The CRM That Finally Stuck

There is a 2018 Substack post by Danielle Morrill (the Mattermark founder) that I read at least once a quarter. She built her own personal CRM in a Google Sheet. The exact tool she wanted, for an audience of one she knew better than anyone. Ten weeks in, she wrote this line:

“…the data entry of my own prototype started to kick my ass… I just randomly stopped for over 2 weeks.”

That is the entire story of personal CRM. And notice — it’s ALWAYS about the data.

I’ve been in this game for a while. Recently I spent the past three years working with CRM data, millions of records, and the lesson hits the same way every time: bad data in equals bad results out. So when I tell you the category collapsed, I know what I’m talking about.

If your business runs on relationships (consulting, agency work, sales, fundraising, financial advisory, legal practice, any operation where keeping track of people IS the job), you’ve felt the wall. You have a hundred and fifty people you should remember meaningful things about. You meant to write it down. You didn’t. The category of software that was supposed to fix this has been on the ropes for a decade.

The personal CRM.

I want to walk through why it broke, what just changed, and what we built for ourselves to actually make this stick. Then I want to talk about why your business has the same problem dressed up in different clothes, and what to do about it.

The graveyard — what happened to the 2010s wave

Between 2018 and 2022 the personal-CRM market filled with attempts. Most of them are dead, pivoted, or acquired now.

• Garden was abandoned by its solo founder Zander Adell sometime in 2020. The App Store listing is still up; the backend has been frozen for years.
• FollowUp got acquired in 2018 and pivoted away from personal CRM into an email-reminder tool.
• Nouri turned into an events-platform-meets-social-CRM hybrid at nouri.ai. Same team, different category.
• UpHabit pivoted to selling enterprise “Relationship Selling” on the Salesforce AppExchange. The category exit dressed up as a product evolution.
• Clay got acquired by Automattic in June 2025 and rebranded as Mesh. Co-founders Matt Achariam and Zach Hamed kept building under Automattic’s “Other Bets” division, alongside Beeper, Gravatar, and Tumblr. The clean exit in a category full of corpses.
• Dex, Nat, and Covve are still hanging around. None of them broke out, but none of them died either.
• Monica was open-sourced in May 2017 by Régis Freyd in Montréal. Today the GitHub repo sits at tens of thousands of stars, with tens of thousands of registered users and hundreds of thousands of contacts being managed across self-hosted instances (figures vary as the project grows). The quiet open-source survivor that outlasted most of the funded competitors.

What broke the category isn’t a mystery, and it’s the same thing Morrill named in 2018: capture is a tax, recall is a search bar. The “second brain” cost more to maintain than the brain it was supposed to extend. Most operators set up the tool, fed it for two weeks, and watched the data go stale until they quit using it.

A 2024 Ask HN thread on whether personal CRM is a tarpit idea had the cleanest articulation: built-in apps win because “a person can use them right away after the talk.” Proximity of capture is the whole game. The friction isn’t a UX problem you can patch. It is structural.

This is what Robin Dunbar’s research on social cognition has been telling us for over thirty years: humans can maintain roughly 150 active relationships at varying depth, and the constraint isn’t storage. It’s computation. Holding the kid’s name, the dog’s surgery, the role they’re considering leaving for, for 150 people at the right cadence, exceeds what the equipment was built to do. The prosthetic was always going to need to be real.

For ten years it wasn’t.

What changed — the input rail finally got rebuilt

Three things happened in the last eighteen months that together fix the part of personal CRM that was structurally broken.

Model Context Protocol (MCP) is an open specification, released by Anthropic in late 2024, that lets AI assistants talk to your own systems through a clean structured interface. Practically: it is a mapping that lets an AI machine learn how to use a particular application or website without having to look at it and click through it the way you and I do. Released open, adopted across vendors, mature enough by 2025 to support a real ecosystem.

Open-source MCP servers for Monica. Three independent developers shipped them within twelve months of each other:

• Jacob-Stokes/monica-mcp — TypeScript, exposes 21 of Monica’s REST operations directly to any MCP-speaking assistant.
• jimternet/monicahq-mcp — Java, exposes 122 categorized operations across 23 entity types if you want the heaviest coverage.
• unbraind/monica-cli — TypeScript, optimized specifically for LLM agents with capability probes and a read-only-by-default safety flag.

None of these is venture-funded. They are independent contributors who saw the same problem and shipped the bridge.

The Agent Skills format, which Anthropic published as an open specification on December 18, 2025. The compatibility showcase currently lists around forty clients — including Cursor, GitHub Copilot, VS Code, OpenAI Codex, Gemini CLI, Goose, OpenHands, and Claude Code — though the spec itself remains under Anthropic’s stewardship rather than a neutral standards body. The practical implication for an operator: write a Monica skill once, deploy it across every assistant you use, get the same workflow surface in whichever tool happens to be in front of you that day. That portability is the part that wasn’t possible eighteen months ago.

What these three add up to, in operator terms: the capture surface is no longer a web form. It’s a conversation. And it can be supercharged with automation.

The technology friction of the past is no longer an issue.

What we built — seventy seconds of friction down to five

I rebuilt my own relationship-memory layer on top of self-hosted Monica using this stack. I want to be honest about where it stands: it isn’t perfect, but it is lightyears ahead of even just two years ago. The wiring itself took me about two hours from my starting point (Monica was already self-hosted, the dev stack was already configured). Your mileage will vary — if you’re starting from zero, expect more.

There are three active workflows I run through with my assistant, sitting on top of a handful of automations that capture and feed the data behind the scenes:

Morning relationship-check. Thirty seconds, conversational. “Who am I overdue with this week? Who in the inner-five is quiet? Who in the next layer is due for outreach? Anything coming up in the next seven days I should be ready for?” The assistant runs the queries, returns three lines, and I act on one of them. That is the entire daily loop.

Quick re-cap after every meeting. Coffee with a client, a long phone call, a walk-and-talk that produced something worth remembering. They mentioned a hiring problem, their kid is starting somewhere new, a project is wrapping up that has implications for what they’ll buy next quarter. I tell the assistant what happened. It maps the fields, picks the right person in Monica, files the activity. I never see the form. If there are notes attached to the meeting in calendar or document form, those get referenced during a nightly summarization pass that updates the relevant contact records.

Pre-meeting brief, delivered to my phone. About an hour before I walk into a meeting (or the morning of, depending on the calendar), a brief lands on my phone. It is not about the meeting. It is about the person in the meeting. Whether it’s coffee or a business proposal: what did we last talk about? What is important to them? Preferences, interests, pet names, what I committed to last time. I show up actually knowing.

None of these workflows are exotic. They are the three Dunbar cadences in action (daily inner-circle, weekly mid-layer, event-driven outer) running inside the conversational interface I am already in all day for everything else. No context switch. No “later, when I have time.”

That last part is the bet.

The bottleneck on AI-augmented operations isn’t the AI

This is the same lesson we keep running into across every AI-augmented operations problem we work on at the studio, and it is worth saying directly.

The bottleneck isn’t the AI.

It’s YOU.

Sometimes it’s the technology too — but mostly it’s you. It’s the briefs you haven’t written. The isolation of this conversation from the rest of your work that nobody set up. The integration plumbing your “architect” didn’t put in place. The discipline your leaders failed to follow so the system can actually get out of the way and do its job.

The AI part is the cheap part. The pipes are the hard part. And the hard part isn’t expensive in tooling. It is expensive in discipline, in clarity of intent, and in the willingness to write down what “done” actually looks like before turning a worker loose on it.

This isn’t a tools problem. It is a leadership problem dressed up as a tools problem. We see it on every engagement.

Where else this pattern applies — the same shape, different clothes

The Monica case is one shape of the pattern. The same shape applies to almost every relationship-heavy operations problem that small and mid-sized businesses are sitting on right now:

• A sales pipeline that lives in spreadsheet purgatory, or worse, split across two different CRMs that don’t talk to each other.
• A donor list scattered across four systems with no canonical source of truth.
• A vendor network and employee workforce where you know everyone but couldn’t tell a new hire who actually does what.
• A student roster in one system with the engagement notes living in another.
• A patient registry where the clinical record and the relationship context don’t share a surface.
• A lead funnel where the meeting notes, the follow-ups, and the qualifying signals exist in three different inboxes.

Every one of these is a relationship-memory problem dressed up as something else. Like I said at the top — it’s the data. Always the data. And every one of these problems is one Agent Skills implementation away from being a tool that actually works — with the caveat that “implementation” also means the unglamorous work around it: data cleansing, compliance (HIPAA for patient data, FERPA for student data), change management, and integration with whatever systems of record already exist. The pattern is portable. The supporting work is real.

The honest objection — what about surveillance?

There is a critique I have to address before this reads like cheerleading.

The argument is that an AI-augmented personal CRM is a kind of surveillance dressed up as friendship — you aren’t befriending people, you’re studying them. Genesis Elizabeth Benedith wrote it cleanly on Substack. The broader academic frame is laid out in Shoshana Zuboff’s The Age of Surveillance Capitalism (2018): when one party in a relationship has the capacity to record, retain, and selectively deploy the shared history in ways the other cannot, the relationship’s terms quietly shift.

She isn’t wrong. The play exists. I have seen people run it, and it feels gross from the receiving side.

But here is the reframe. Showing up for someone (actually remembering what they told you, asking how their dog is, knowing their kid’s name) is not surveillance. It is the thing we used to expect from each other before we all got too tired and too distracted to keep up. Surveillance is a posture. The same notebook can be a dossier or it can be a card you send on someone’s anniversary. The notebook doesn’t pick.

You do. And the tool quietly makes either one easier, so pretending the asymmetry doesn’t exist would be dishonest.

We built this so we can show up better for the people whose details we would otherwise lose. If we notice the system training us toward extraction instead of attention, we will say so out loud, in public, on this blog. That is the deal.

The shape was right the whole time

None of this is finished. The skills layer on top of Monica is still being written; we are months into a working version, not years into a hardened one. The receipts will get longer as the system runs.

What I can say now is the shape was right the whole time. Tiago Forte gave us the Building a Second Brain framework for offloading ideas. Nobody figured out how to apply the same architecture to people. The capture rail was wrong, and the wrong capture rail killed the thing.

With the rail fixed, the rest of the architecture is older than most of us. Andy Clark and David Chalmers laid the philosophical groundwork in The Extended Mind (1998): an external tool functioning like memory under normal use is part of memory. Dunbar laid the empirical groundwork through the 1990s. Vannevar Bush laid it all the way back in 1945. Bush called the device a Memex. He was talking about documents, but he saw the shape.

It just took eighty years to find the right input device.

Technology is here. Today is the future. The infrastructure exists. The pattern is portable. The leverage is here.

The only question left is whether you’re going to do anything about it.

Related reading

• Our AI development services — building the kind of agent + MCP + Agent Skills scaffolding above for real business processes (relationship management, lead routing, inbound triage, knowledge-base authoring).
• Our automation services — the operational scaffolding side: briefs, isolation, audit trails, guardrails that make AI safe to deploy on your day-to-day work.
• The hour that cleared the backlog — sister piece in this case-study series. Same lesson (the bottleneck is the scaffolding, not the AI) applied to a different operational problem.