An Enlarged Intimate Supplement to His Memory

Click to zoom

In 1945, the war had just ended. Vannevar Bush, the man who had coordinated six thousand American scientists during WWII, who had overseen the research that led to radar and the atomic bomb, who advised Roosevelt directly, sat down and wrote an essay for The Atlantic called "As We May Think."

The problem on his mind wasn't weapons. It was information. The scientific community was producing knowledge faster than anyone could organize, find, or use it. Papers buried other papers. Discoveries went unconnected. The sum of human knowledge was growing, but our ability to access it wasn't keeping up.

His proposed solution was a machine he called the Memex.

"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."

An enlarged intimate supplement to his memory. Read that again. He wasn't describing a filing cabinet. He was describing something closer to a second brain, a system that doesn't just store what you know, but makes it retrievable at the exact moment you need it.

Eighty years later, we have machines Bush couldn't have imagined. And most people use a CRM they forget to update.

The moment I knew

I was on a call with someone I'd spoken to three months earlier. They'd mentioned a specific problem their team was struggling with, something about their data pipeline breaking under load during quarterly reporting. I remembered the conversation vaguely, the way you remember a dream, but I couldn't recall the details. What exactly was breaking? What had I said I'd follow up on? Had I even followed up?

I checked Slack. Nothing useful. Scrolled through iMessage. Found a thread but it was mostly scheduling. Searched my email. Found a forwarded article I'd sent them but no notes on the actual conversation. The context existed somewhere across all these apps, fragmented, buried, useless when I actually needed it.

I stumbled through that call. Afterward, I sat there thinking about how absurd it was. I'd had the conversation. I'd had the information. I just couldn't find it when it mattered.

I kept thinking about the Memex, not as a product idea, but as a personal problem. I was losing context on people I cared about, deals I was working, conversations that mattered, not because I didn't have the information, but because it lived in too many places and none of them talked to each other.

Starting small

The first version was embarrassingly simple. A git repo with markdown files, one per person. Name, company, how we met, and a running log of notes I'd type up after calls. It was basically a fancy address book.

But even that changed things. Having one place to look, one file per person, meant I actually wrote things down. And because it was git, every change was tracked. I could see how a relationship evolved over time, not just where it stood today.

The structure grew naturally. A /people/ directory for everyone I interact with. /projects/ for active work. /active_leads/ for pipeline deals. /outreach/ for prospecting. Each file followed the same format because I'd built rules into the repo that enforced consistency.

It was useful. But I was still typing everything manually. And manual systems, no matter how well-designed, decay the moment you get busy. I needed the information to flow in without me thinking about it.

Plugging in

iMessage was first. I wrote a script that reads the Messages database directly, copies it to a temp file for safety, and pulls any conversation by contact name or phone number. Suddenly I could search every text I'd ever sent or received. That alone was worth the effort.

Then email. Apple Mail stores everything in a local SQLite database, so I wrote a script to query it directly. Faster than any API. Sent and received, searchable by sender, subject, or date range.

Slack came next, because half my work conversations were happening there and none of it was making it into the CRM. API export across all channels and DMs, filtered however I needed.

The real unlock was meeting transcripts. I use Granola, which records meetings automatically and stores transcripts locally. Once those plugged in, every meeting became searchable text.

The most recent addition is Wispr Flow. It runs in the background all day handling my voice dictation, every message I speak, every command, every stray thought I dictate while walking between meetings. It logs everything to a local database. I built a script to pull it all by date range. 115,000 words captured so far. Things I said out loud that I would have otherwise lost completely.

Six data sources in total, each one added because I'd lost something I shouldn't have. Each one a patch over a specific failure of memory.

As of today, the system holds 106,000 iMessages, 51,000 emails, 2,100 voice dictations totaling 117,000 spoken words, 138 recorded and transcribed calls and meetings, and 122 individual person files, all versioned across 116 git commits. It keeps growing every day, and none of it required me to type a single note.

The overnight shift

The system ran on manual pulls for a while. I'd run the scripts when I needed something. But the whole point was to not rely on my own discipline, so I set up a cron job that runs every night.

While I sleep, it pulls the last day of iMessages across a hundred threads, sent and received emails from the past week, Slack conversations filtered by date, meeting transcripts from Granola, and every Wispr Flow dictation from the day. Then an AI agent processes all of it, reads every person file, cross-references the new data, adds interaction notes, flags follow-ups, creates tasks.

