Revisiting the Memex
A soul, heartbeat, and cron jobs
I built my first end-to-end agent yesterday!
But, before I get into that, take a second to read this quote from Vannevar Bush’s As we may think manifesto:
Consider a future device for individual use, which is a sort of mechanized private file and library. It needs a name, and, to coin one at random, "memex" will do. A memex is a device in which an individual stores all his books, records, and communications, and which is mechanized so that it may be consulted with exceeding speed and flexibility. It is an enlarged intimate supplement to his memory.
Bush wrote that in 1945…more than 80 years ago.
And yes, we’ve seen versions of his vision come to life in the past few decades. But I believe we have finally reached the point where the Memex has stopped being just a library where we store our thoughts, and started acting as a partner that we can consult and collaborate with.
The funniest part is that, at the core level, all this agent craziness just comes down to having text files in a folder and some automated subroutines that run at a set cadence or when triggered.
And yes, I know. “Agents are supposed to do much more like check your e-mail, send texts, and book your trip”. Blah blah blah. You’re missing the forest for the trees. Those things are all tools and chores that agents can learn faster than you think.
I mentioned this in my Context Compounds post and want to double down on it here:
The takeaway is this: as agentic tooling becomes widely accessible to the masses, the differentiating factor will be the context your AI setup has on you.
Or, as Nat puts it, the soul of your agent.
In my last post, Knowledge Work PRs, I explained the idea I had for an agent to help me strengthen my Obsidian.
Well, yesterday, in just a few hours of chit chatting with Claude Desktop (not kidding, I literally built the whole thing with voice), I was able to:
Set up a private GitHub repo for my Obsidian
Sync all ~450 files in my vault as a starting point (I also uploaded them separately to Google Drive to serve as a cold storage)
Using the Filesystem MCP server, gave Claude read-write access to my Obsidian vault
Have Claude generate an ‘agent-rules.md’ file that the agent has to read before doing anything (this is a crucial step that I’ll discuss in another post)
Then let Claude create the
agentsubroutine called ‘tend the vault.md’ to solve my problem. I literally gave it the link to my Knowledge Work PR post from Wednesday and it created instructions based on that.And then have Claude create a ‘PR folder’ so that anytime I triggered the subroutine, it would document the changes in a new file in that directory
Finally, I was able to have it test itself by running the subroutine end-to-end using my Commonplace notes from Feb 18th. It generated a list of my atomic notes it thought needed to be updated and drafts of the edits as well
I gave feedback on all the edits using voice and once we both did the first pass, it generated the final PR for me and I went ahead and pushed the changes to my repo
This was the result of the first PR Claude created after it ran the ‘tend the vault’ subroutine it wrote itself. All of those files are linked and I can just navigate to them to see the updates. For example, if you click on the ‘Engineering Agency’ file and go to the bottom, you’ll see this edit written by Claude.
What was most insane to me is that it was also able to create notes from scratch as well. There were some edits Claude suggested to me in the first draft of the PR that I shut down because they felt forced and would be better as standalone thoughts. It agreed and the result was fantastic…here’s a note that it created with tags, a body, and a link to a relevant note.
My point in sharing all this is not to flex the technical capabilities of what happened - there really wasn’t much to it. We’ve all seen vibecoded projects that are 1000x crazier. Rather, the key thing to emphasize is the fact that the agent worked so well because it had so much context to learn from my notes right off the bat.
The "memex mechanization" was just a few hours of voice-chatting with Claude. The "soul" was the 450 notes I had already written.
And the best part is that, as of yesterday, I’ve opened the agent floodgate. From here on out, I’ll have subroutines for as many repetitive and background tasks that make sense for my workflow. Not just to automate chores, but to continuously improve the quality and strength of my knowledge management’s soul.
Over time, as the agents and I learn to work together, my modern day memex (Obsidian lol) won’t just act as an archive of my thoughts but will truly serve as the second brain Bush predicted almost a century ago.