Pulse: Long-Horizon Tasks Need Streaming Personalization

Qijia Zhuang; Jialiang Liang; Yuxiang Ren

cs.AIOct 31, 2026

Pulse: Long-Horizon Tasks Need Streaming Personalization

Authors: Qijia Zhuang, Jialiang Liang, Yuxiang Ren

Organizations: Nanjing University · Xinjiang University

Abstract

Large language model agents increasingly undertake complex, long-horizon tasks that benefit from sustained personalized support through long-term memory. Existing memory-augmented methods retrieve and use memory primarily from the initial request, with limited awareness of ongoing execution and emerging personalization needs. Agents may therefore overlook user preferences and project constraints during multistep tasks. We propose Pulse, a streaming personalization method for long-horizon tasks. An independent, lightweight memory agent continuously monitors execution, identifies emerging personalization needs and potential preference conflicts, and uses user and project memories to provide targeted interventions at appropriate stages. Reinforcement learning optimizes when and how the memory agent intervenes, improving the timeliness and accuracy of personalized support. Pulse scores 71.9% on $π$ -Bench and 46.4% on MCP-Persona, achieving the highest overall scores among the compared methods on both datasets while improving task execution efficiency.

Figures & tables

Figure 1. Hotel booking with request-level memory recall (top) and Pulse (bottom). The traditional workflow relies on a user reminder; Pulse prompts wake-up service verification or hotel reselection before submission. Vertically stacked hotel-booking workflows with the shared query and memories combined in one input panel above. The upper workflow selects a hotel with wake-up calls unconfirmed and revises it after a user reminder. The lower Pulse workflow prompts verification or reselection before submission.

Figure 2. Overview of Pulse. Left: historical interactions and task traces form a user Topic graph and project memory containing workflows $\mathcal{W}$ and current information $\mathcal{S}$ . Center: Pulse augments the request with initial evidence $\mathcal{E}_{0}$ , incrementally observes messages, candidate calls, and tool results, gates retrieval, and selects silence, supplementation, or pre-commit intervention. Right: GRPO optimizes the execution-time policy using action correctness and an incorrect-intervention penalty. Three-part Pulse architecture. Historical conversations and task traces populate user Topic graphs and project workflows and state. One Pulse agent first personalizes the request, then listens to exposed execution events. A retrieval gate either reuses valid context or queries memory and selects evidence. Behavior selection yields Silent, Supplement for the next model call, or Intervene before a candidate tool call executes. GRPO trains the execution-time policy from a base model using group samples, a reward of plus one for a correct action, minus one for an incorrect intervention, and zero for other errors, with a frozen reference policy for KL regularization.

Group	Method	User Domain					Overall	Token Usage
		Researcher	Marketer	Pharmacist	Law Trainee	Financier	COMP	Store Token (k)	Retrieve Token (k)	Time (s)
Memory	Mem0	51.3	53.0	53.6	46.6	50.9	50.5	5.6	3.7	242.1
	LightMem	50.6	40.9	61.8	47.7	53.8	51.6	4.8	3.1	277.3
	A-MEM	45.5	48.5	48.2	48.3	53.8	49.0	8.3	6.1	243.2
Harness	OpenClaw	56.4	42.4	88.2	60.9	63.9	63.3	5.5	5.0	275.9
	dsh	46.8	47.0	67.3	70.1	56.8	58.7	6.4	5.9	264.0

Table 1.

\pi

-Bench results. COMP is reported as a percentage; Store and Retrieve are thousands of tokens per task. Red and blue mark the best and second-best values, respectively. Time is seconds per task.

Group	Method	Single-Server				Cross-Server			Overall	Token Usage
		CP	CM	SM	EM	Lark	Rednote	Hodgepodge	Acc	Store Token (k)	Retrieve Token (k)	Time (s)
Memory	Mem0	32.6	18.0	33.1	43.1	28.6	25.5	6.3	29.5	8.0	3.4	181.1
	LightMem	35.2	20.7	35.5	45.1	31.6	27.7	18.8	32.2	3.8	1.9	139.8
	A-MEM	27.9	14.7	28.5	39.2	23.6	21.2	6.3	25.0	5.3	4.1	164.6
Harness	OpenClaw	37.8	19.3	37.8	47.1	34.3	30.4	18.8	34.3	4.9	3.1	157.8
	dsh	40.4	21.3	40.1	46.1	37.0	32.6	18.8	36.5	5.3	3.7	147.3

Table 2. MCP-Persona results. Acc is reported as a percentage; Store and Retrieve are thousands of tokens per task. Red and blue mark the best and second-best values, respectively. Time is seconds per task.

