MemDuet: Learning Dual-Process Memory for Long-Context Reasoning

Jiangcheng Song; Haodong Jing; Yulin Mu; Lesong Tao; Jixin Wang; Jin Cui; Tianjing Ren; Runxuan Xie

cs.CLOct 14, 2026

MemDuet: Learning Dual-Process Memory for Long-Context Reasoning

Authors: Jiangcheng Song, Haodong Jing, Yulin Mu, Lesong Tao, Jixin Wang, Jin Cui, Tianjing Ren, Runxuan Xie,

Organizations: State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University · Faculty of Electronic and Information Engineering, Xi’an Jiaotong University · Imperial College London

Abstract

Reasoning over infinitely long documents is becoming a reality with recurrent memory agents. However, increasing context length exacerbates the trade-off between reasoning performance and inference efficiency. Inspired by Kahneman's account of fast and slow thinking, we propose MemDuet, a dual-process memory framework that learns to coordinate a small language model for fast, parallel evidence selection from raw context chunks (System 1) with a large language model for memory-update verification and memory updates in document order (System 2). We train MemDuet in three stages: supervised fine-tuning of System 1 for evidence selection, Group Relative Policy Optimization (GRPO) of System 2 for verification and memory rewriting, and joint optimization of both policies to coordinate selection and updating for long-context reasoning. Experiments on diverse long-context reasoning benchmarks (RULER-HQA, LongBench v2, and 2WikiMultiHopQA) demonstrate a favorable accuracy-efficiency trade-off as context length grows. Averaged across evaluation settings on the three datasets, MemDuet achieves an end-to-end inference speedup of 9.74x and an output-token reduction of 73.91% relative to RL-MemAgent.

Figures & tables

Figure 1: Accuracy–efficiency trade-off on RULER-HQA. Accuracy, inference time, and output tokens are equally averaged over the nine shared length settings (7K–1.75M). Qwen and DS-Distill use the available settings. Circle areas use a square-root-compressed scale for mean output tokens per question; circle labels give actual output-token ratios to MemDuet (1

\times

Figure 2: MemDuet’s overall inference workflow. System 1 predicts Skip / Update in parallel for raw chunks under a shared memory state. Candidate chunks retain their original text. System 2 verifies whether a selected chunk warrants a memory update given the current memory, then generates and commits memory from the first accepted chunk in document order. Later chunks are reevaluated under the updated memory.

Figure 3: Three-stage training of MemDuet. Stage I trains System 1 with weighted classification and a paired-memory hinge loss. Stage II freezes System 1 and optimizes System 2 with GRPO. Stage III jointly optimizes both systems using task rewards.

Accuracy (%, $\uparrow$ )
Method	7K	14K	28K	56K	112K	224K	448K	896K	1.75M	3.5M
Search-R1 7B	72.66	71.88	67.71	73.96	66.67	62.50	64.58	67.71	67.19	—
QwenLong-L1 32B	72.66	75.00	72.66	60.94	31.25	17.19	13.28	11.72	N/A	N/A
DS-Distill 32B	70.31	66.41	65.62	46.88	23.44	13.28	7.81	7.03	N/A	N/A
RL-MemAgent 7B	81.25	81.25	82.03	80.47	79.69	75.78	76.56	74.22	77.34	71.88
LycheeMemory ( $-G$ ) 7B	80.47	81.25	82.03	81.25	80.47	79.69	75.00	75.78	78.12	—

Table 1: Accuracy and inference efficiency on RULER-HQA across context lengths from 7K to 3.5M tokens. Relative to RL-MemAgent, speedup (

\uparrow

) is

T_{\mathrm{base}}/T_{\mathrm{method}}

, and the output-token ratio (

\downarrow

) is

O_{\mathrm{method}}/O_{\mathrm{base}}

, where

T

and

O

denote inference time and generated-token count, respectively. LycheeMemory (

-G

+G

) disables/enables its native memory gate. ∗ denotes results from our reproduction. Bold and underlined values indicate the best and second-best reported results at each length within each metric block. “N/A” denotes unavailable results; “-” denotes results not reported in the original paper.

Figure 4: Efficiency scaling on RULER-HQA from 7K to 3.5M tokens. The upper and lower panels report latency and total output tokens from both systems per example, respectively, on logarithmic axes.

LongBench v2
Metric	Method	Short	Medium	Long	Ultra
	MemAgent	33.62	31.91	25.00	28.28
	LycheeMemory ( $-G$ )	25.86	31.91	29.00	28.28
Option Acc. $\uparrow$	MemDuet	31.03	31.38	30.00	29.29
	LycheeMemory ( $-G$ )	$0.45\times$	$0.54\times$	$0.55\times$	$0.59\times$
Speedup $\uparrow$	MemDuet	$1.69\times$	$1.96\times$	$3.00\times$	$8.03\times$

Table 2: LongBench v2 (top; option accuracy, %) and long-context 2WikiMultihopQA (bottom; sub-EM, %) results. LycheeMemory uses no gate (

-G

); bold marks the best values per length and metric.

