ActiveMem: Dynamic Latent Memory Trees for Long-Horizon Agents

Song-Li Wu; Jingyi Wang; Zhaocheng Du; Weinan Gan

cs.AISep 27, 2026

ActiveMem: Dynamic Latent Memory Trees for Long-Horizon Agents

Authors: Song-Li Wu, Jingyi Wang, Zhaocheng Du, Weinan Gan

Organizations: Tsinghua University · Huawei Noah’s Ark Lab

Abstract

Large Language Model (LLM) agents increasingly rely on external memory to support long-horizon reasoning and decision making. Existing memory systems typically retrieve historical trajectories or summaries as independent context fragments, overlooking the procedural dependencies underlying multi-step execution. As memory scales, such flat retrieval introduces context fragmentation and cross-task interference, leading to structurally inconsistent reasoning trajectories. We propose ActiveMem, a hierarchical memory framework that recursively organizes agent experiences into dependency-aware latent execution trees. ActiveMem abstracts trajectories into reusable subtask nodes while explicitly preserving execution transitions, enabling coherent reasoning-path retrieval conditioned on the current execution state. To support continual adaptation, ActiveMem further learns dynamic memory expansion, retrieval, and pruning policies through reinforcement learning. Experiments across various agent benchmarks demonstrate that ActiveMem consistently improves task completion, reasoning stability, and memory efficiency over existing memory-based agents. Moreover, ActiveMem enables compact open-weight models to achieve competitive performance with substantially larger proprietary systems.

Figures & tables

Figure 1: Comparison between Flat Memory and ActiveMem. (a) Flat memory methods retrieve disparate context fragments and mixed workflows from a flat pool based on simple similarity, leading to context fragmentation and cross-task leakage (e.g., blindly applying hotel booking steps to a flight workflow). (b) ActiveMem organizes context using a structured memory pool with a tree hierarchy. It enforces task isolation to filter out irrelevant domains and preserves causal chains mapping to the HTML/UI, enabling the agent to accurately plan domain-specific steps and execute the correct transfer flight action.

Backbone	Method	ALFWorld	TriviaQA	PopQA	KodCode	BigCodeBench	GPQA	GSM8K	MATH
SmolLM3-3B	Vanilla	18.96	10.47	8.23	37.05	35.96	9.35	47.63	16.22
	CoT	17.60	12.88	9.95	38.45	39.42	20.70	58.91	56.33
	SFT	32.36	55.25	37.22	59.25	40.79	19.70	63.48	45.65
	GRPO	55.35	65.88	45.16	68.48	72.44	22.73	80.03	61.23
	REINFORCE	53.13	63.20	46.81	65.53	67.14	23.44	82.03	58.75
	REINFORCE++	53.95	63.20	44.10	65.90	68.80	22.73	81.50	59.89

Table 1: Results on SmolLM3-3B and Qwen3-8B. All values represent the performance metric for each task (e.g., accuracy %). We highlight the best and second best results.

Method	ALFWorld	TriviaQA	PopQA	KodCode	BigCodeBench	GPQA	GSM8K	MATH
ActiveMem ${}_{\text{GRPO}}$	95.57	87.46	73.20	83.42	84.39	54.49	94.81	92.76
w/o Tree Action Head	89.88	81.93	67.80	80.95	79.46	48.53	86.41	89.83
w/o Merge/Delete	93.11	84.24	71.24	81.54	83.64	52.13	88.02	87.11
w/o Hierarchical Advantage	90.74	83.25	69.93	82.46	81.35	50.98	86.74	90.86
w/o Latent Injection	88.96	81.67	66.81	81.92	78.44	51.76	85.98	87.35
w/o HLMT	83.25	78.64	58.92	73.41	69.88	44.53	82.47	86.39

Table 2: Ablation Study on Qwen3-8B. We evaluate the contribution of each major component in ActiveMem. All values denote task performance (%).