Group	Variant	$\pi$ -Bench				MCP-Persona
		COMP (%)	Store (k)	Retrieve (k)	Time (s)	Acc (%)	Store (k)	Retrieve (k)	Time (s)
Memory	Flat Memory	49.9	7.4	4.3	228.4	27.9	7.3	3.1	143.2
	Flat + User	53.6 $+3.7$	6.7 $-0.7$	4.2 $-0.1$	248.5 $+20.1$	30.2 $+2.3$	6.8 $-0.6$	2.9 $-0.2$	158.6 $+15.4$
	Flat + Project	54.4 $+4.4$	6.5 $-0.9$	4.1 $-0.2$	245.7 $+17.3$	30.2 $+2.3$	6.6 $-0.7$	2.8 $-0.3$	154.3 $+11.1$
	Flat Memory + Pulse w/o RL	61.2 $+11.3$	7.4	4.6 $+0.3$	202.7 $-25.7$	35.9 $+8.0$	7.3	3.2 $+0.1$	124.4 $-18.8$
Retrieval	Start-only	56.4	5.3	3.5	215.7	33.1	5.3	2.2	110.4

Table 3. Component ablations. Store and Retrieve are k tokens per task; Time is seconds per task.

Figure 3. Effects of the incorrect-intervention reward. The

-1

penalty achieves higher interruption success,

\pi

-Bench COMP, and MCP-Persona Acc than reward

0

. The first three groups use the left axis (%); the last two use the right axis (s/task). Grouped bar chart comparing No RL, RL with reward zero for an incorrect interruption, and RL with reward minus one. Interruption success is 75.1, 68.4, and 81.2 percent, with 494, 557, and 346 total interruptions, respectively; pi-Bench COMP is 66.5, 69.2, and 71.9 percent; MCP-Persona accuracy is 41.1, 43.8, and 46.4 percent. The first three groups use the left axis from 0 to 100 percent. The last two show time on the right axis from 0 to 300 seconds per task: No RL, pi-Bench 233.9 and MCP-Persona 128.7; reward 0, 247.1 and 137.4; reward minus one, 203.8 and 118.3.

Figure 4. MCP-Persona accuracy across early, middle, and late stages divided by execution step. Pulse (Full) leads at every stage and retains 44.2% accuracy in the late stage. Line chart of MCP-Persona accuracy in early, middle, and late stages divided by execution step. Mem0: 35.8, 29.2, 23.6 percent. ZCode: 38.2, 35.4, 33.8 percent. Pulse without RL: 43.3, 41.2, 38.7 percent. Full Pulse: 48.0, 46.9, 44.2 percent. The vertical axis spans 20 to 50 percent.

Figure 5. Policy choices in MCP-Persona 109. A habitual Notion preference triggers an unnecessary pause on the left. The right branch selects Silent and continues preparing the requested Obsidian note. Six interactions compare two decisions at the same proposed Obsidian append. The left branch treats a frequent-use preference for Notion as an exclusive constraint and pauses a valid write. The right branch recognizes that the memory does not conflict with the task, selects Silent, and continues attendee retrieval, calendar lookup, route calculation, and note updates.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

		Checkpoint completeness (%)					Overall	Cost per task
Group	Variant	Researcher	Marketer	Pharmacist	Law Trainee	Financier	COMP	Store (k)	Retrieve (k)	Time (s)
Memory	Flat Memory	50.0% [1pt] 78/156	48.5% [1pt] 32/66	51.8% [1pt] 57/110	47.7% [1pt] 83/174	51.5% [1pt] 87/169	49.9% [1pt] 337/675	7.4	4.3	228.4
	Flat + User	52.6% [1pt] 82/156	51.5% [1pt] 34/66	58.2% [1pt] 64/110	50.6% [1pt] 88/174	55.6% [1pt] 94/169	53.6% [1pt] 362/675	6.7	4.2	248.5
	Flat + Project	53.2% [1pt] 83/156	51.5% [1pt] 34/66	60.9% [1pt] 67/110	52.3% [1pt] 91/174	54.4% [1pt] 92/169	54.4% [1pt] 367/675	6.5	4.1	245.7
	Flat Memory + Pulse w/o RL	57.1% [1pt] 89/156	53.0% [1pt] 35/66	66.4% [1pt] 73/110	63.8% [1pt] 111/174	62.1% [1pt] 105/169	61.2% [1pt] 413/675	7.4	4.6	202.7
Retrieval	Start-only	54.5% [1pt] 85/156	53.0% [1pt] 35/66	62.7% [1pt] 69/110	54.0% [1pt] 94/174	58.0% [1pt] 98/169	56.4% [1pt] 381/675	5.3	3.5	215.7

Appendix

Table B1.

\pi

-Bench per-category ablations. Percentages appear above completed/total checkpoints; Store and Retrieve are thousands of tokens per task. Time is seconds per task.

