If you’ve spent any time in productivity circles recently (and I assume as a reader of this blog you are) you’ve probably noticed a growing trend: people are building a “second brain” powered by LLMs and personal knowledge bases.
The recent wave was largely sparked by a post from former OpenAI researcher Andrej Karpathy, who described how he uses local markdown files together with AI as an external memory system. Combined with tools like “Obsidian”, vector search, and increasingly capable language models, the idea is certainly compelling.
But like every productivity system, the important question isn’t whether you can build it—it’s when this pattern actually makes sense, and when it’s simply overengineering.
So before you install Obsidian and start migrating your entire life into hundreds of .md files (which is quite cumbersome, to be honest), it’s worth asking whether the setup is actually worth the effort. And if yes, what is the right use case for you.
The Idea in a Nutshell
If you haven’t seen the concept yet, Andrej Karpathy’s original post and this accompanying video (there are many more great ones on this topic, but do not get lost in details ;D) gives a great overview of the workflow. Personally, I found this blogpost with a basic “how to” description, most helpful for getting started.
The basic idea is straightforward: Instead of relying solely on the model’s training data or the public internet, you maintain your own knowledge base consisting of markdown documents, which you can of course generate from other data formats than markdown (i.e. PDF articles, code repositories, spreadsheets, images and more). Whenever you ask your AI assistant a question (usually via AI harness, like Hermes Agent, Claude Code, etc.), it retrieves the relevant notes and uses them as additional context.
The result is that you receive very accurate answers based on the material you’ve fed to your LLM wiki and whenever you feed new information you update the knowledge base, so that it can create new wiki pages that might help you to understand the information you provided in another context. Ideal for learning and querying complex domains!
When It Actually Pays Off
This approach shines when you benefit from a specific body of information that is not easily searchable on the internet.
For example:
- your personal knowledge on a specific domain (i.e. securing a specific tech stack)
- your team’s internal documentation
- company-specific processes
- project documentation (including all these notorious meeting notes)
- research you’ve accumulated over months or years
In other words, information that only exists because you curated it.
A second important requirement is that this knowledge base serves a defined goal.
If you’re simply dumping random notes into a wiki “just in case,” the return on investment is often low.
On the other hand, if better access to your knowledge directly improves important decisions—because choosing the right information, approach, or skill has a significant impact on a project outcome, or on your team’s problem—then an LLM-enhanced knowledge base can become incredibly valuable.
Therefore, as a rule of thumb: the more specific and less possible to query your required information via simple web search, the more worthwhile it becomes to organize your knowledge and put work into building your LLM wiki.
When It’s Probably Overengineering
If most of the information you use can already be found quickly on the web, or you rarely revisit your notes, maintaining a sophisticated LLM knowledge base may create more work than value.
There is also the practical aspect: Maintaining an LLM wiki requires discipline, organization, and continuous maintenance for it to bring value to you. No one wants to get answers based on outdated information. So before committing to the system, it’s worth considering whether the benefits justify that ongoing effort.
One Important Security Consideration
One thing that’s easy to overlook is how you collect information in the first place. There are a few markdown web clippers in your browsers market places that request very broad permissions on the content you are clipping. They can potentially access everything on the page you’re clipping—including secrets or confidential data.
If you’re collecting information from websites that use basic authentication via user-password or that collect personal data, please be cautious. A simple precaution is to use a dedicated browser for this kind of “harvesting” work. Separating your clipping workflow from your normal browsing (where you might accidentally clip data from your online banking account 🙈 ) reduces the risk of accidentally exposing sensitive information.
The Bottom Line
LLM-powered personal knowledge bases are an exciting evolution of the “second brain” idea coming from Tiago Forte and his very insightful book, although AI researcher Andrej Karpathy brought up the idea of LLM powering one’s knowledge base.
But they’re not universally useful.
They work best when you have valuable knowledge that isn’t publicly available and when that knowledge supports a clear objective.
Otherwise, you may end up building an impressive system that mostly manages information you could have found with a quick web search.
As with most productivity tools, the smartest solution isn’t the most sophisticated one—it’s the one whose maintenance cost is justified by the value it creates.
Checklist
- Identify one concrete project, decision, or domain where better access to your own knowledge would create meaningful value.
- Focus on not more than 3 separate wikis for a start.
- Test the workflow with a focused set of notes before migrating your entire knowledge library into your LLM wiki.
- Set up a process to keep information organized and current (ingest new data and update the index at the same time and lint regularly); otherwise, outdated notes can make the wiki less useful.
- Avoid giving web clippers unnecessary access to confidential information, passwords, or personal data and use a dedicated browser for web clipping.