RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory

Dabin Sheng; Zhe Wu; Jinming Zhao; Junliang Xing; Yuanchun Shi

cs.AIOct 16, 2026

RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory

Authors: Dabin Sheng, Zhe Wu, Jinming Zhao, Junliang Xing, Yuanchun Shi

Organizations: Key Laboratory of Pervasive Computing, Ministry of Education, Department of Computer Science and Technology, Tsinghua University · Qiyuan Lab

Abstract

Time-aware retrieval is crucial for long-horizon LLM agents to effectively leverage ever-growing long-term memory. Existing methods focus on lexical or semantic similarity, leading to target memories being crowded out by temporally mismatched yet semantically similar candidates. Even time-aware approaches inject temporal cues only during memory organization or answer generation, overlooking temporal constraints at the retrieval stage, where candidate selection occurs - the root cause of the displacement. Inspired by the temporal contiguity effect in human episodic memory, we propose TIDER (Temporal Interval-driven Dual-channel Evidence Retrieval), a time-aware agentic retrieval framework. Central to TIDER is a reinforcement learning (RL)-trained temporal interval locator that infers the temporal windows in which the evidence lies. RL optimizes window localization as a discrete, non-differentiable retrieval decision, using composite rewards to convert sparse answer supervision into fine-grained retrieval feedback. These temporal windows first define the temporal retrieval channel; the retrieved candidates are then merged with results from the semantic channel and reranked by proximity to the windows. This temporal grounding enables TIDER to retrieve temporally aligned evidence that would otherwise be drowned out by the volume of semantically similar yet temporally mismatched memories. With a lightweight 1.7B locator, TIDER achieves state-of-the-art accuracy on both long-horizon memory benchmarks - 61.52% on LifeBench (+8.82%) and 78.96% on LoCoMo (+3.66%) - which indicates that temporal cues are crucial for pinpointing target evidence amid a vast pool of semantically similar historical events in long-term daily-life memory.

Figures & tables

Figure 1: Time-sensitive queries require evidence from a specific window, but semantic-only retrieval ranks semantically similar memories from the wrong time. TIDER predicts a temporal window to define the temporal channel, and promotes temporally aligned evidence into the retrieved context.

Figure 2: Overview of TIDER. TIDER uses a temporal interval locator to decide whether a query requires temporal constraints and, if so, to predict explicit temporal windows. The windows define the temporal channel, whose candidates are merged with semantic candidates and reranked by temporal distance; non-temporal queries fall back to the semantic-only path. The locator is trained with GRPO under rewards for switch detection, window localization, evidence coverage, and answer correctness.

Method	IE $\uparrow$	MR $\uparrow$	ND $\uparrow$	TKU $\uparrow$	Overall $\uparrow$	F1 $\uparrow$
TsContriever [ 28 ]	31.28 $\pm$ 1.05	23.33 $\pm$ 0.00	27.43 $\pm$ 0.60	23.91 $\pm$ 0.00	27.21 $\pm$ 0.42	49.36
MemOS [ 12 ]	54.57 $\pm$ 1.43	48.89 $\pm$ 0.48	42.01 $\pm$ 1.20	31.16 $\pm$ 1.26	47.30 $\pm$ 0.88	45.90
Zep [ 21 ]	43.61 $\pm$ 0.40	39.44 $\pm$ 0.96	32.99 $\pm$ 0.60	30.43 $\pm$ 2.17	38.40 $\pm$ 0.37	38.55
Mem0 [ 5 ]	57.99 $\pm$ 1.05	48.33 $\pm$ 2.89	50.35 $\pm$ 1.59	44.20 $\pm$ 6.64	51.80 $\pm$ 1.26	51.94
TiMem [ 11 ]	17.58 $\pm$ 1.43	17.50 $\pm$ 1.44	21.53 $\pm$ 3.01	10.87 $\pm$ 2.17	17.73 $\pm$ 0.57	18.89
NeSTR [ 13 ]	56.16 $\pm$ 1.37	52.22 $\pm$ 1.27	53.12 $\pm$ 1.80	42.03 $\pm$ 3.32	52.70 $\pm$ 0.65	61.25

Table 1: Performance comparison on the LifeBench benchmark. Categories include Information Extraction, Multi-hop Reasoning, Nondeclarative Memory Reasoning, and Temporal and Knowledge Updating. Best non-Oracle results are bolded ; underline indicates second best.

