A practical guide to building a knowledge base that grows over time, using Google NotebookLM, no technical setup required.
There is a pattern to how most people use AI for research. Open a new chat, paste some context of who you are, what you are working on, what you already know, ask your question, get an answer, close the tab. The next day, open a new chat. Paste the same context again.
Some AI tools now offer memory features that carry basic context between sessions. Claude remembers your name and what you are working on. ChatGPT does something similar. But that is not the same as a knowledge base. Those tools remember you, not your research. They have no understanding of how today's article connects to the report you read last month, no accumulated picture of the topic you have been building for the past six months. The synthesis still happens in your head, or not at all.
NotebookLM sits somewhere in between. It stores the sources you upload and answers questions grounded in them, which is more persistent than a standard chatbot. But each notebook is isolated. Sources added this week do not automatically update summaries written last month. There is no second layer that connects, cross-references, and maintains a coherent picture over time — unless you build one.
That is what this guide is about.
Most AI tools are stateless by design. They process whatever is in the current session and nothing else. NotebookLM is partly an exception to this, it stores your uploaded sources and answers questions grounded in them, which is already more persistent than a standard chatbot. But even within NotebookLM, each notebook is isolated. Your research from one project cannot reference material in another. Sources added this week do not automatically update summaries written last month. The synthesis has to happen in your head, or not at all.
For occasional queries, this is fine. For anyone building knowledge on a topic over weeks or months; tracking a market, monitoring a regulatory area, developing a body of expertise, it creates a ceiling. The tool stays flat while the complexity of what you are working on grows.
The problem is not that AI tools are bad at answering questions. It is that the work of connecting, updating, and maintaining a knowledge base still falls entirely on the person using the tool.
What changes with a structured approach is that the AI takes on more of that maintenance work. You still decide what sources are worth adding. But the summarising, cross-referencing, and updating happens in the tool rather than in a notebook you maintain manually or, more often, not at all.
The pattern is straightforward once you see it. Instead of uploading raw sources and querying them directly, you build a second layer: a set of structured summary pages that the AI creates and updates as new sources arrive. Your questions draw on those pages rather than the raw source pile. Over time, the pages get richer, and your queries get better answers.
Think of it like the difference between a filing cabinet and a reference library. A filing cabinet stores everything. A reference library organises everything into a form that is actually useful to consult. The AI is doing the librarian's work.
In NotebookLM, this translates to a practical workflow that anyone can follow without any technical background.
Start by creating a dedicated folder in Google Drive with two subfolders: one for raw sources and one for wiki pages. This separation matters, it keeps the original material distinct from the synthesised layer and makes the notebook easier to manage as it grows.
Open a new notebook in NotebookLM. This will be your working notebook for a specific topic area, a market you are tracking, a technology you are evaluating, a regulatory area you need to stay across.
Create a short Google Doc, one or two pages is enough that describes how you want the wiki organised. This is the schema. It does not need to be formal. It just needs to tell the AI what kinds of pages to create and what to include on each one.
A basic schema might define three page types: a source summary page for each new source you add, a topic page for recurring themes or concepts, and an index that lists everything in the wiki with a one-line description. Add this doc to your NotebookLM notebook as a source. The AI will refer to it when you ask it to create or update pages.
Add a source to your raw sources folder and to the notebook. Then ask NotebookLM to do three things: summarise the source according to the schema, identify the key topics and entities it covers, and note any claims that relate to material already in the wiki.
Take that output, save it as a Google Doc in your wiki pages folder, and add it back into the notebook as a source. You now have both the raw material and the synthesised page in the same notebook, and the AI can draw on either.
This is the step that creates the compounding effect. When a new source arrives, do not just add it to the notebook and move on. Bring it in alongside the relevant wiki page, ask the AI to review both and suggest updates, apply the changes, and save the updated doc.
Over time, the wiki pages become progressively richer. An entity page that started with three bullet points from one source might have a paragraph of nuanced context after ten sources have touched on the same topic. That depth is what makes the system more useful than a standard document query.
Once you have a few wiki pages built up, shift your query habit. Instead of asking NotebookLM to find something in your sources, ask it to draw on the wiki pages to answer your question. The answers will be more synthesised and better referenced because the work of connecting sources has already been done.
Every few weeks, ask the AI to scan the index and flag anything that looks outdated, any topics that appear across multiple pages without a dedicated page of their own, and any sources that were added but not fully reflected in the wiki. NotebookLM's source-grounded approach makes this a reliable check, it will work from what you have actually added rather than filling gaps from general knowledge.
NotebookLM has some limits worth knowing before you build on it.
Each notebook is isolated, so if your knowledge base spans multiple topic areas you will either need to consolidate or accept that cross-topic connections require a manual step. A shared index document added to multiple notebooks can partially bridge this.
The free tier supports up to 50 sources per notebook. For a structured wiki with a schema document, source summaries, and topic pages all counting toward that limit, 50 fills up faster than you might expect. The Plus tier at around $20 per month raises this to 300 sources, which is workable for most individual use cases.
NotebookLM connects to Google Drive, Docs, PDFs, URLs, and YouTube. If you regularly work with source types outside this list, you may need to convert or paste them in manually.
None of these are blockers for most people. They are just worth knowing so you can design your notebook structure with them in mind from the start.
This approach is most valuable when you are working with a topic that evolves over time and where the connections between sources matter as much as the individual sources themselves.
In each of these cases, the effort of maintaining the wiki is lower than the effort of rebuilding context from scratch every time. That is the trade-off worth making.
The setup takes less than an hour. Create the Drive folders, write a simple schema document, add your first few sources, and ask NotebookLM to build the first wiki pages. The system does not need to be comprehensive from day one. It just needs to be consistent — adding and updating with each new source rather than letting the pile grow unmanaged.
If you are looking at a larger implementation, a shared knowledge base for a team, an integration with existing workflows, or a more automated version of this pattern, we are happy to talk through how to approach it.
Get in touch at itsavirus.com to explore what makes sense for your context.