Framework	Acc. (%) $\uparrow$			Speed $\uparrow$	Tokens $\downarrow$
	112K	448K	896K	Speed $\uparrow$	Tokens $\downarrow$
MemDuet	87.50	82.81	82.03	$12.16\times$	$.0719\times$
Inference Mechanisms
w/o parallel execution	87.50	82.81	82.03	$11.84\times$	$.0661\times$
w/o draft–verify	84.38	78.91	78.13	$9.89\times$	$.0769\times$
w/o state reevaluation	84.38	80.47	78.13	$11.58\times$	$.0712\times$

Table 3: Ablations of MemDuet’s inference mechanisms and three-stage training on RULER-HQA. Accuracy is reported at 112K, 448K, and 896K tokens. Speed and Tokens equally average per-length ratios relative to RL-MemAgent. W/o Stage I retains Stages II and III; w/o Stage II retains Stages I and III; w/o Stage III retains Stages I and II.

System 1	F1 (%) $\uparrow$	Acc. (%) $\uparrow$	Speed $\uparrow$	Tokens $\downarrow$
Qwen2.5-0.5B	78.49	80.08	$5.69\times$	$.2690\times$
Qwen2.5-1.5B	81.68	82.89	$7.28\times$	$.2500\times$
Qwen3.5-0.8B	86.05	82.03	$14.06\times$	$.2149\times$
Qwen3.5-2B	84.47	83.36	$14.26\times$	$.2336\times$

Table 4: System 1 on RULER-HQA (7K–3.5M tokens). Acc., Speed, and Tokens are ten-length means; the latter two use per-length ratios to RL-MemAgent. F1 measures Update selection. Appendix C.2 gives full results and timing normalization.

$k$	Speed $\uparrow$
	112K	448K	896K	1.75M
[1.5pt][1.5pt]1	$1\times$	$1\times$	$1\times$	$1\times$
[1.5pt][1.5pt]2	1.006 $\boldsymbol{\times}$	1.032 $\boldsymbol{\times}$	$1.010\times$	$1.087\times$
[1.5pt][1.5pt] 4	$0.968\times$	$0.981\times$	1.080 $\boldsymbol{\times}$	$1.107\times$
[1.5pt][1.5pt]8	$0.920\times$	$0.867\times$	$0.919\times$	1.111 $\boldsymbol{\times}$
[1.5pt][1.5pt]16	$0.768\times$	$0.777\times$	$0.838\times$	$0.927\times$

Table 5: Effect of the parallel chunk window width

k

. Speedups are measured relative to the same-length serial configuration (

k=1

); values below

1\times

indicate slower execution. Bold marks the highest speedup at each length, and shading denotes the default

k=4

API family	Method	Option Acc. (%) $\uparrow$	Latency (s) $\downarrow$	Cost ( $)$ \downarrow$
DeepSeek-V4	MemAgent	57.46	634.23	0.1974
	MemDuet	57.65	178.55	0.0825
GPT-5.6	MemAgent	61.63	438.92	0.5978
	MemDuet	61.03	128.46	0.2398

Table 6: Training-free transfer to proprietary models on LongBench v2, reporting option accuracy. MemDuet combines off-the-shelf System 1 and System 2 APIs using binary accept-or-reject verification. Latency and cost include both systems.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Training / validation instances	32,768 / 128
Questions $\times$ rollouts	$128\times 16=2{,}048$
Question limit	1,024 tokens
Chunk / memory / answer limits	5,000 / 1,024 / 1,024 tokens
Maximum chunks / selector window	8 / 4
Trainable parameters	All System 2; System 1 frozen

Appendix

Table A1: Stage II training configuration.

Setting	System 1: selector	System 2: verifier / memory updater
Backbone	Qwen3.5-2B	Qwen2.5-7B-Instruct
Trainable parameters	Existing LoRA adapters; backbone frozen	New LoRA adapters; backbone frozen
LoRA rank / scaling $(r/\alpha_{\mathrm{LoRA}})$	32 / 64	32 / 32
LoRA target modules	Same as Stage I	All-linear
Dropout during joint RL	Disabled in rollout and policy replay	Disabled in rollout and policy replay
Training budget	5 dataset epochs

Appendix

Table A2: Stage III joint LoRA training configuration.

Model	Input	Cached input	Output
DeepSeek-V4.1-Flash	0.15	0.003	0.60
DeepSeek-V4-Pro-0813	0.66	0.022	1.98
GPT-5.6 Luna	0.20	0.020	1.20
GPT-5.6 Sol	4.00	0.400	20.00

Appendix

Table A3: API prices in USD per million tokens. DeepSeek prices are off-peak rates.