Method	Single-Hop $\uparrow$	Multi-Hop $\uparrow$	Temporal $\uparrow$	Open-Domain $\uparrow$	Overall $\uparrow$	F1 $\uparrow$
TsContriever [ 28 ]	77.41 $\pm$ 0.45	57.92 $\pm$ 0.33	62.93 $\pm$ 0.25	39.58 $\pm$ 0.85	68.47 $\pm$ 0.16	48.94
MemOS [ 12 ]	76.07 $\pm$ 0.10	56.85 $\pm$ 0.69	69.47 $\pm$ 0.25	45.14 $\pm$ 0.49	69.24 $\pm$ 0.11	45.02
Zep [ 21 ]	50.42 $\pm$ 0.29	42.20 $\pm$ 0.77	19.11 $\pm$ 0.29	38.19 $\pm$ 0.49	41.62 $\pm$ 0.21	26.88
Mem0 [ 5 ]	62.09 $\pm$ 0.42	50.14 $\pm$ 0.32	59.25 $\pm$ 0.41	37.70 $\pm$ 0.41	57.79 $\pm$ 0.34	42.52
TiMem [ 11 ]	81.43 $\pm$ 0.05	62.20 $\pm$ 0.82	77.63 $\pm$ 0.34	52.08 $\pm$ 0.74	75.30 $\pm$ 0.16	54.40
NeSTR [ 13 ]	82.60 $\pm$ 0.25	63.00 $\pm$ 1.02	66.56 $\pm$ 0.18	47.57 $\pm$ 0.60	73.48 $\pm$ 0.23	53.54

Table 2: Performance comparison on the LoCoMo benchmark. Reported categories include Single-hop, Multi-hop, Temporal Reasoning, and Open-domain Knowledge. Best results are bolded ; underline indicates second best.

Figure 3: LifeBench accuracy stratified by gold-window count. Bold labels report TIDER’s absolute accuracy; colored labels report each baseline’s delta from TIDER within the same category.

Variant	$\mathcal{S}_{q}$	Temporal	Decay $\gamma$	IE $\uparrow$	MR $\uparrow$	ND $\uparrow$	TKU $\uparrow$	Overall $\uparrow$	F1 $\uparrow$
TIDER (full)	✓	✓	✓	67.12	60.83	60.42	47.83	61.52	69.23
w/o decay rerank	✓	✓	$\times$	63.70	58.33	55.21	41.30	57.60	67.02
w/o temporal channel	✓	$\times$	$\times$	58.22	52.50	51.04	34.78	52.21	47.62
w/o summary (semantic)	$\times$	$\times$	$\times$	55.48	51.67	48.96	39.13	51.45	63.05

Table 3: Module ablations on LifeBench. Successively removing the distance-decay reranker, the temporal channel, and the timeline summary

\mathcal{S}_{q}

from TIDER; the bottom row is a pure semantic baseline. Best in column bold .

Figure 4: Locator substitution on LifeBench. Cells show accuracy differences (%) from TIDER; column headers show TIDER’s baseline scores.

Figure 5: Hyperparameter sensitivity analysis on LifeBench.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Recall@10	Recall@20	nDCG@20	MRR@20
TsContriever	17.91	23.84	20.48	42.98
Semantic retriever (NeSTR input)	32.74	39.35	37.21	67.55
TIDER	33.18	48.36	48.67	73.71

Appendix

Table 4: Retrieval metrics on the same LifeBench memory pool

\mathcal{M}

. All values are reported as percentages and use LLM-judged binary relevance. Best results are bolded .

