MemoryAthena: Adaptive Routing over Latent and Generated Memories

The paper investigates combining generated memory with retrieved memory through MemoryAthena, which utilizes three pathways: direct Engram retrieval (E), generation from retrieved cues (GE), and generation from backbone states (GH). Using E as an anchor, a lightweight causal routing head is trained via counterfactual token likelihood advantages to determine when generated representations should intervene through bounded interpolation. On question answering, MemoryAthena raises the five-task average from 37.65 to 39.28, while the six-task general NLP average increases from 76.73 to 79.13, utilizing approximately 201M parameters for the complete memory-side system.

MemoryAthena raises the five-task average from 37.65 to 39.28 over the direct pathway of the same checkpoint.

The six-task general-NLP average increases from 76.73 to 79.13.

The complete memory-side system contains approximately 201M parameters, excluding the frozen backbone.