Proactive Memory for Event-Driven LLM Agents

Jinyang Li; Zichuan Fu; Xian Wu; Wenlin Zhang; Yimin Deng; Kaifeng Guo; Guojing Li; Junjia QI

cs.AIOct 27, 2026

Proactive Memory for Event-Driven LLM Agents

Authors: Jinyang Li, Zichuan Fu, Xian Wu, Wenlin Zhang, Yimin Deng, Kaifeng Guo, Guojing Li, Junjia QI,

Organizations: City University of Hong Kong · Tencent

Abstract

Proactive LLM agents anticipate service opportunities from ongoing observations and accumulated history. Existing approaches emphasize recent context or retrieval of related memories, providing limited support for validating recurring behavioral regularities and assessing their applicability to the current event. We propose Proactive Memory (ProactMem), which organizes history into recent events, temporal trajectories, and behavioral regularities. Trajectory facts are checked against source events, and regularities are assessed through historical replay before being matched to current conditions. We also construct ProactiveStream, a longitudinal benchmark covering three application scopes, where agents process isolated event streams chronologically under episode-aware opportunity labels. On ProactiveStream, ProactMem achieves an overall effectiveness (F1) of 52.6%, a 21.4% relative improvement over the strongest of five external baselines. Code is available at: https://anonymous.4open.science/r/ProactMem-5351/.

Figures & tables

Figure 1: Illustrative Sports scenario: ProactMem uses a supported evening-walk regularity beyond recent events.

Panel of Figure 2 — Figure 2: ProactMem architecture. (A) Source validation and replay build memory. (B) Retrieved memory and current conditions inform assessment and intervention control. Events are stored after the decision (dashed arrows).

Scope	Primary Source	Entity	Streams	Input Events	Temporal Coverage	Reference Interventions
Sports	HeartSteps	participant	9	1,456	367 observed participant-days	365
Home	CASAS	home	4	1,624	14 consecutive days/home	190
Code	GH Archive	repository	4	1,459	longitudinal repository histories	174
Total	three scopes	isolated stream	17	4,539	three longitudinal settings	729

Table 1: Composition and provenance of ProactiveStream. Reference interventions mark service opportunities.

Scope	Method	Frequency ( $\times$ )	Relevance $\uparrow$	Coverage $\uparrow$	Effectiveness $\uparrow$	Intrusion $\downarrow$	Balance $\uparrow$
Sports	ProactiveAgent	2.75	30.0	82.5	43.9	64.5	59.0
	ContextAgent	2.16	31.0	67.1	42.4	50.0	58.6
	Mem0	3.40	28.8	97.8	44.5	81.0	58.4
	A-MEM	2.41	30.6	74.0	43.3	56.0	59.0
	Graphiti	2.20	29.3	64.4	40.3	52.0	56.2
	ProactMem	2.15	35.8	76.7	48.8	46.1	65.3

Table 2: Full-scope comparison (GPT-5.6 Luna, %; frequency in

\times

). Bold/underline: best/second-best per scope.

Model	Sports		Home		Code
Model	Effectiveness $\uparrow$	Intrusion $\downarrow$	Effectiveness $\uparrow$	Intrusion $\downarrow$	Effectiveness $\uparrow$	Intrusion $\downarrow$
GPT-5.6 Luna	48.8	46.1	58.6	10.3	55.4	18.1
GPT-5.6 Terra	54.1	39.4	64.4	9.1	60.3	14.9
GPT-5.6 Sol	58.6	34.4	69.3	7.5	66.4	11.8

Table 3: Within-series model scaling (%). Bold and underline mark the best and second-best scores per column.

Panel of Figure 3 — Figure 3: Event-only comparison of four backbones on selected streams (0–100%). Specificity $=1-\mathrm{intrusion}$ .

Panel of Figure 4 — Figure 4: Memory-layer ablation (GPT-5.6 Luna) under a shared decision prompt and intervention control policy.

	Sports		Home		Code
Method	Relevance	Effectiveness	Relevance	Effectiveness	Relevance	Effectiveness
w/o replay gate	31.1	42.7	43.6	54.4	33.8	48.1
w/o trajectory validation	34.4	47.5	45.9	56.7	36.1	51.5
before suppression	–	–	–	–	36.7	53.6
ProactMem	35.8	48.8	48.6	58.6	40.1	55.4

Table 4: Validation and Code intervention control (GPT-5.6 Luna, %). Each w/o row drops one check; before suppression scores recorded proposals with later states fixed. Bold and underline mark the best and second-best.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Interface and Settings
ProactiveAgent	Released prompt and proposal/null contract on the study backbone; causal event–decision history capped at 1,000 events per request including the current event ( Lu et al., 2025 ; THUNLP, 2026 ) .
ContextAgent	In-context proactive-service judgment on public events ( Yang et al., 2025 ) .
Mem0	Open-source mem0ai==2.0.12 extraction, consolidation, and retrieval; local all-MiniLM-L6-v2 embeddings; retrieval of up to eight memories for the same entity with search threshold $0.1$ ( Chhikara et al., 2025 ; Mem0, 2026 ) .
A-MEM	Linked notes with association updates augment bounded context ( Xu et al., 2025 ) .
Graphiti	Graphiti 0.30.2 with Neo4j and local 384-dimensional all-MiniLM-L6-v2 embeddings; committed events populate separate graphs for each stream. Queries from the current event use native hybrid edge search with reciprocal-rank fusion, retrieving up to ten relation facts with validity fields and source IDs ( Zep, 2026 ) .