User	#Items	#Clusters	Items / cluster (avg)	Timeline span (days)
user-0	136,208	15,603	8.73	990
user-1	128,538	14,847	8.66	833
user-2	132,862	15,369	8.64	2,726
user-3	131,961	15,243	8.66	823
user-4	121,168	14,036	8.63	7,014
user-5	88,335	10,656	8.29	1,308

Appendix

Table 5: LifeBench memory-cluster statistics. “#Items” counts raw turns absorbed into clusters; “#Clusters” is the number of surviving event clusters after temporally aware greedy merging; “Items / cluster (avg)” is the arithmetic mean cluster size; “Timeline span (days)” is the span between the earliest and latest cluster date.

Category	N	Top-30 localization coverage
Non-temporal	50	87.60
Single-window	228	89.45
Dual-window	103	73.51
Multi-window	27	78.85
Overall	408	84.51

Appendix

Table 6: Coarse temporal-window localization coverage of the locator’s top-30 timeline-summary input on LifeBench. Coverage is judged by an LLM reviewer and averaged per question.

Parameter	Value
Advantage estimator	GRPO
KL coefficient $\beta$ (loss)	0.1
Group size $G$ (rollouts per prompt)	8
Rollout temperature	0.7
Peak learning rate	$5\times 10^{-6}$
LR schedule	cosine, warmup ratio 0.1

Appendix

Table 7: GRPO hyperparameters for training the locator

\pi_{\theta}

. The SFT-aligned Qwen3-1.7B policy is further optimized with GRPO using the configuration below.

Bucket ( $\|W_{\text{gold}}\|$ )	DAR factor	Effective probability
0 (non-temporal)	10.0	0.159
1 (single-window)	1.0	0.424
2 (dual-window)	1.0	0.160
$\geq 3$ (multi-window)	4.0	0.257

Appendix

Table 8: Difficulty-aware rollout reweighting (DAR) by gold-window count. “Effective probability” applies the DAR factors to the empirical training distribution and renormalizes the result.

Variant	IE $\uparrow$	MR $\uparrow$	ND $\uparrow$	TKU $\uparrow$	Overall $\uparrow$	F1 $\uparrow$
TIDER (SFT+GRPO, 1.7B)	67.12	60.83	60.42	47.83	61.52	69.23
w/o $r_{\text{win}},r_{\text{det}}$	60.96	56.67	58.33	32.61	55.88	67.46
w/o $r_{\text{out}}$	66.44	56.67	58.33	43.48	59.07	69.20
w/o $r_{\text{evd}}$	64.38	60.83	56.25	47.83	59.56	68.81

Appendix

Table 9: Reward-component leave-one-out ablations on LifeBench. The first row is full TIDER; subsequent rows remove the indicated term(s) during GRPO. Best results are bolded .

Figure 6: GRPO training curves of the 1.7B locator on LifeBench.

Figure 7: Capacity ablation on the GRPO training curve. Per-step reward (EMA

\alpha{=}0.95

, bias-corrected) for the 4B locator and the 1.7B.

Locator	IE $\uparrow$	MR $\uparrow$	ND $\uparrow$	TKU $\uparrow$	Overall $\uparrow$	F1 $\uparrow$
TIDER (4B)	65.07	61.67	59.38	50.00	61.03	69.57
TIDER (1.7B)	67.12	60.83	60.42	47.83	61.52	69.23
$\Delta$ ( $1.7$ B $-4$ B)	$+2.05$	$-0.84$	$+1.04$	$-2.17$	$+0.49$	$-0.34$

Appendix

Table 10: Locator capacity ablation on LifeBench (user-0/1,

n{=}408

). Per-category accuracy is reported alongside Overall accuracy and average F1. Per-column best results are bolded ; the

1.7

B locator is our adopted default.

