From Lossy to Verified: A Provenance-Aware Tiered Memory for Agents

Long-horizon agents often rely on write-time summaries to keep interaction histories manageable. But compression happens before the system knows what a future query will depend on. As a result, summary-only memory can remain topically relevant while omitting the query-critical detail required for a faithful answer. Always grounding on raw logs avoids this failure mode, but treating raw history as the default evidence source is costly and often unnecessary. We argue that long-horizon agent memory should therefore support \emph{inference-time evidence allocation}: for each query, the system should use the cheapest evidence that is still sufficient for faithful and traceable answering. We instantiate this principle in \textbf{TierMem}, a provenance-linked two-tier memory framework with three components: summary-first retrieval, selective escalation to immutable raw logs, and verified write-back of evidence-backed findings. Across two long-horizon memory benchmarks, TierMem improves the accuracy--efficiency frontier over summary-only memory while substantially reducing the cost of always-raw grounding. On LoCoMo, TierMem reaches 0.851 accuracy versus 0.873 for a raw-grounded, while reducing average input tokens by 54.1\% and latency by 60.7\%. Code is available at \url{https://github.com/FreedomIntelligence/Tiermem}.