Research map / Multi-agent systems and games / LLM agents
RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory
Dabin Sheng, Zhe Wu, Jinming Zhao, Junliang Xing, Yuanchun Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an RL-guided temporal localization method with dual-channel retrieval for time-aware memory retrieval in long-horizon LLM agents.
Problem
Time-aware retrieval for long-horizon LLM agents over growing long-term memory; existing methods rely on lexical/semantic similarity and inject temporal cues only at organization or answer generation, so temporally mismatched but semantically similar memories crowd out targets.
Approach
Applies temporal constraints at the retrieval stage itself, using reinforcement learning to guide temporal localization within a dual-channel retrieval scheme (details beyond the abstract are not stated).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the distance-decay reranker (ablation) | TIDER · LifeBench | Overall accuracy (%): 57.60 -> 61.52 | Improved |
| Add the temporal channel (ablation) | TIDER · LifeBench | Overall accuracy (%): 52.21 -> 57.60 | Improved |
| Add the timeline summary S_q (ablation) | TIDER · LifeBench | Overall accuracy (%): 51.45 -> 52.21 | Improved |
| Add the timeline summary without a temporal channel (ablation) | semantic-only retrieval · LifeBench | F1: 63.05 -> 47.62 | Hurt |
| Add GRPO to the SFT-initialized Qwen3-1.7B locator (ablation) | TIDER locator · LifeBench | Overall accuracy (%): 55.64 -> 61.52 | Improved |
| Use a much larger zero-shot GPT-5.2 locator instead of the trained 1.7B locator (ablation) | TIDER locator · LifeBench | Overall accuracy (%): 61.52 -> 57.84 | Hurt |
| Supply gold oracle windows instead of predicted windows (ablation) | TIDER · LifeBench | Overall accuracy (%): +1.70% | No clear effect |
| Supply oracle windows on LoCoMo (ablation) | TIDER · LoCoMo Temporal | Temporal accuracy (%): 71.65 -> 79.13 | Improved |
| Add detection and window-localization rewards in GRPO (ablation) | TIDER · LifeBench | Overall accuracy (%): 55.88 -> 61.52 | Improved |
| Add outcome reward in GRPO (ablation) | TIDER · LifeBench | Overall accuracy (%): 59.07 -> 61.52 | Improved |
| Add evidence reward in GRPO (ablation) | TIDER · LifeBench | Overall accuracy (%): 59.56 -> 61.52 | Improved |
| Increase retrieval budget top-k from 20 to 100 (ablation) | TIDER · LifeBench | Overall accuracy (%): +0.49% | No clear effect |
Limitations
- Full text unavailable; only abstract-level information available
Closest papers by meaning
- Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.89)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.88)
- Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.88)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.88)
- SimpleMem: Efficient Lifelong Memory for LLM Agents (ICML 2026 · similarity 0.88)
- REMem: Reasoning with Episodic Memory in Language Agent (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
