Extractors propose. One adjudicator commits.
Extractors propose. One adjudicator commits. The builder is never the judge. A proposed fact is not a fact you can act on until that commit. Independent crews that extract and judge their own output amplify errors. This site follows the same split: the page that states a claim is not the page that adjudicates it.
A debating crew over your knowledge can vote an error into the map. Mnemma does not run that topology.
Proposed facts wait. A separate adjudicator commits them — or they stay proposed. The brief can be wrong in public, with a receipt. It cannot quietly elect itself into memory.
This is not a Mnemma benchmark. It is a published wiring result we take as design: independent crews amplify their own errors; a centralized supervisor contains them.
Extractors propose. One adjudicator commits.
The builder is never the judge.
A published wiring study held prompts, tools, and compute fixed and moved only how agents were connected. Independent crews amplified errors 17.2× versus a single agent; a centralized supervisor held that amplification to 4.4×.
Proposed facts do not vote themselves into the map.
Customer pages cite this page. They do not re-host the paper names as a pitch.
Machine-readable → index.json · /ai-info · llms.txt