# Doyle

- **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:** [Narayan S S](https://cerebralvalley.ai/u/nrynss)
- **GitHub:** https://github.com/nrynss/doyle
- **Demo video:** https://www.youtube.com/watch?v=cBD3CQrDqUc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2769

Doyle is truth maintenance for agent memory.

Coding agents remember what they learn: how an API pages, what a COBOL batch charges, which keys retire an old console. Each memory rests on something outside it, and when that thing changes, or was misread in the first place, nothing tells the agent. It keeps acting on the stale belief, and so does every other agent reading the same memory.

Doyle records what each memory rests on and keeps the agent off the stale ones until a person has looked:
- A curator model cites every memory, in Muse Code and in Lambo, against its evidence: an OpenAPI contract, a COBOL program, or the screens of a legacy console.
- When evidence changes, Doyle flags every memory resting on it. The evidence can be a contract diff, a diff of two parsed COBOL versions, a differential run (the COBOL batch and the agent's rewrite on the same input, mismatches traced to the paragraphs that ran), or a diff of console screens that Muse Wonder read from stills and a training video.
- A re-check model proposes a correction for each flag: correct, retire or keep.
- Doyle's guard, a Muse Code hook, denies any step that touches a flagged element, with the reason. It is deterministic, with no model on the hot path.
- A person decides on a review page. The correction is written back to every memory system, and the stale note is marked superseded.

The demo, recorded live: Muse Wonder retires a COBOL billing system with 4,812 accounts. Its memory says the 1.5% late fee has "no exceptions", but the batch exempts vaccination visits. Without Doyle, Wonder found the bug itself and shipped correctly, but the wrong memory survived for the next session. With Doyle, the differential run found 535 mismatches ($8,308.80 overcharged) blamed on one COBOL paragraph. The guard denied the rewrite and the ship script, and Wonder stopped and asked a person. After the review, session two worked from the corrected memory, shipped a two-line fix, and Doyle's re-run found 0 mismatches. Of 9 flags, 3 were false alarms; all are counted in the repo.

Written in Rust during the event. The gate and the demo run offline from a fresh clone, on committed recordings.

## More from Meta Global AI Developer Hackathon

- [Future Tech Industries](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2766)
- [Pitchside](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2767)
- [Three Billion](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2768)
- [Diwan](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2770)
- [FlatSpot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2771)
- [Hypot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2772)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2769. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt