Long‑term agent memory must retrieve answer‑relevant evidence that is scattered across turns and sessions. Existing approaches increasingly rely on structured memories and adaptive retrieval, yet they often pass retrieved or expanded units directly to the reader, coupling memory addressing with evidence presentation and leaving the final context either redundant or incomplete.
To bridge this gap, we introduce MemCompile, a query‑conditioned evidence compilation framework for long‑term conversational memory. Given a query, the Address Plane locates relevant support using compact source‑linked descriptors and capacity‑aware Cells, while a normalization boundary resolves heterogeneous retrieval hits into stable SourcePacket identities. The Evidence Plane then allocates a bounded evidence budget across complementary packets, admits limited additional support, and renders the retained claims of each selected packet into an answer‑ready context.
Empirically, MemCompile achieves 80.52% Judge Accuracy on the LoCoMo benchmark and extends the quality‑cost Pareto frontier into a higher‑quality regime. Its strong performance on LongMemEval‑S further suggests that query‑conditioned compilation generalizes across long‑term conversational‑memory benchmarks.
Review: By decoupling retrieval from presentation and introducing budget‑constrained evidence compilation, MemCompile markedly improves the completeness and answerability of long‑term dialogue systems, offering a promising direction for future memory‑augmented models.