Appendix

Table 5: External-method memory interfaces and implementation settings; all methods share the backbone.

Scope	Source Stream	Events
Sports	HeartSteps participant 05	184
	HeartSteps participant 35	191
Home	CASAS home 0004	406
	CASAS home 0003	406
Code	Unitech/pm2	365
	thephpleague/commonmark	347

Appendix

Table 6: Streams in the cross-family comparison.

Scope	Method	Frequency ( $\times$ )	Coverage $\uparrow$
Sports	w/o replay gate	2.19	68.2
	w/o trajectory validation	2.23	76.7
	ProactMem	2.15	76.7
Home	w/o replay gate	1.65	72.1
	w/o trajectory validation	1.62	74.2
	ProactMem	1.52	73.7

Appendix

Table 7: Additional validation and Code intervention control scores. Frequency uses multiples (

\times

); coverage uses percentages. Bold and underline mark the best and second-best coverage per scope, including ties.

Scope	Positives	$\boldsymbol{I}$	TP	FP	FN	TN	Earlier Diagnostic Effectiveness	Separate-Team Effectiveness
Sports	27	86	24	62	3	61	42.6	42.5
Home	15	39	15	24	0	111	68.7	55.6
Code	34	42	18	24	16	92	57.6	47.4

Appendix

Table 8: Reference sensitivity on 150 targets per scope. Effectiveness (%) scores identical outputs under earlier and separate-team references. Positives and confusion counts use the latter;

I

counts final interventions.

Scope	Covered/New (Coverage)	$\boldsymbol{I}$	$\boldsymbol{F}$	$\boldsymbol{V}$	$\boldsymbol{R}$	Unmatched	Late	Unresolved	SR	RIR
Sports	23/26 (88.5)	86	23	1	1	60	0	1	27.9	1.2
Home	9/9 (100.0)	39	10	6	17	6	0	0	41.0	43.6
Code	14/31 (45.2)	42	14	3	4	16	4	1	40.5	9.5

Appendix

Table 9: ProactMem episode diagnostic on 150 targets per scope. Coverage, SR, and RIR use percentages. Unresolved denotes ambiguous semantic or condition matches; uncertain follow-ups (

U

) are zero in all scopes.

LLM Agent Memory

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

See all 376 papers

ProMemAgent: Long-Horizon Proactive Assistance with Context-Aware Prospective Memory from Sensory Perceptions
Oct 19, 2026Lilin Xu, Linyang He, Kaiyuan Hou +4Proactive Agents Long-Horizon Agent Tasks
Columbia University
Beyond Static Summarization: Proactive Memory Extraction for LLM Agents
Jan 8, 2026Chengyuan Yang, Zequn Sun, Wei Wei +1LLM Agent Memory Persistent Memory for Language Models
State Key Laboratory for Novel Software Technology, Nanjing University, China · National Institute of Healthcare Data Science, Nanjing University, China
ProEvent: An Event-centric Benchmark for Proactive Agents
Jul 20, 2026Guanzhen Li, Liangming Pan, Leye WangProactive Agents LLM Agent Evaluation
School of Computing, Peking University

Oct 19, 2026J/K move · Enter open · S save

ProMemAgent: Long-Horizon Proactive Assistance with Context-Aware Prospective Memory from Sensory Perceptions

Lilin Xu, Linyang He, Kaiyuan Hou, Yuang Fan+3

Columbia University

Proactive Agents Long-Horizon Agent Tasks Agent Memory Long-Horizon Agent Evaluation

Open paper AskPDF arXiv

Large language model (LLM) agents are moving beyond reactive responses toward proactive assistance that anticipates user needs without explicit requests. However, existing proactive agents primarily address immediate needs, while many everyday situations involve delayed intentions that can only be acted on hours or days later. Humans handle such intentions through prospective memory, the ability to remember to act on an intention when the right moment arises. In this paper, we introduce ProMemAgent, the first proactive agent to incorporate prospective memory over everyday sensory contexts for long-horizon proactive assistance. ProMemAgent organizes past sensory perceptions (e.g., egocentric video and audio) into context-aware memory, generating prospective cues and linking earlier events with subsequent developments for future retrieval. Given the current context, it retrieves relevant past events and performs memory-augmented proactive reasoning to determine whether assistance is needed. We further introduce CPMBench, the first benchmark for evaluating long-horizon proactive assistance over sensory contexts, comprising 1,000 instances across eight everyday categories. Extensive experiments show that ProMemAgent achieves the best memory retrieval performance and consistently outperforms all baselines on long-horizon proactive assistance.

Abstract

LLM Agent Memory

Explore similar work

ProMemAgent: Long-Horizon Proactive Assistance with Context-Aware Prospective Memory from Sensory Perceptions

Beyond Static Summarization: Proactive Memory Extraction for LLM Agents

ProEvent: An Event-centric Benchmark for Proactive Agents