of 9. The project has 1,197 checks across 29 suites, all green.

Small context. Total recall.

## More from Meta Global AI Developer Hackathon

- [teardown](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4120)
- [HireSense by LV8N Tech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4121)
- [Team Crew Daniel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4122)
- [WattDocket](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4124)
- [Padhai Saathi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4125)
- [Runtime Terror](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4126)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4123. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt