Time‑aware retrieval is essential for long‑horizon LLM agents to exploit ever‑growing long‑term memory. Existing approaches rely on lexical or semantic similarity, causing temporally mismatched yet semantically similar memories to crowd out the target. Even methods that incorporate temporal cues do so only during memory organization or answer generation, overlooking temporal constraints at the retrieval stage where candidate selection occurs. Inspired by the temporal contiguity effect in human episodic memory, we introduce TIDER (Temporal Interval‑driven Dual‑channel Evidence Retrieval), a time‑aware retrieval framework. At its core lies a reinforcement‑learning trained temporal interval locator that predicts the time windows containing the evidence. RL treats window localization as a discrete, non‑differentiable decision and uses composite rewards to turn sparse answer supervision into fine‑grained retrieval feedback. These windows define the temporal retrieval channel; candidates from this channel are merged with those from a semantic channel and reranked according to proximity to the windows. Temporal grounding enables TIDER to surface evidence that would otherwise be drowned out by a sea of semantically similar but temporally misaligned memories. With a lightweight 1.7B parameter locator, TIDER achieves state‑of‑the‑art accuracy on two long‑horizon memory benchmarks—61.52% on LifeBench (+8.82%) and 78.96% on LoCoMo (+3.66%)—demonstrating the crucial role of temporal cues in pinpointing target evidence.
Review: TIDER’s RL‑driven window modeling fills a critical gap in retrieval‑time temporal reasoning, offering an interpretable and efficient solution for long‑term memory access.