Model

2-Objective

8-Objective

16-Objective

\uparrow

\uparrow

Peak (

\times 10^{2}

)

\downarrow

Time (s)

\downarrow

\uparrow

\uparrow

Peak (

\times 10^{2}

)

\downarrow

Time (s)

\downarrow

\uparrow

\uparrow

Peak (

\times 10^{2}

)

\downarrow

Time (s)

\downarrow

Qwen2.5-14B-Inst

0.732

0.902

15.6

\pm

0.19

5.49

\pm

0.16

1.55

1.87

44.7

\pm

0.37

16.2

\pm

0.27

0.567

0.703

38.4

\pm

0.71

29.7

\pm

0.75

Qwen2.5-7B-Inst

0.268

0.366

19.6

\pm

0.33

4.60

\pm

0.08

0.87

1.10

49.5

\pm

0.40

13.9

\pm

0.18

0.165

0.213

43.3

\pm

0.62

15.5

\pm

0.23

Qwen2.5-7B-Inst (A-MEM)

0.286

0.371

14.1

\pm

0.10

24.6

\pm

0.51

1.13

1.43

18.6

\pm

0.10

53.7

\pm

1.26

0.730

0.961

18.8

\pm

0.14

91.2

\pm

2.44

Qwen2.5-7B-Inst (truncate)

0.262

0.336

8.28

\pm

0.06

5.89

\pm

0.16

0.97

1.23

11.8

\pm

0.10

11.9

\pm

0.20

0.396

0.497

13.3

\pm

0.16

22.1

\pm

0.60

Search-R1

0.452

0.531

13.0

\pm

0.08

4.09

\pm

0.23

0.064

0.08

24.7

\pm

0.19

4.25

\pm

0.16

0.009

0.011

20.9

\pm

0.03

4.75

\pm

0.18

Table 3: Quantitative comparison of models on multi-objective, multi-hop QA tasks. Arrows indicate the optimal direction for each metric. Underlined entries denote catastrophic model collapse (i.e., extremely low performance). Variants marked with (truncate) adopt ActiveMem’s prompt and rollout pipeline with context truncation. Variants labeled with (A-MEM) employ the prompt, rollout pipeline, and the external memory module from ActiveMem ( Xu et al. 2026 ) . ActiveMem-QA refers to our proposed ActiveMem optimized specifically on the 2-objective QA.

Figure 2: Generalization study of ActiveMem. We train ActiveMem SFT on a single dataset (either ALFWorld or TriviaQA) and evaluate it across four diverse datasets: TriviaQA, ALFWorld, ScienceWorld, and FEVER.

Figure 3: The generalization study. We train ActiveMem on GSM8K or KodCode and evaluate it on all four datasets.

Figure 4: Scaling trajectories of task performance and memory efficiency as a function of task complexity (number of objectives). ActiveMem (trained on the 2-objective QA dataset) consistently outperforms all baselines while maintaining a nearly constant memory footprint. Note that the apparent plateau in peak token usage for baseline models at 16 objectives is an artifact of catastrophic model collapse (e.g., premature trajectory termination), rather than an indication of sustained efficiency.

Settings	Hyperparameters
Training (SFT)	train_batch_size = 4
	learning_rate = 1e-5
	epochs = 2
	warmup_ratio = 0.1
	optim = adamw_torch
	schedular = cosine

Table 4: Hyperparameters used in the training of ActiveMem.

Method	ALFWorld	TrivialQA	PopQA	KodCode	BigCodeBench	GPQA	GSM8K	MATH
Vanilla	22.54	32.10	16.08	24.55	40.35	11.62	39.51	36.63
CoT	18.30	28.67	18.39	32.32	38.59	15.67	56.79	45.22
SFT	36.57	63.84	39.20	55.83	37.72	11.11	54.83	38.84
GRPO	43.55	68.21	43.15	62.11	70.34	15.65	68.10	47.42
REINFORCE	43.25	66.50	41.87	60.20	67.80	12.50	67.40	46.89
REINFORCE++	43.66	66.90	44.69	63.33	69.50	13.80	69.04	47.33

Table 5: Results on Qwen2.5-1.5B. All values represent the performance metric for each task (e.g., accuracy %). We highlight the best and second best results.

Figure 5: The generalization study. We train ActiveMem on GSM8K or KodCode and evaluate it on all four datasets

Trained On	Method	AQuA	GPQA	GSM8K	KodCode
Trained On	Vanilla	41.34	11.62	39.51	24.55
AQuA	SFT	42.52	16.67	42.10	18.20
	ExpeL	41.73	12.67	40.16	16.30
	MemGen	43.31	19.70	39.80	19.55
	ActiveMem	47.10	22.51	51.28	24.38
GPQA	SFT	38.55	17.17	45.74	18.50

Table 6: Continual learning results of Qwen2.5-1.5B-Instruct across four reasoning and programming datasets (AQuA, GPQA, GSM8K, KodCode). The model is sequentially trained on each dataset (AQuA

\rightarrow

GPQA

\rightarrow

GSM8K

\rightarrow

KodCode), and after each training stage, evaluation is conducted on all four benchmarks.