Performance: Accuracy (%, $\uparrow$ )
Method	7K	14K	28K	56K	112K	224K	448K	896K	1.75M	3.5M
RAG 7B	67.19	66.41	66.41	67.19	64.84	64.06	62.50	61.72	62.38	—
Search-R1 7B	72.66	71.88	67.71	73.96	66.67	62.50	64.58	67.71	67.19	—
QwenLong-L1 32B	72.66	75.00	72.66	60.94	31.25	17.19	13.28	11.72	N/A	N/A
QwenLong-L1.5 [-2pt]30B-A3B	67.97	68.75	66.41	63.28	61.72	59.38	43.75	39.06	28.13	—
Qwen2.5-Instruct 32B	69.53	64.84	60.16	51.56	44.53	21.88	14.06	7.03	N/A	N/A

Appendix

Table A4: RULER-HQA length scaling. The three panels report accuracy (%; local values use sub-EM), time speedup relative to MemAgent-7B, and relative output tokens (MemAgent =

1\times

). Output tokens include both systems’ outputs. ∗ denotes results from our reproduction. Within each column and panel, bold and underlined values indicate the best and second-best distinct values, respectively. “-” denotes unreported results; N/A denotes unavailable results.

Acc. (%) $\uparrow$
System 1	7K	14K	28K	56K	112K	224K	448K	896K	1.75M	3.5M	Mean
Qwen2.5-0.5B	86.72	77.34	88.28	78.13	79.69	83.59	77.34	77.34	75.78	76.56	80.08
Qwen2.5-1.5B	86.72	84.38	89.06	82.03	85.16	82.81	82.03	79.69	78.91	78.13	82.89
Qwen3.5-0.8B	87.50	82.03	89.06	82.03	80.47	81.25	77.34	83.59	78.91	78.13	82.03
Qwen3.5-2B	87.50	83.59	89.06	79.69	87.50	85.94	82.81	82.03	79.69	75.78	83.36
Speed $\uparrow$

Appendix

Table A5: Complete System 1 backbone results on RULER-HQA (7K–3.5M tokens). Mean is the arithmetic average of the ten per-length values. Speed and Tokens are ratios relative to RL-MemAgent.

Dual-Process Reasoning

See all 9 papers

PI-Mem: Pushing Long-Context Reasoning to 3.6M Tokens with Parallel-Iterative Memory
Aug 4, 2026Dawei Liu, Haixu Song, Shuang Cheng +9Memory-Augmented Language Models Long-Context QA
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Tsinghua University +3
CoEM: Empowering Long-Context Reasoning with Commit-on-Evidence Memory
Sep 29, 2026Jingguang Li, Yebo Wu, Zuyi Guo +7Long-Context Reasoning in Language Models Memory-Augmented Language Models
Korea Advanced Institute of Science & Technology · University of Macau · Massachusetts Institute of Technology
MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading
May 11, 2026Baibei Ji, Xiaoyang Weng, Juntao Li +3Long-Context Inference Memory-Augmented Language Model Agents
Soochow University · Peking University

Aug 4, 2026J/K move · Enter open · S save

PI-Mem: Pushing Long-Context Reasoning to 3.6M Tokens with Parallel-Iterative Memory

Dawei Liu, Haixu Song, Shuang Cheng, Shijie Wang+8

Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Tsinghua University+3

Memory-Augmented Language Models Long-Context QA LLM Inference Acceleration

Open paper AskPDF arXiv

Long-context reasoning remains a critical bottleneck for large language models, as recent recurrent-memory approaches face two inherent challenges: sequential chunk-wise updates can overwrite early critical evidence with later irrelevant content, and serial inter-chunk dependencies limit parallelism and cause latency to increase with context length. To address these issues, we propose PI-Mem (Parallel-Iterative Memory), a mechanism that processes all chunks in parallel and iteratively refines a shared memory over a bounded number of turns. In each turn, PI-Mem reads all chunks in parallel conditioned on the current memory, selects new or complementary evidence from each chunk, and merges the selected evidence into a compact shared memory for the next turn. To discourage redundant turns, we optimize the workflow through reinforcement learning with an auxiliary turn-efficiency reward, enabling the model to adaptively exit once sufficient evidence has been accumulated. We evaluate PI-Mem with Qwen3.5-35B-A3B and Qwen2.5-7B on the HotpotQA benchmark across context lengths up to 3.6 million tokens and find that it outperforms the recurrent-memory baseline by +6.25 and +7.81 absolute points while achieving 6.1

\times

and 2.1

\times

inference speedups, respectively. These results demonstrate that PI-Mem breaks the accuracy--efficiency trade-off in long-context reasoning and provides a scalable approach to complex multi-hop question answering over extremely long documents.

Abstract

Dual-Process Reasoning

Explore similar work

PI-Mem: Pushing Long-Context Reasoning to 3.6M Tokens with Parallel-Iterative Memory

CoEM: Empowering Long-Context Reasoning with Commit-on-Evidence Memory

MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading