AI Glasses Start Learning to Take the Initiative

A New Shift in Interaction

By ZJ, VR Tuoluo

A deadline casually mentioned in an ordinary chat can be impossible to recall days later. In the past, you would have had to dig through chat logs or ask a colleague; now, a single question to your AI glasses can bring the information back.

Scenarios like this are already appearing in real-world use of today's AI glasses. With long-term memory features, some AI glasses can, with user permission, record everyday conversations and organize key information into searchable memories. Going further, even if users do not explicitly create schedules or to-do lists, the AI can try to extract relevant information from conversations and proactively remind them later.

To-do reminders from TCL RayNeo iO   Image source: VR Tuoluo

Over the past few years, AI glasses makers have invested heavily in making devices understand user questions, recognize objects in view, and respond via voice or displays. Entering 2026, more and more vendors are exploring how AI can help before the user even speaks. Meanwhile, academia is studying how AI can judge when to intervene proactively, and even ask users questions on its own.

As AI moves from answering questions to participating in daily life, how far have proactive AI glasses actually come?

01

From Proactive Recording to Proactive Reminders, AI Glasses Start Finding Things to Do on Their Own

At Meta Connect 2026 in September this year, Meta announced it would bring its personal AI agent Muse to its AI glasses.

Image source: Meta

Under Meta's vision, when users see products on a shelf, event flyers, or a shopping list, Muse can combine what's in front of them to understand the need and call external apps to handle follow-up tasks. Muse can also continuously track user goals and make plans. However, these features have not yet officially launched on AI glasses; judging by the mobile app's current performance, Muse is already remarkably impressive.

Unlike traditional AI chat apps, Muse provides each user's chat window with its own dedicated cloud computing environment — effectively giving the AI a computer it can operate autonomously. This means Muse can not only call large language models to answer questions, but also operate web pages, use software, and execute multi-step tasks like a real person, even continuing to work after the user leaves the chat interface — greatly expanding the practical use cases for AI agents.

Latest data shows that Meta's AI agent app Muse surpassed 5 million downloads just 22 days after launch, far ahead of ChatGPT's 56 days and Grok's 103 days, setting a record for the fastest growth in the AI sector. On the hardware side, what Meta is describing is essentially the deeper convergence of AI glasses and personal agents. In the Chinese market, proactive AI has also begun entering more specific everyday scenarios.

Image source: VR Tuoluo

These products show that proactive AI currently develops along two main directions: one is having the AI remember users' past conversations and behaviors and recall them when needed later; the other is combining real-time environment, time, and user state to proactively provide reminders or suggestions.

A prime example of the former is the TCL RayNeo iO, released in August this year. Its "All-Day Smart Memory" feature can, with user permission, continuously record everyday conversations and organize people, events, times, and other details into long-term memory. In VR Tuoluo's earlier review, one use case was demonstrated: an event date confirmed in a meeting days ago may have slipped the user's mind, but by asking the glasses, they can retrieve the answer from earlier conversation records.

TCL RayNeo iO Image source: VR Tuoluo

Going further, the system can extract to-do items from conversations and surface them again in later summaries or reminders. For users, ad hoc arrangements that were never explicitly recorded can also be preserved.

The other path focuses more on real-time scene awareness. For example, the Qwen AI Glasses S1 can combine weather, location, time, and wearing status to remind users to take an umbrella before leaving home, or suggest moving around after prolonged sitting. Rokid's AI Assistant 2.0 also adds automatic itinerary logging and scheduled proactive reminders, letting the glasses provide relevant prompts based on travel plans.

Image source: Rokid

In addition, Even G2 and Halliday G2 bring proactive prompts into conversations, meetings, and other scenarios, while the Li Wei Ke X-AI Memory Glasses attempt to reduce manual note-taking through automatic scene recognition and long-term memory.

These features are already changing how AI glasses are used, but the degree of proactivity still varies. Weather and itinerary reminders mainly rely on known triggers; long-term memory requires extracting useful information from scattered conversations; real-time conversation assistance must understand what is happening and judge what deserves a prompt. As for cross-scenario autonomous planning and task execution, this remains at the exploratory stage.

As AI glasses take on more proactive services, the importance of continuous sensing and long-term memory grows. AI needs to know what the user is doing and what has happened in the past to judge when and how to help. How these capabilities are actually implemented is the next problem proactive AI glasses must solve.

02

What's Still Blocking Proactive AI from Making Decisions?

Current products show that proactive AI can already work in well-defined scenarios, but making AI understand complex daily activities and judge on its own when to help still presents many challenges.

First is the accuracy of long-term memory. For example, a user confirms a project deadline in a morning meeting, discusses revision requirements with a colleague in the afternoon, and the next day the client makes a last-minute change to the delivery date. To remind the user at the right time, the AI must link events occurring at different times and update earlier memories with new information.

The paper "EgoLife: Towards Egocentric Life Assistant," published at CVPR in 2025, addresses exactly this problem. The research team collected roughly 300 hours of life data from 6 participants over a week using AI glasses, covering conversations, shopping, cooking, and other scenarios, and proposed the EgoButler system to tackle understanding and retrieval of long-term first-person visual information.

Image source: paper "EgoLife"

The study demonstrated the feasibility of combining multimodal models with retrieval mechanisms to handle long-term life information, but also exposed current models' limitations in person recognition and cross-time information linking. This means that even if AI glasses can record vast amounts of everyday information, they may not accurately find the connections within it, let alone keep long-term memory consistently up to date.

More complex than remembering information is judging when to intervene proactively. Many proactive AI features still rely on preset conditions or specific operating modes. For example, weather changes can trigger reminders, and during meetings the AI can continuously analyze conversation. But in more open living environments, user needs are not always so clear-cut.

In July this year, researchers from Dalian University of Technology, the University of Tokyo, and Alaya Lab published the paper "Vinci2: Providing Proactive Assistance in Continuous Egocentric Videos." The paper focuses on proactive decision-making in continuous sensing — that is, AI must combine the user's current activity, past experiences, and long-term habits to judge on its own whether to offer help.

Image source: paper "Vinci2"

The team built an evaluation benchmark called EgoServe, containing more than 3,000 proactive service instances across 10 scenario categories, including immediate safety alerts, task guidance, memory retrieval, and long-term habit suggestions.

At the same time, the team proposed the EgoMemo system, which continuously organizes information from first-person video through multi-scale temporal summaries, semantic knowledge graphs, and visual information retrieval, then decides whether to intervene. For instance, when the system observes a user wiping a knife bare-handed, it can issue a safety alert; if it notices the user repeating a certain action at different times, it can combine long-term records to suggest optimizations.

Image source: paper "Vinci2"

Experimental results show that EgoMemo achieved an overall F1 score of 8.0 on the EgoServe benchmark, higher than GPT-5-mini's 4.7 and Qwen3-VL-Plus's 3.5. However, even the best-performing EgoMemo still scored low overall. This result reflects that combining long-term memory does improve proactive AI service performance, but accurately judging when to intervene in complex, continuous real-world scenarios remains very difficult.

Beyond memory and timing, proactive AI faces another problem: how to acquire more personalized knowledge about the user?

The paper "EgoAsk: Egocentric Teaching of Personalized Object Knowledge for Household Robots," published in September this year, proposes having AI proactively identify its own knowledge gaps and choose the right moments to ask users questions.

The research primarily targets household robots. Because robots can hardly know each household's living habits in advance, the team tried sharing users' first-person views via AI glasses to identify the objects users are handling and the activities they are performing, then having the robot proactively ask about an object's purpose or ownership. In an experiment with 18 participants, this approach reduced the burden on users of constantly judging what the robot needed to learn, and also reduced some of the need to recall relevant situations afterward.

Image source: paper "EgoAsk"

Although EgoAsk has not been applied to consumer AI glasses, it offers another research approach: when AI is uncertain about something, it can proactively confirm with the user during everyday interaction and gradually fill in personalized knowledge.

These studies show that technical challenges such as long-term memory, proactive decision-making, and personalized learning now have different solution approaches. However, current results are still mainly validated in specific tasks and experimental environments, and there remains a gap before stable deployment in complex everyday settings.

For AI glasses makers, integrating these capabilities into a single system and keeping it accurate and stable in daily use is the next problem to solve. And once these technologies truly enter consumer products, makers will also have to face real constraints around hardware resources, user experience, and privacy protection.

03

What Challenges Remain Before Proactive AI Truly Fits into Daily Life?

For AI glasses that must be worn for long periods, continuous sensing first creates pressure on power consumption and computing capacity.

Proactive AI must constantly acquire and process information about the user's surroundings, especially when first-person visual understanding is involved — cameras, sensors, and related computing modules must work continuously. But AI glasses themselves have strict limits on battery capacity, heat dissipation space, and weight. Even if some computing tasks are offloaded to a phone or the cloud, data transmission, network latency, and energy consumption remain unavoidable.

This also explains why many vendors currently start with scenarios like voice, meetings, time, and location. Compared with continuously processing full video frames, the cost of acquiring and processing this information is relatively manageable, and it is easier to scope the proactive service.

Image source: Halliday

Beyond hardware limits, proactive AI must also balance accuracy against intervention frequency. For example, a user mentioning in chat that "we might schedule a meeting next week" should be handled differently from a firm agreement to "meet at 10 a.m. next Monday." If the AI treats an unconfirmed plan as a firm arrangement and keeps reminding the user afterward, it may actually create a new burden.

Similarly, even accurate reminders may not come at the right moment. When a user is talking with a client, handling an urgent matter, or having a private conversation, a sudden prompt can be disruptive.

In addition, long-term memory raises privacy and data security concerns. What AI glasses record may include conversations and behavior of colleagues, friends, and other people in public places. As devices begin continuously collecting and organizing this information, when recording starts, where data is stored, who can access it, and whether users can delete memories at any time all affect real-world product use.

In the future, as AI glasses connect further with personal agents and begin helping users send messages, modify schedules, and even make purchases and payments, operating permissions and confirmation mechanisms will also need to be clearer. These issues determine how quickly proactive AI lands and whether users are willing to use such features long term.

04

Final Thoughts

As the example at the beginning of this article shows, AI glasses can now help users retrieve information from conversations days ago, extract to-do items from scattered dialogue, and proactively provide reminders in well-defined scenarios. And as long-term memory, environmental awareness, and agent capabilities continue to develop, future AI glasses have the chance to handle more continuous everyday tasks.

But for glasses that must be worn every day, whether proactive service becomes a genuine user habit still needs validation in more real-world scenarios. After all, once AI begins proactively participating in daily life, it must also learn to judge what is worth remembering, when to remind, and when to stay quiet.

Of course, from the perspective of technological development, today's AI glasses marketed on proactive services have already leapt far beyond products from two years ago that mainly relied on user commands and single-turn Q&A, and are beginning to truly evolve toward the ideal of an all-day, companion-like personal AI assistant.

       Original title : AI Glasses Are Learning to Take the Initiative