Figure 8: Hyperparameter robustness on LifeBench. Five hyperparameters are swept individually around the default configuration (

K{=}30

\lambda_{\text{boost}}{=}0.3

\gamma_{\text{min}}{=}0.3

\rho{=}0.05

\tau_{\text{buf}}{=}14

). Each panel shows Overall accuracy (%) versus the swept value; the orange star marks the chosen default; the upper dashed line is TIDER’s default-config score; the lower dotted line is the No-Temporal lower bound.

Memory representation	Semantic retrieval	TIDER retrieval
Fixed chunks + session timestamps	50.49	58.58
Full temporally anchored memory pool $\mathcal{M}$	52.21	61.52

Appendix

Table 11: A

2\times 2

factorial comparison of memory representation and retrieval method on LifeBench. Values are Overall accuracy (%).

Retrieval budget $k$	10	20	30	40	50	100
TIDER	52.70	61.52	60.78	58.33	58.58	62.01
Semantic Only	46.86	52.21	52.90	57.49	56.04	57.73

Appendix

Table 12: Retrieval-budget sweep on LifeBench with the same memory pool and answer generator. Values are Overall accuracy (%).

Parameter	Settings	Accuracy (%)
$\Delta_{\text{date}}$ (days)	3 / 7 / 14	61.76 / 61.52 / 59.07
$\theta_{\text{sem}}$	0.75 / 0.85 / 0.95	59.56 / 61.52 / 61.27
Cluster-size cap	20 / 40 / 80	57.84 / 61.52 / 60.54
Summarizer	GPT-4.1-nano / GPT-4o-mini / GPT-5-mini	60.29 / 61.52 / 63.73

Appendix

Table 13: Memory-construction sensitivity on LifeBench. Settings and accuracies in each row correspond in order; accuracy is reported in percent.

Method	Non-temporal	Single-window	Dual-window	Multi-window	Overall
MemOS	52.67 $\pm$ 1.15	48.83 $\pm$ 1.10	42.07 $\pm$ 1.48	44.44 $\pm$ 11.11	47.30 $\pm$ 0.88
TiMem	23.33 $\pm$ 1.15	16.96 $\pm$ 1.10	17.80 $\pm$ 1.48	13.58 $\pm$ 4.28	17.73 $\pm$ 0.57
Zep	48.67 $\pm$ 7.02	38.74 $\pm$ 0.67	33.98 $\pm$ 1.68	33.33 $\pm$ 0.00	38.40 $\pm$ 0.37
Mem0	63.33 $\pm$ 3.06	51.32 $\pm$ 1.16	47.25 $\pm$ 4.04	51.85 $\pm$ 3.70	51.80 $\pm$ 1.26
TsContriever	38.00 $\pm$ 2.00	28.07 $\pm$ 1.16	23.62 $\pm$ 1.12	13.58 $\pm$ 4.28	27.21 $\pm$ 0.42
NeSTR	56.00 $\pm$ 2.00	55.12 $\pm$ 0.67	46.60 $\pm$ 1.68	49.38 $\pm$ 2.14	52.70 $\pm$ 0.65

Appendix

Table 14: LifeBench accuracy stratified by gold-window-count category. Best results are bolded ; underline indicates second best.

Comparison	$\Delta$ Acc	$b$	$c$	Exact $p$	Holm $p$
TIDER vs. MemOS	13.48	92	37	$1.40{\times}10^{-6}$	$4.19{\times}10^{-6}$
TIDER vs. TiMem	43.14	188	12	$8.11{\times}10^{-42}$	$4.87{\times}10^{-41}$
TIDER vs. Zep	22.79	121	28	$5.59{\times}10^{-15}$	$2.24{\times}10^{-14}$
TIDER vs. Mem0	8.33	66	32	$7.70{\times}10^{-4}$	$8.94{\times}10^{-4}$
TIDER vs. TsContriever	34.07	149	10	$6.25{\times}10^{-33}$	$3.12{\times}10^{-32}$
TIDER vs. NeSTR	8.09	59	26	$4.47{\times}10^{-4}$	$8.94{\times}10^{-4}$

Appendix

Table 15: Paired significance tests for TIDER against each baseline on LifeBench.

