Proactivity is a core expectation for AGI. Prior work remains largely confined to laboratory settings, leaving a clear gap in real-world proactive agent: depth, complexity, ambiguity, precision and real-time constraints. We study this setting, where useful intervention requires inferring latent needs from ongoing context and grounding actions in evolving user memory under latency and long-horizon constraints. We first propose DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System) as a general paradigm for streaming proactive AI agent. We instantiate this paradigm in Pask, with streaming IntentFlow model for DD, a hybrid memory (workspace, user, global) for long-term MM, PAS infra framework and introduce how these components form a closed loop. We also introduce LatentNeeds-Bench, a real-world benchmark built from user-consented data and refined through thousands of rounds of human editing. Experiments show that IntentFlow matches leading Gemini3-Flash models under latency constraints, while identifying deeper user intent.
Figures & tables
Figure 2: Illustration of the proposed DD-MM-PAS, a potentially general paradigm for proactive AI, alongside its diverse application instances. The framework is composed of three interconnected modules: Demand Detecting (DD) acts as the proactive engine to predict latent human needs from real-time signals; the Memory Module (MM) accumulates persistent, long-term data to enable deep, person-level understanding; and the Proactive Agent System (PAS) serves as the system backbone, managing real-time execution, information fusion, and memory scheduling. Together, these components transition AI from a passive, query-based responder into an active initiator of assistance across professional, academic, and daily contexts.
Figure 3: IntentFlow model architecture for streaming information–based human deep intent reasoning and demand detection. The system processes textual information fragments from the frontend and predicts one control token per step: <silent> (wait for the next message), <fast intervention> (low-latency understanding help, e.g., concept explanation, typically <1 s), or <full assistance> (invoke memory modules and global context for deep intent detection.
Figure 4: Different DD outcomes trigger different system modes: <silent> avoids unnecessary interruption, <fast intervention> provides immediate help, and <full assistance> enables agentic operations for higher-quality support.
Figure 5: Overview of the LatentNeeds dataset and benchmark composition, showing the distribution across three major domains (meetings, learning, and daily life) and their corresponding subcategories. The middle part summarizes the data construction pipeline, including public-data grounding, multi-agent information-stream generation, iterative human-intent generation, structured reformatting, and AI-based post-filtering, together with the collection and refinement of real-world sessions. The bottom part outlines the training strategy: separate supervised fine-tuning of the Demand Detector and MemLoader, followed by joint reinforcement learning for deeper human-intent alignment.
Figure 6: Overview of the internal architecture of Pask-MM . Inspired by hierarchical memory designs in modern agent systems, the framework organizes memory into three functional components with different access frequencies: user memory Muser , workspace memory Mwsp , and global memory Mglobal . Muser stores stable user profiles for personalization, Mwsp maintains short-term session dynamics for real-time inference, and Mglobal records long-term episodic knowledge for retrieval and validation. During interaction, workspace memory is updated continuously to support low-latency reasoning, while Muser and Mglobal evolve only after session completion through offline maintenance and compression.
Figure 7: System architecture of Pask-PAS , illustrating how frontend devices, server infrastructure, and AI backends form an always-on loop for perception, understanding, and proactive action.
Figure 8: Balanced accuracy as a function of conversation depth. Turns are grouped into non-overlapping buckets of 4 consecutive turns (1–4, 5–8, …, 57–60). (a) Per-bucket accuracy: each point is the balanced average of demand recall and non-demand precision within that bucket alone, revealing local fluctuations. (b) Cumulative accuracy: each point aggregates all turns from the start up to and including the current bucket, yielding a smoothed view of overall degradation. IntentFlow maintains > 80% balanced accuracy across all buckets, while the strongest baseline (Gemini-3-Flash) drops from 85.6 to 70.8 ( Δ =–17.3%).
Model
Type
Work
Learning
Daily
Overall
Business
Product
Tech
Work
STEM
Program.
Human.
Personal
Tools &
Content
Metrics
Strategy
Engineer.
Collab.
Lecture
Tutorial
Business
Life
Workflow
Knowl.
GPT-5-Mini
Demand.
