Paper

Title: Did You Check the Right Pocket? Cost-Sensitive Store Routing for Memory-Augmented Agents
arXiv: https://arxiv.org/abs/2603.15658
Published: 2026-03-08 (ICLR 2026 Workshop on Memory and State in LLM-Based Agents)

Key Technique

Formalizes memory retrieval as a store-routing problem: select which memory store to query (episodic, semantic, graph, working) based on query type and cost. Provides cost-sensitive decision framework trading answer accuracy against retrieval cost.

Oracle router achieves higher accuracy with substantially fewer context tokens vs uniform (query-all-stores) retrieval.

Why Relevant to Zeph

Zeph queries multiple stores per turn: SQLite conversation history, Qdrant semantic search, MAGMA graph BFS. Currently all stores are queried for every turn regardless of query type. This paper provides a principled framework for selective store routing.

Implementation sketch: classify query intent (factual lookup vs contextual vs graph-relational) and route to the cheapest sufficient store. Could reduce Qdrant calls for simple factual queries and graph BFS for non-relational queries.

Priority Rationale

P3: sound theoretical framework, but Zeph's store query costs are relatively low per-turn. Lower priority than #2441/#2442/#2443.