Long‑term memory enables LLM agents to reuse past experiences, yet retrieving merely relevant memories does not guarantee sufficient evidence for answering a query. Crucial information may be scattered across multiple records, hidden by abstraction, or disconnected from the relational structure needed for reasoning. We define this as the relevance‑sufficiency gap and introduce the QEMem framework, which reframes memory utilization as the construction of query‑specific evidence sets, shifting the focus from individual relevance to collective evidential support.
QEMem comprises three complementary mechanisms:
- Query‑guided discovery retrieves evidence across multiple memory granularities to capture fragmented information;
- Source‑grounded reification organizes the retrieved pieces into structured ViewCards, explicitly assigning evidential roles and source references;
- Sufficiency‑guided refinement selectively recovers missing information when the evidence set remains incomplete.
The framework operates on existing memory stores without altering how memories are formed. In experiments on the LoCoMo benchmark using Claude Opus 4.5, QEMem achieves 91.43% LLM‑judge accuracy, outperforming the strongest baseline MemR$^3$ by 7.6 percentage points and NEMORI by 10.7 points under a shared memory store. Further analysis shows improved evidence coverage and strong performance on questions that require cross‑session evidence integration.
Review