Model

2-Objective

8-Objective

16-Objective

\uparrow

\uparrow

Peak (

\times 10^{2}

)

\downarrow

Time (s)

\downarrow

\uparrow

\uparrow

Peak (

\times 10^{2}

)

\downarrow

Time (s)

\downarrow

\uparrow

\uparrow

Peak (

\times 10^{2}

)

\downarrow

Time (s)

\downarrow

Qwen2.5-14B-Inst

0.732

0.902

15.6

\pm

0.19

5.49

\pm

0.16

1.55

1.87

44.7

\pm

0.37

16.2

\pm

0.27

0.567

0.703

38.4

\pm

0.71

29.7

\pm

0.75

Qwen2.5-7B-Inst

0.268

0.366

19.6

\pm

0.33

4.60

\pm

0.08

0.87

1.10

49.5

\pm

0.40

13.9

\pm

0.18

0.165

0.213

43.3

\pm

0.62

15.5

\pm

0.23

Qwen2.5-7B-Inst (A-MEM)

0.286

0.371

14.1

\pm

0.10

24.6

\pm

0.51

1.13

1.43

18.6

\pm

0.10

53.7

\pm

1.26

0.730

0.961

18.8

\pm

0.14

91.2

\pm

2.44

Qwen2.5-7B-Inst (truncate)

0.262

0.336

8.28

\pm

0.06

5.89

\pm

0.16

0.97

1.23

11.8

\pm

0.10

11.9

\pm

0.20

0.396

0.497

13.3

\pm

0.16

22.1

\pm

0.60

Search-R1

0.452

0.531

13.0

\pm

0.08

4.09

\pm

0.23

0.064

0.08

24.7

\pm

0.19

4.25

\pm

0.16

0.009

0.011

20.9

\pm

0.03

4.75

\pm

0.18

Table 7: Quantitative comparison of models on multi-objective, multi-hop QA tasks. Arrows indicate the optimal direction for each metric. Underlined entries denote catastrophic model collapse (i.e., extremely low performance). Variants marked with (truncate) adopt ActiveMem’s prompt and rollout pipeline with context truncation. Variants labeled with (A-MEM) employ the prompt, rollout pipeline, and the external memory module from ActiveMem ( Xu et al. 2026 ) . ActiveMem-QA refers to our proposed ActiveMem optimized specifically on the 2-objective QA.

Model	Avg Final Reward $\uparrow$	Peak Token ( $\times 10^{3}$ ) $\downarrow$	Dependency ( $\times 10^{6}$ ) $\downarrow$	Inference Time Per Traj (s) $\downarrow$
GPT-4o	25.48	$5.30\pm 1.23$	$3.99\pm 1.16$	N/A
GPT-4o (truncate)	13.82	$0.99\pm 0.99$	$0.81\pm 0.23$	N/A
GPT-4o (A-MEM)	24.50	$1.84\pm 0.06$	$0.31\pm 0.11$	N/A
Qwen2.5-7B-Instruct	18.42	$5.64\pm 1.34$	$3.38\pm 0.89$	$12.31\pm 1.82$
Qwen2.5-14B-Instruct	12.34	$5.44\pm 0.92$	$3.30\pm 0.61$	$18.17\pm 2.32$
Agent-FLAN-7B	40.35	$3.37\pm 1.12$	$2.18\pm 1.62$	$9.95\pm 6.19$

Table 8: Quantitative evaluation on the WebShop navigation benchmark. Inference latency for proprietary GPT models is excluded to ensure equitable comparison, as API-dependent metrics are not directly comparable to local execution. Performance scores for Agent-R are sourced directly from the original publication due to the unavailability of its model weights. The -WebShop suffix designates model variants explicitly fine-tuned on the WebShop domain.

