ROAM sorts memory updates into four relationship types

A new paper called ROAM splits every incoming memory fact into four relationship types before deciding what to store. That matters for anyone storing facts an agent will act on later. An outdated fact can otherwise sit in context right next to the one that replaced it. ROAM lifts answer accuracy by up to 29.8 points over baseline.

Each new observation gets classified against what's already stored: independent, equivalent, subsuming, or conflicting. ROAM then sorts observations into two roles, Primary and Evidence, and retrieves only the Primary ones. A superseded fact no longer competes with the fact that replaced it.

Answer-critical source recall improves 15.6 points, and confounder tokens make up 11.5 points less of what gets retrieved. The paper reports the accuracy gains hold across different manager model scales.

The case for the approach goes beyond the benchmark table. A standard atomic-memory call bundles three separate jobs into one shot. It reads what a new fact means, decides what to do with it, and writes the stored content. Any one of those three can go wrong without the other two catching it. Typing the relationship between old and new information first turns that judgment call into a step that can be checked before storage happens.

A team doesn't need the full system to use the idea behind it. Retrieving only the facts marked current keeps a contradiction from sitting in context beside the fact it replaced.