\Delta

Acc is reported in percentage points.

Method	Wrong	Retrieval miss	Temporal mismatch	Partial evidence	Reason failure	Retrieval-side failure
TIDER	157	27 (17.2%)	107 (68.2%)	5 (3.2%)	18 (11.5%)	139 (88.5%)
MemOS	212	41 (19.3%)	132 (62.3%)	13 (6.1%)	26 (12.3%)	186 (87.7%)
TiMem	333	174 (52.3%)	142 (42.6%)	7 (2.1%)	10 (3.0%)	323 (97.0%)
Zep	250	88 (35.2%)	115 (46.0%)	18 (7.2%)	29 (11.6%)	221 (88.4%)
Mem0	191	45 (23.6%)	123 (64.4%)	9 (4.7%)	14 (7.3%)	177 (92.7%)
TsContriever	296	143 (48.3%)	122 (41.2%)	8 (2.7%)	23 (7.8%)	273 (92.2%)

Appendix

Table 16: LLM-judged primary error causes over incorrect LifeBench answers. Percentages are computed within each method’s incorrect answers.

Temporal IR

Temporal information retrieval finds information relevant to when events occurred or facts were valid, not just what they describe.

See all 65 papers

RaMem: Contextual Reinstatement for Long-term Agentic Memory
Jun 22, 2026Wei Yang, Bryce Kan, Shixuan Li +5Agent Memory LLM Agent Memory
University of Southern California
Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory
Oct 9, 2026Junyou Guo, Chongyang Tao, Yan He +2LLM Agent Memory Conversational Memory
School of Computer Science and Engineering, Beihang University, Beijing, China · Alibaba Group
RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory
Aug 13, 2026Jingbo Ji, Lingyi Li, Xilong Cheng +4Agent Memory Persistent Memory for Language Models
Communication University of China, Beijing, China. · Zhilian Yinghe Technology Co., Ltd., Beijing, China · State Key Laboratory of Media Convergence and Communication, Beijing, China

Jun 22, 2026J/K move · Enter open · S save

RaMem: Contextual Reinstatement for Long-term Agentic Memory

Wei Yang, Bryce Kan, Shixuan Li, Li Li+4

University of Southern California

Agent Memory LLM Agent Memory

Open paper AskPDF arXiv

Long-term memory has become increasingly important for LLM agents that operate across extended interactions and evolving task contexts. Recent memory systems have made past experiences more persistent, compact, and retrievable, but retrieval alone does not ensure that a memory provides valid evidence for the current query. When experiences are compressed into reusable fragments, memories from different situations may appear equally relevant if they involve recurring entities or user states. We refer to this failure as context collapse: memories lose the surrounding context needed to judge whether they provide valid evidence for the current query. To address this problem, we propose Contextual Reinstatement for Agentic Memory (RaMem), a framework that turns retrieved memory fragments into contextually verifiable evidence. RaMem operates through four coordinated stages: (i) evidence anchoring grounds each memory in its original episodic conditions, especially event time, mention time, session span, and participants; (ii) recall condition induction derives the evidence conditions implied by the query; (iii) validity-aware retrieval uses these conditions to prioritize context-compatible memories while retaining content-relevant candidates as fallback evidence; and (iv) context-preserved synthesis keeps the selected memories' structured context available to the generator. Experiments on long-term memory benchmarks show that RaMem consistently improves performance over strong memory baselines, with average F1 gains of more than 10% across several backbones.

Abstract

Temporal IR

Explore similar work

RaMem: Contextual Reinstatement for Long-term Agentic Memory

Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory

RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory