Introduction
At the I/O conference on May 20, 2025, Google unveiled the mobile version of the NotebookLM app. Initially offered as a web-based tool for AI note-taking and research assistance, users could import materials into a notebook, after which the AI would generate summaries, answer questions, and more.
After a period of refinement, the features are now well-polished. On the left, you see the added resources; the central main screen is for conversations (powered by Gemini 2.5 Flash), and on the right is the notebook or, as its name suggests, the studio.
You can import various file types into NotebookLM. For PDFs and audio files, it automatically processes them (extracting text or converting speech to text) into plain text. Regular users seem to be allowed up to 50 documents per notebook, while paid users can add up to 300. Below, you can see that in addition to dragging and dropping documents, you can also import directly from Google Drive, via website links, or YouTube videos—and the simplest method: pasting text.
In the top-right corner of the source addition area, there’s a brand-new feature called Explore Sources. Often, when you want to delve into a topic without having any material at hand, you simply type in the topic you're interested in, and it will help you search for relevant information.
In cases where the material you have imported is limited and does not provide a comprehensive view of the subject you want to explore, you can click “I’m Curious,” and it will help you explore related resources based on what you’ve already added (though its performance in Chinese is generally average).
When you first import materials, it automatically generates a summary. In the studio section, you can also let it create an FAQ, Briefing Doc, or Study Guide with just one click based on your materials. This is especially useful for learning: using quizzes to help you better understand, using summary documents to grasp the general content, and then using study guides to map out your learning path.
I believe the recommended order of usage for these tools is as follows:
- Briefing Doc: First, it provides an overview of the entire topic, giving you an initial understanding and forming a basic cognitive framework.
- Study Guide: Once you have a general idea of the content, you need to study step by step.
- FAQ: You can test your understanding of specific points as you learn through questions.
The central section (the Q&A dialogue area) serves as a 24/7 on-demand teacher that answers your questions. The AI model uses your existing materials to respond, and all answers include clickable reference snippets to support cross-document reasoning, ensuring accurate answers while helping learners revisit the relevant points.
In the top-right corner of the studio, there’s an “Audio Overview” feature. I suspect this is akin to the pioneer of many current AI podcast tools. Since NotebookLM’s launch, numerous AI podcast apps have emerged. After generating audio, two AI hosts engage in a 5–15 minute conversational explanation based on the materials; the feature supports offline downloads and background playback, and currently offers output in over 50 languages. This means you can learn about a topic through audio during your commute or other spare moments.
Overall, NotebookLM covers the full chain from input to output for a given topic and is backed by Google’s powerful underlying models, making it an outstanding learning tool. Although it supports multilingual output, in my tests the English version works best, and its RAG mechanism can sometimes be unstable with other languages.
Mobile Version
When 96% of internet users carry their "information gateway" in their pockets, a knowledge assistant available only on desktop is like a good book locked in a cabinet—you can see it but you can’t use it.
A staggering 96.3% of global internet users access the web via mobile phones, and 63% of global web traffic comes from mobile devices. There’s no doubt that desktops are no longer the primary battlefield; mobile phones have become the primary information endpoint. Your readers, students, and—nine out of ten people—will learn on their phones.
Modern life is full of fragmented moments: waiting in line, commuting, standing by for someone. These are scenarios where your computer isn’t as handy as your mobile phone. Moreover, in today’s fast-paced world, people no longer want to spend long stretches of time reading lengthy texts.
With an average of over 4 hours of mobile usage per day, a quick tap to open NotebookLM is all it takes to capture your attention.
Before the mobile version was available, whenever we encountered a good PDF online while out and about—if I didn’t have my laptop with me—I would first save it to WeChat’s file transfer assistant. Then, when I finally got home and opened my PC, I would log into WeChat and transfer it to NotebookLM, sometimes even forgetting about it. This added extra steps and significantly increased the operational friction.
With the mobile version, you can share directly to the app, making the process streamlined and immediate.
In addition to being more convenient and engaging for users, most features are essentially the same as the online version, though the functionality in the studio section isn’t yet complete. In the future, I believe that for NotebookLM to truly evolve into a comprehensive learning assistant, it will need to introduce new experiences or features.
Modern knowledge management should not only focus on acquiring, organizing, and generating learning outputs but also on consolidating long-term memory and its practical application. NotebookLM opens up the possibility of building a whole new digital learning ecosystem. In my view, it could also consider adding functions such as a “Spaced-Prompt Scheduler” and “Cross-Notebook Collaboration.”
Spaced-Prompt Scheduler (Automated Spaced Repetition Scheduling)
Traditional spaced repetition systems (like Anki) use algorithms to schedule reviews of flashcards at progressively increasing intervals to maximize memory retention. Perhaps NotebookLM could incorporate a similar feature, automatically generating quiz questions or flashcards from user materials. With this feature, after a user studies a particular topic, NotebookLM could automatically generate a set of review questions and intelligently schedule the next prompt based on the user’s responses.
For example, if a user reads a report in one day, NotebookLM could generate 10 key-point questions; if the user is uncertain about certain answers, the system will prompt similar questions the next day (to reinforce memory), while correct answers might be scheduled for review several days later.
Thus, users wouldn’t have to manually create a flashcard library—NotebookLM acts as an intelligent review assistant, dynamically adjusting the review schedule. Given that NotebookLM can understand semantics and generate diverse questions, it could even rephrase questions in subsequent reviews to ensure users truly grasp the concepts rather than just memorizing answers.
This AI-driven spaced repetition blends the human forgetting curve with machine precision reminders, potentially improving long-term retention.
Human memory fades very quickly if not reinforced—the phenomenon of exponential knowledge loss without review was identified by Ebbinghaus in the 19th century, famously known as the forgetting curve.
The red line represents the rapid decline in memory retention without review, while the blue line shows how spaced repetition resets memory and gradually extends retention time, significantly enhancing long-term memory.
The so-called “Forgetting Tax” is a vivid metaphor for the cost we incur due to forgetting. When knowledge isn’t reinforced in time, we eventually have to spend extra time and energy re-learning it—much like an accumulating tax that must eventually be paid. Understanding this cost helps us balance the trade-off between timely review and the higher energy expense of relearning.
The “Forgetting Tax” can be seen as the large area of forgotten information (illustrated by the gray arrow under the red line) that requires re-learning. In contrast, the blue line shows that by “paying a small fee” (a little review) at the right time, you avoid a massive cumulative loss.
Numerous studies and practices since Ebbinghaus (for example, the SuperMemo algorithm and Anki software) have demonstrated the effectiveness of spaced repetition in combating forgetting.
As early as 1932, educational psychologists suggested that “review sessions should be spaced at increasing intervals… approximately 1 day, 2 days, 4 days, 8 days, and so on.” This laid the groundwork for the various spaced repetition algorithms that followed.
The rationale is to schedule each review just as the memory is about to fade.
Reviewing too early wastes time, while reviewing too late means the material has already faded. A scientifically arranged review schedule maximizes retention and minimizes re-learning time.
Forgetting in itself isn’t frightening—what’s dangerous is ignoring the cost of forgetting. If we let knowledge fade, we will eventually have to invest much more effort in catching up. This "tax" might manifest as all-nighters before exams or as the inefficiency of repeatedly searching for information at work. Conversely, by regularly “paying” a small tax (following the forgetting curve with timely revisions), we can prevent a snowballing loss of knowledge.
For instance, medical students could submit lecture notes to NotebookLM, which would then generate a daily set of Q&A prompts based on the material, automatically scheduled according to Ebbinghaus intervals (1 day, 3 days, 7 days, 14 days…). This “short, frequent practice” model allows students to continuously reinforce their memory with minimal time investment, without the hefty cost of re-learning.
Cross-Notebook Collaboration
Another promising direction is cross-notebook collaboration. In reality, knowledge is rarely confined to a single topic; interdisciplinary connections and cross-project interactions are the norm.
If NotebookLM could enable different notebooks to be interconnected, allow collaborative construction by multiple users, and support cross-domain searches—similar to connecting scattered islands with bridges and undersea cables—knowledge and inspiration could flow, collide, and appreciate on a much larger scale.
Turning “isolated knowledge” into “knowledge archipelagos.”
For cross-notebook collaboration, the first challenge is to avoid having the same material copied repeatedly, so that an update in one place is reflected everywhere. This would eliminate the need for widespread copy-pasting and reduce version conflicts. When writing reports, citations would update along with the original sources, eliminating discrepancies between the main text and data tables.
Currently, the chat is limited to a single notebook. For true cross-notebook collaboration, there would need to be a universal workspace where conversations could pull data from one or more notebooks to supplement information.
Perhaps the interface might look something like this:
On the left are various notebooks corresponding to different folders. On the right, the workspace allows you to select one or multiple notebooks for discussion. This creates a creative space that is clear and intuitive.
The next aspect involves symbiosis—knowledge that grows on its own. Lately, linked note-taking systems like Obsidian have become popular for connecting different notes, thereby forming an interconnected web of knowledge.
Tools should not merely be containers but collaborative partners—proactively suggesting connections, helping you discover blind spots, and even scheduling reviews.
AI could run similarity matching in the background. For example, when you write a paragraph or mention a concept during a conversation, the AI could, in the background, match it with previously recorded ideas.
This can reduce redundant work; for instance, if a research idea was recorded three months ago, the AI could flag its relevance.
Of course, this process needs to be gradual. If AI immediately decides that something is connected based solely on vectorized embeddings, it might produce inaccurate associations since creators may not view the connection as relevant.
Therefore, initially, the AI could provide a list of possibly related items for the user to decide whether to link them. Once the AI learns your preferences, it could gradually take over this task.
YouMind: An “AI Creative Studio” That’s More Than Just Notes
This brings to mind YouMind, an AI software I started using last year. YouMind brands itself as an AI Creation Studio: it integrates information collection, AI research, card organization, writing generation, and external sharing into one “board” (project board), aiming to help creators complete their work from inspiration to final draft in one go.
The board here is somewhat like a notebook, but within the board, there are groups. If we consider each group as a notebook, YouMind allows for cross-notebook AI chat because it also conducts chat based on a board. But if we regard the board itself as a notebook, it no longer supports cross-notebook conversation; however, it has already integrated the entire process from input to learning to output, forming a creative IDE.
The middle section is for viewing files or editing, and on the right is an AI toolbox for chatting or customizing prompts. The default tools include a Reader (which summarizes knowledge for you), Chat, Note (for taking notes), and Translator.
Currently, NotebookLM focuses more on the learning aspect, while YouMind leans towards organization and output. For learning, NotebookLM’s underlying model performs more finely in key-point retrieval compared to YouMind. On the other hand, YouMind allows you to customize prompts, letting you convert any material into your desired output. Its entire suite functions like a creative studio.
I see that the official team is also planning a mobile version—exciting! 👀
Conclusion
Last week, while on the subway heading home, I came across a lengthy psychology research paper. In the past, I would send the link to my WeChat Transfer Assistant to save it as a to-do item, then upload it later at home. However, last week coinciding with the launch of NotebookLM’s mobile version, I directly uploaded the file via the mobile app. Within minutes, I generated an AI podcast using its Audio Overview feature—the entire process took less than ten minutes, and the resulting audio was just over ten minutes long.
By the time the subway reached my stop, I had a preliminary understanding of the report's core ideas. It was at that moment I truly realized that the future of knowledge management truly comes with you wherever you go.
If desktop note-taking tools taught us how to store information, the combination of mobile and AI is teaching us how to understand and reuse information in a timely manner. Learning is shifting from long, stationary sessions to fragmented micro-cycles—from observation, capture, questioning, to output.
Future knowledge workers might resemble information commanders rather than mere data movers. We can effortlessly deploy AI to arrange and review vast amounts of content in real time, leaving deep thinking to us humans.