RoMem: Fixing AI Agents' Quietly Outdated Memories
Key Info
RoMem is a drop-in reranker designed to fix agents' quietly outdated memories: instead of deleting stale facts, it rotates them out of alignment at query time so agents rely on current information.
Highlights
- Plug-and-play: integrates as a reranker in roughly three lines of code.
- Spin, don't delete: outdated facts are shadowed, not erased, and rotate out at query time.
- Adaptive freshness: a pretrained gate estimates how fast each relation changes (e.g., born_in changes slowly, works_at quickly), and generalizes zero-shot to unseen relations.
- Time as rotation: time is modeled as a continuous rotation rather than a discrete label.