MemRouter: Embedding-Based Memory Routing Removes Per-Turn Decoding from Agent Write Path

arXiv·medium signal

A write-side memory router that decouples memory admission from the downstream answer backbone, replacing per-turn memory-management decoding with an embedding-based routing policy. Makes per-turn storage decisions entirely in embedding space via lightweight classification heads over a frozen LLM backbone. Yields higher matched-harness accuracy and substantially lower write-side latency than LLM-based managers, and the same router can be reused across QA backbones.