The first morning I woke up to a fully updated CRM, with notes from yesterday's meetings already filed under the right people, with follow-up tasks already created, I realized I'd built something that didn't exist before. Not a better CRM. A system that actually remembers.

Full recall

My favorite feature came out of frustration. I had a meeting in twenty minutes and I couldn't piece together everything I knew about this person across five different apps. So I wrote a script that does it all at once:

python3 person_dump.py 'Their Name'

Every iMessage thread, every email, every call transcript with that person, across all channels, all history, in seconds. One command, one output, complete context.

I ran it before that meeting. I knew what we'd discussed three months ago. I knew the exact problem they'd described with their data pipeline. I knew what I'd promised to send. I referenced it naturally in the conversation, and they were genuinely surprised I'd remembered. I hadn't remembered. The system had.

That deal closed. And I knew exactly why.

The conversation layer

I don't actually run these scripts by hand anymore. I run a Claude Code instance that sits on top of the repo, and the repo itself teaches the agent how to use everything. Skills, rules, integration docs, all baked into the codebase. The agent reads them and just knows.

So in practice, I just talk to it. "What did I discuss with Ashish last week?" and it pulls the iMessages, the meeting transcripts, the emails. "Prep me for my call with David" and it runs the person dump, summarizes the relationship, surfaces anything I need to follow up on. "What have I been saying about the pipeline?" and it queries my Wispr Flow dictations from the past week, filtered by topic.

I never think about which script to run or what flags to pass. The system documents itself, the agent reads the documentation, and I just ask for what I need. This is what makes it feel like the thing Bush imagined, not a tool you operate, but a system you converse with. The data is always there, always accumulating. The agent is the layer that makes it useful the moment you need it.

Context only appreciates

This system has made me hundreds of thousands of dollars. But the money isn't really the point. The point is what the money reveals about how value works.

Context is an asset that only goes up.

Not all context is equal. Relationship context, what people care about, what they've told you, what you promised them, that compounds quietly in the background until the moment it matters. Records of your own habits and outcomes compound the same way. You start seeing patterns you couldn't see before, what works, what doesn't, who to prioritize, when to reach out.

A conversation from six months ago might seem irrelevant today. Next quarter it could close a deal. Someone mentions a pain point in passing on a call, you forget about it, and three months later you have exactly the solution they need. But only if you captured it, only if the system remembered what you didn't.

You can't predict when context becomes valuable. That's precisely why you capture everything.

Building for the future

Here's the part most people miss, and it's the part I think about most.

Agents are getting better every month. The models that process language, search for connections, and take action on your behalf, they improve continuously. Which means the context you collect today becomes more useful over time, not less.

Think about what that means. Every transcript sitting in your system, every message thread, every voice note you dictated while driving. The raw material stays the same. But the tools that can search it, connect it, and act on it get sharper every year. Context you captured in January might be unsearchable noise today and a goldmine by December, simply because a better agent learned how to find the pattern.

That's why local matters, why ownership matters. Files in a git repo, markdown that any tool can read, not locked in a SaaS database that might sunset, change their API, or decide your data is their training set.

When the next generation of AI tools arrives, and it will, the people with rich local context will have something no one else can buy. Everyone else will be starting from scratch, feeding new systems with whatever they can remember.

The supplement to memory

Bush's essay went on to shape the next eighty years of computing. Doug Engelbart read it and invented the mouse and hypertext. Tim Berners-Lee read it and built the web. The Macintosh, the iPhone, the browser you're reading this in, all of it traces a line back to a single essay written in the summer of 1945 by a man who had just helped win a world war and turned his attention to the next problem worth solving.

What strikes me most isn't that Bush predicted the technology. He didn't, not really. Microfilm and electromechanical levers look nothing like git repos and language models. What he got right was the shape of the need. That a person's records and communications would grow beyond what memory could hold, and that the system built to extend that memory would become inseparable from the person's ability to think, to connect, to act.

He described it in a single phrase that has stayed with me since I first read the essay.

"An enlarged intimate supplement to his memory."

Eighty years later, that's exactly what I built. The materials are different. The vision is the same.