71.1
69.1
73.2
65.8
78.3
60.3
69.5
63.4
67.3
46.5
66.5
No-Dem.
88.8
92.5
83.7
76.8
79.2
91.2
91.7
91.2
93.8
90.6
88.0
Avg.
79.9
80.8
78.5
71.3
78.8
75.8
80.6
77.3
80.5
68.5
77.2
GPT-5-Nano
Demand.
80.5
87.4
88.4
77.2
58.0
53.4
51.6
77.5
74.4
63.1
71.2
Table 1: Main results on the IntentFlow Proactive Demand Detection Benchmark. Each cell reports the turn-level accuracy score (0–100). Demand = accuracy on demand turns; No-Dem. = accuracy on non-demand turns; Avg. = balanced average (1:1 demand/non-demand). Columns are grouped by domain: Work (Business Metrics, Product Strategy, Tech Engineer., Work Collab.), Learning (STEM Lecture, Program. Tutorial, Human. Business), Daily (Personal Life, Tools & Workflow, Content & Knowl.). Bold = best per column; Italic = second best.
Model
Demand
Work
Learning
Daily
Overall
Business
Product
Tech
Work
STEM
Program.
Human.
Personal
Tools &
Content
Metrics
Strategy
Engineer.
Collab.
Lecture
Tutorial
Business
Life
Workflow
Knowl.
GPT-5-Mini
Req.
72.9
71.6
71.0
66.9
100.0
49.1
33.3
60.7
67.8
47.8
64.1
Ins.
65.9
60.3
83.3
63.5
75.6
70.5
70.7
65.1
64.0
46.4
66.5
Avg.
69.4
65.9
77.2
65.2
87.8
59.8
52.0
62.9
65.9
47.1
65.3
GPT-5-Nano
Req.
80.5
87.0
88.9
76.2
68.8
50.9
33.3
80.4
77.6
60.9
70.5
Table 2: Performance by demand type on the IntentFlow benchmark. Each cell reports the turn-level accuracy score (0–100) on demand turns only. Req. = Requirement-type demands (decision support, task planning, problem solving, summarization, information lookup); Ins. = Insight-type demands (risk warning, knowledge gap, callback reminder, context synthesis, trend insight, sentiment analysis). Bold = best per column; Italic = second best.
Model
Type
Turn Position (bucket)
Δ (%)
1-4
5-8
9-12
13-16
17-20
21-24
25-28
29-32
33-36
37-40
41-44
45-48
49-52
53-56
57-60
GPT-5-Mini
Demand.
80.3
76.6
74.7
68.4
69.3
67.2
67.7
64.5
60.2
67.0
65.4
59.4
60.0
50.9
51.0
-36.5
No-Dem.
93.1
92.7
86.3
87.9
82.8
90.5
84.2
85.5
89.1
87.2
83.6
91.4
86.7
95.6
89.6
-3.8
Avg.
86.7
84.6
80.5
78.1
76.0
78.9
76.0
75.0
74.7
77.1
74.5
75.4
73.4
73.2
70.3
-19.0
GPT-5-Nano
Demand.
88.6
78.6
82.0
80.9
67.5
69.0
66.1
69.1
71.0
75.5
66.7
76.8
70.8
75.4
60.8
-31.4
No-Dem.
75.2
68.8
77.2
73.1
71.1
71.6
71.1
73.1
69.6
67.5
70.9
68.8
74.7
72.1
77.1
2.5
Table 3: Per-turn-bucket performance under the multi-turn setting. Turn positions are grouped into buckets of 4 consecutive turns. Demand = accuracy on demand turns; No-Dem. = accuracy on non-demand turns; Avg. = balanced average. Δ (%) = relative change from first to last bucket. Bold = best per column; Italic = second best.
Model
Per-Turn Latency (ms)
T =1–12
T =13–24
T =25–36
T =37–48
T =49–60
Dem.
N-D.
Avg.
Dem.
N-D.
Avg.
Dem.
N-D.
Avg.
Dem.
N-D.
Avg.
Dem.
N-D.
Avg.