Environment	System	EM $\uparrow$	F1 $\uparrow$	Peak Token ( $\times 10^{2}$ ) $\downarrow$	Dependency ( $\times 10^{5}$ ) $\downarrow$	Inference Time $\downarrow$
Wiki RAG	Qwen2.5-7B-Inst (truncate)	0.287	0.382	$6.28\pm 0.05$	$1.65\pm 0.04$	$2.26\pm 0.04$
	Qwen2.5-7B-Inst (A-MEM)	0.246	0.373	$8.47\pm 0.12$	$0.92\pm 0.03$	$11.2\pm 0.40$
	Qwen2.5-7B-Inst	0.269	0.390	$9.32\pm 0.19$	$1.17\pm 0.04$	$2.31\pm 0.04$
	Qwen2.5-14B-Inst	0.422	0.534	$8.89\pm 0.21$	$2.22\pm 0.10$	$6.73\pm 0.24$
	Search-R1	0.445	0.516	$11.0\pm 0.25$	$1.50\pm 0.05$	$\mathbf{2.23\pm 0.14}$
	DeepResearcher	0.419	0.503	$13.3\pm 0.34$	$7.04\pm 0.33$	$3.86\pm 0.09$

Table 9: Performance comparison across diverse environments on single-objective tasks. Arrows denote the optimal direction for each metric. The (SFT) notation indicates variants trained via Supervised Fine-Tuning while adopting ActiveMem’s prompt templates and rollout pipeline. Notably, baseline models undergo task-specific optimization for their respective domains: DeepResearcher is exclusively tailored for the Online Web-QA task targeting the F1 score, whereas Search-R1 is specifically optimized for the Wiki-RAG task using Exact Match (EM) as the training signal.

Agent	BrowseComp	BrowseComp-ZH	WideSearch	GAIA
Proprietary Agents
Claude-4-Sonnet ( Anthropic 2025 )	14.7	22.5	62.0	68.3
Claude-4-Opus ( Anthropic 2025 )	18.8	37.4	-	-
OpenAI-o4-mini ( OpenAI 2025b )	28.3	44.3	-	-
OpenAI-o3 ( OpenAI 2025b )	49.7	58.1	60.0	70.5
OpenAI Deep Research ( OpenAI 2025a )	51.5	42.9	-	67.4

Table 10: Main results. ActiveMem-30B-A3B achieves remarkable performance, surpassing open-source agents with much larger model size such as AgentFold-30B-A3B, DeepSeek-V3.1-671B-A37B and matching proprietary agents such as OpenAI-o4-mini, indicating the potential of this new paradigm.

LLM Agent Memory

Large language model agent memory lets agents retain, organize, retrieve, and update information across interactions and long-running tasks.

See all 357 papers

ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning
Jun 9, 2026Yunhan Jiang, Wenbin Duan, Shasha Guo +3Long-Horizon LLM Agents Agent Memory Management
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences
Weighted Memory Tree: Remembering What Matters for Long-Horizon LLM Agents
Aug 21, 2026Quang Dao, Purvi Kathalkar, Kenneth EatonLLM Agent Memory Memory Management for LLM Agents
Rose-Hulman Institute of Technology · Georgia Institute of Technology · Georgia Tech Research Institute
AdMem: Advanced Memory for Task-solving Agents
Jun 5, 2026Runzhe Wang, Huilin Lu, Shengjie Liu +2Continual Learning for LLM Agents LLM Agent Memory
Princeton University · Amazon · Arm

Jun 9, 2026J/K move · Enter open · S save

ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning

Yunhan Jiang, Wenbin Duan, Shasha Guo, Liang Pang+2

State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences

Long-Horizon LLM Agents Agent Memory Management Memory-Augmented Language Model Agents

Open paper AskPDF arXiv

Memory is essential for enabling large language model (LLM) agents to handle long-horizon reasoning tasks. Existing memory mechanisms are largely centralized, typically organizing retrieved information and interaction history within a single model context. This design imposes a fundamental trade-off: scaling reasoning trajectories risks context overload, whereas aggressive content pruning may result in irreversible information loss. Seeking a better trade-off, we draw inspiration from human cognitive systems, especially the functional complementarity between the prefrontal cortex (executive control) and the hippocampus (memory management), suggesting that such a trade-off need not be inherent, but may instead stem from centralized memory organization. To this end, we propose ActiveMem, a heterogeneous framework that decouples agent memory from the core reasoning process. Specifically, a high-level Planner utilizes distilled semantic gists to execute reasoning, while a lightweight, distributed memory system operates in parallel to actively accumulate and consolidate these gists throughout the task. Experiments on BrowseComp-Plus and GAIA show that ActiveMem achieves state-of-the-art accuracy with significantly reduced overhead, demonstrating the effectiveness of distributed active memory for long-horizon reasoning.

Abstract

LLM Agent Memory

Explore similar work

ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning

Weighted Memory Tree: Remembering What Matters for Long-Horizon LLM Agents

AdMem: Advanced Memory for Task-solving Agents