ArXiv PASK: proactive AI agents with long-term memory that predict user intent
Why it matters
A new paper, PASK, introduces a framework for proactive AI agents that combine intent detection, hybrid memory, and self-initiated action. The IntentFlow model reached the level of the leading Gemini 3 Flash models in recognizing latent user needs.
From reactive to proactive agents
Most of today’s AI agents are reactive — they wait for the user to tell them what to do. PASK (Proactive Agent System with Knowledge) goes a step further: the agent itself detects what the user needs and proposes actions before the user has to ask.
Imagine an assistant that notices you have 3 meetings in a row and automatically prepares summaries of each, suggests water breaks, and orders an Uber for the next location — all without you saying a single word.
How does PASK work?
PASK implements a new paradigm called DD-MM-PAS:
1. Demand Detection (DD) — Need detection
- Uses the IntentFlow model to infer latent user needs
- Analyzes current context (conversation, calendar, location, habits)
- Identifies what the user would likely want NEXT
2. Memory Modeling (MM) — Hybrid memory
- Workspace — short-term current session
- User memory — long-term user knowledge (preferences, habits)
- Global memory — general knowledge about the world
3. Proactive Agent System (PAS) — Proactive system
- Closed-loop infrastructure that combines detection and memory
- Generates proactive actions without waiting for a query
Performance
The authors built LatentNeeds-Bench — a benchmark of real user data. Their IntentFlow model:
- Reaches the level of Gemini 3 Flash models under identical latency constraints
- Identifies deeper user intents better than baseline models
- Functions in real-time conditions for everyday use
Why is this important?
PASK directly addresses practical challenges that previous research under laboratory conditions did not adequately solve:
- Depth — understanding subtle or implicit needs
- Complexity — combining multiple sources of context
- Ambiguity — handling scenarios where the user does not know exactly what they want
- Precision — avoiding incorrect or intrusive suggestions
- Real-time constraints — the response must come in milliseconds, not seconds
Broader implications
Proactive agents are the next major iteration of AI assistants. Current assistants (Claude, ChatGPT, Gemini) are outstanding reactively, but the user must know WHAT to ask. Proactive agents reduce this cognitive load — the assistant makes suggestions itself.
The risk is obvious: too much proactivity = intrusiveness. PASK tries to balance this through fine-grained intent detection, but the ultimate test will be in real-world application.