GPT-5-Mini
10.4k
6.7k
8.1k
10.2k
7.1k
8.4k
9.9k
7.3k
8.4k
9.3k
6.4k
7.6k
8.6k
5.8k
7.1k
GPT-5-Nano
7.4k
6.3k
6.7k
6.9k
5.7k
6.2k
7.0k
6.2k
6.5k
7.2k
6.2k
6.6k
6.3k
5.8k
6.0k
GPT-oss-120b
7.7k
6.3k
6.8k
8.3k
6.6k
7.3k
9.1k
8.4k
8.7k
7.9k
6.8k
7.3k
8.6k
7.4k
8.0k
Gemini-3-Flash
3.6k
3.0k
3.2k
3.9k
3.7k
3.8k
4.3k
4.1k
4.2k
4.3k
4.3k
4.3k
4.4k
4.3k
4.4k
Table 4: Per-turn inference latency by conversation position. Turn positions are grouped into buckets of 12 consecutive turns ( T ). Dem. = average latency on demand turns (model generates a response); N-D. = average latency on non-demand turns (model outputs [NO_DEMAND] ); Avg. = overall average. Demand turns consistently incur higher latency due to longer generated outputs. Bold = fastest; Italic = second fastest.
Figure 9: Long-term memory case study and user study. Top: Case studies illustrating three types of long-term memory in Pask-MM : (1) background memory for understanding user roles and needs, (2) knowledge memory via Mglobal for retrieving past facts, and (3) preference memory for personalization. Bottom-left: User ratings (20 participants per scenario) are mostly around 3–4, with learning rated highest, work moderate, and daily lowest with some scores near 1–2, highlighting increasing difficulty from structured to open-ended settings. Bottom-right: Quiz results (5 participants per group) show that Pask improves average scores from about 6 to 7–7.5 after a 5-minute learning task, demonstrating its effectiveness in knowledge acquisition.
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.
The rise of personal assistant agents, e.g., OpenClaw, highlights the growing potential of large language models to support users across everyday life and work. A core challenge in these settings is proactive assistance, since users often begin with underspecified requests and leave important needs, constraints, or preferences unstated. However, existing benchmarks rarely evaluate whether agents can identify and act on such hidden intents before they are explicitly stated, especially in sustained multi-turn interactions where user needs emerge gradually. To address this gap, we introduce π-Bench, a benchmark for proactive assistance comprising 100 multi-turn tasks across 5 domain-specific user personas. By incorporating hidden user intents, inter-task dependencies, and cross-session continuity, π-Bench evaluates agents' ability to anticipate and address user needs over extended interactions, jointly measuring proactivity and task completion in long-horizon trajectories that better reflect real-world use. Experiments show (1) proactive assistance remains challenging, (2) a clear distinction between task completion and proactivity, and (3) the value of prior interaction for proactive intent resolution in later tasks.
Haoran Zhang, Luxin Xu, Zhilin Wang +11
Shanghai Jiao Tong University · Shanghai AI Laboratory · Fudan University +7
A long-lived LLM agent, such as OpenClaw, earns its value by acting on a user's preferences and constraints across sessions, not just the current request. Yet today's agents keep what a user volunteers but rarely ask for what stays unspoken, leaving a proactivity gap in long-lived LLM agents: an agent cannot act on a preference it never obtained. As users delegate more of their affairs to agents, the impact of this gap grows. We isolate one concrete, controllable slice of this gap as Ask-to-Remember (ATR): the agent decides whether to ask now for a reusable user preference that the current task does not need but a later session with the same user will. ATR is hard even to evaluate: the right question is underdetermined and its payoff deferred to tasks that may never arise. ATRBench, to the best of our knowledge the first ATR benchmark, makes it measurable by fixing each user's preferences as hidden ground truth, so success demands asking, not recall. Across eight frontier LLM agents, defaults fall at least 62 points below an oracle handed the relevant preference, and prompting closes little of it. Diagnostics identify acquisition as the bottleneck. ATRBench surfaces this proactivity gap in current agents and offers a diagnostic testbed for closing it.
Bin Wu, Guanyun Zou, Bingbing Wang +2
Beijing University of Posts and Telecommunications · Nanjing University of Aeronautics and Astronautics · Harbin Institute of Technology (Shenzhen) +1