Checkpoint accuracy (%)

Overall

Cost per task

Group

Variant

Lark

Rednote

Hodge- podge

Acc

Store (k)

Retrieve (k)

Time (s)

Memory

Flat Memory

31.5% [1pt] 60.5/192

15.3% [1pt] 11.5/75

30.2% [1pt] 26/86

39.2% [1pt] 20/51

27.5% [1pt] 60.5/220

25.0% [1pt] 23/92

6.3% [1pt] 0.5/8

27.9% [1pt] 202/724

7.3

3.1

143.2

Flat + User

33.9% [1pt] 65/192

18.0% [1pt] 13.5/75

33.7% [1pt] 29/86

43.1% [1pt] 22/51

29.1% [1pt] 64/220

26.1% [1pt] 24/92

12.5% [1pt] 1/8

30.2% [1pt] 218.5/724

6.8

2.9

158.6

Flat + Project

32.8% [1pt] 63/192

19.3% [1pt] 14.5/75

32.0% [1pt] 27.5/86

44.1% [1pt] 22.5/51

30.0% [1pt] 66/220

25.5% [1pt] 23.5/92

18.8% [1pt] 1.5/8

30.2% [1pt] 218.5/724

6.6

2.8

154.3

Flat Memory + Pulse w/o RL

38.5% [1pt] 74/192

22.0% [1pt] 16.5/75

38.4% [1pt] 33/86

50.0% [1pt] 25.5/51

36.1% [1pt] 79.5/220

32.1% [1pt] 29.5/92

25.0% [1pt] 2/8

35.9% [1pt] 260/724

7.3

3.2

124.4

Retrieval

Start-only

37.8% [1pt] 72.5/192

20.7% [1pt] 15.5/75

33.7% [1pt] 29/86

49.0% [1pt] 25/51

32.5% [1pt] 71.5/220

26.6% [1pt] 24.5/92

18.8% [1pt] 1.5/8

33.1% [1pt] 239.5/724

5.3

2.2

110.4

Appendix

Table B2. MCP-Persona per-category ablations. Percentages appear above completed/total checkpoints; Store and Retrieve are thousands of tokens per task. Time is seconds per task.

Memory systems
Method	Performance (%)		Cost per task
	Web EM	Pref R	Store (k)	Retrieve (k)	Time (s)
Mem0	59.6	69.6	3.9	2.9	69.3
LightMem	34.7	54.2	1.7	1.8	62.1
A-MEM	40.4	54.2	6.3	3.7	72.9
Agent harnesses

Appendix

Table B3. Persona2Web supplementary results. Web EM and Pref R are percentages; Store and Retrieve are thousands of tokens per task. Time is seconds per task.

$\pi$ -Bench: COMP
Method	Performance	Cost per task
	(%)	Retrieve (k)	Time (s)
Mem0	56.4	4.7	257.0
dsh	63.7	4.5	229.2
Pulse w/o RL	66.5	3.6	233.9
Pulse	71.9	3.9	203.8

Appendix

Table B4. Shared Pulse memory backend. Performance is in percent; Retrieve is k tokens per task, and Time is seconds per task.

Long-Horizon LLM Agents

Long-horizon large language model agents plan, use tools, and preserve task state across extended, multi-step interactions.

See all 471 papers

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents
May 25, 2026Yeonjun In, Wonjoong Kim, Sangwu Park +2Memorization in Language Models Persistent Memory for Language Models
KAIST
Who Schedules Agent Memory? A Sidecar Approach to Long-Horizon Tasks
Oct 31, 2026Qijia Zhuang, Rui Liu, Yuxiang RenLLM Agent Memory Memory Management for LLM Agents
Nanjing University · Nanjing University, Nanjing, China · Northeastern University +1
TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation
Apr 9, 2026Xinliang Frederick Zhang, Lu WangPersonalized Language Models Memory-Augmented Language Models
Computer Science and Engineering, University of Michigan, Ann Arbor, MI

May 25, 2026J/K move · Enter open · S save

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents

Yeonjun In, Wonjoong Kim, Sangwu Park, Kanghoon Yoon+1

KAIST

Memorization in Language Models Persistent Memory for Language Models Memory Agent Benchmarks Long-Horizon LLM Agents Agent Memory Management

Open paper AskPDF arXiv

Existing large language model (LLM) based memory systems apply universal, static policies that overlook a fundamental reality: the contexts that are worth storing in memory are different across users. This misalignment wastes limited memory budget on transient interactions while failing to preserve critical context for long horizon tasks. To address this gap, we investigate an underexplored question: can LLM based memory systems learn personalized memory policies? We introduce PerMemBench, the first benchmark for evaluating personalized memory systems, featuring multi year, multi domain interaction histories across diverse user personas. We further present the first empirical study of memory personalization, proposing session level storage gating, a lightweight framework that selectively bypasses memory operations for transient sessions. Our study confirms that personalization yields substantial retention gains under perfect gating, yet reveals that accurate gating remains an open and critical challenge.

Abstract

Long-Horizon LLM Agents

Explore similar work

Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents

Who Schedules Agent Memory? A Sidecar Approach to Long-Horizon Tasks

TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation