Sanden and I have spent the last couple of years building software for meetings, and the thing I've become most convinced of is that meetings aren't the interesting part.

The interesting part is that companies forget.

Every company I talk to has more meeting data than it has ever had. Every call can be recorded. Every conversation can be transcribed. AI can summarize what happened, extract the action items, and drop a tidy set of notes in Slack four minutes after the call ends.

And they still forget. They forget why a decision was made six months ago. They ask a customer the same discovery question twice. Engineering rediscovers a constraint that was already litigated in a roadmap debate last spring. A new PM reconstructs history by asking around. Someone remembers an important tradeoff came up in a meeting, but not which meeting, not who raised it, and not what happened next.

We didn't need better notes. We were treating conversations as documents when we should have been treating them as memory.

Companies are built in conversation

Before Earmark, Sanden and I were at ProductPlan, shipping product and running teams. What struck me then - and what customers confirm now, over and over - is how much of what makes a company valuable never originates in a system. It originates in a room.

A customer explains why they almost churned. An engineer names a constraint nobody had considered. A PM connects feedback from three different accounts. An exec changes what matters this quarter. A rep hears the same objection for the fifth time and finally believes it.

That's where companies learn. But every system of record we have is designed to capture the output of learning, after the fact. Someone writes the ticket. Someone updates Salesforce. Someone drafts the PRD. Each step compresses the original conversation a little more. The reasoning falls away first, then the connections between conversations, and by the time anything reaches a system of record, most of what made it valuable is gone.

That's the "about the work" work that buried us at our last company. It's also the reason we started this one.

Recording everything didn't fix it

The first generation of AI meeting tools made capture nearly free. That mattered. It was a real step. But recording, transcribing, and summarizing mostly traded one problem for a bigger one: enormous archives that humans still have to navigate.

A company with thousands of meetings does not benefit from having thousands of transcripts. Nobody wants a second inbox. And even a genuinely good summary has the same ceiling - it represents one meeting.

Companies don't operate one meeting at a time. The question is almost never "what happened on Tuesday's call?" It's "what are customers consistently telling us about onboarding?" or "why did we pick this architecture?" or "what's changed since the last account review?"

Those questions can only be answered across conversations, not inside one.

Conversation as memory

So imagine every meaningful conversation at your company contributing to one shared memory instead of another folder of recordings.

That memory knows what the organization has discussed, learned, decided, promised, questioned, and changed its mind about. When a customer names a problem, it doesn't evaporate when the call ends - it becomes part of what the company knows about that customer, that product area, that problem. When a team makes a call, the decision stays attached to the reasoning and evidence behind it. When the same theme shows up in five customer conversations, the pattern surfaces without anyone rereading five sets of notes. When someone asks a question in November, the answer includes why, not just what.

That's a fundamentally different object than meeting notes.

Memory compounds; notes don't

This is the part I find genuinely exciting.

Traditional meeting software produces independent artifacts. Ten meetings, ten recordings, ten summaries. The tenth is worth exactly as much as the first.

Memory works the other way. The tenth conversation should make the first nine more useful. A new customer interview validates a pattern that was previously just a hunch. An engineering discussion explains a constraint someone waved at months earlier. A pricing call retroactively gives meaning to a quarter of sales objections.

Context compounds, because every new conversation improves your understanding of everything that came before it. Eventually the system answers questions no single meeting could answer, and starts to represent not what was said, but what the company has learned.

Worth saying plainly: this wasn't economically possible until very recently. When we started running multiple agents live against real conversations, a single meeting cost us about $70 in inference. We got that under a dollar through caching and a lot of unglamorous optimization work. That number is why "remember everything, connect everything" was a nice idea in 2023 and is a product in 2026.

This is what agents have been missing

The current wave of AI agents is wildly capable and mostly context-starved.

Give an agent a generic prompt and you get a generic assistant. Give it your company's accumulated context and it starts behaving like a colleague - one who knows that this enterprise customer raised the same security concern in May, that the feature was deprioritized on purpose and here's the reasoning, that the proposed solution quietly contradicts an architectural decision from last quarter, that this piece of feedback is the fourth data point on an existing hypothesis.

At that point memory stops being retrieval and becomes infrastructure. The interesting future isn't an AI that can summarize your meeting. It's an AI that understands enough of your company's history to be useful during the next one.

The meeting is where the work should happen

There's one more assumption I think is on its way out: that AI gets to be useful only after everyone hangs up.

Nearly all meeting software follows the same shape. Have the meeting. End the meeting. Generate the summary. Figure out what to do with it. But if the system understands the conversation while it's happening, the meeting stops being a place where work is described and becomes a place where work gets done.

A product conversation can produce the spec while the details are still being argued. A customer call can update the account context while the customer is still explaining the problem. A technical discussion can capture the decision and its reasoning before anyone leaves the room.

That's what we build at Earmark, and it's why we refuse to put a bot in your call — the assistant should be part of the conversation's infrastructure, not a guest in it. The line between talking about the work and doing the work starts to dissolve. Instead of generating notes that remind humans to do work later, the conversation produces the artifacts, the context, and the actions directly.

The end state isn't better notes

We started by thinking hard about meetings. We even started somewhere stranger than that - our first product was a presentation coach for the Apple Vision Pro, which is a story for another post. What we kept running into, from every direction, was the same thing: conversations hold an enormous share of a company's intelligence, and virtually no software we use is designed to preserve that intelligence in a form that gets more valuable over time.

I think that changes over the next few years. Companies move from recording conversations to understanding them. From storing transcripts to building memory. From isolated summaries to shared organizational context. From AI that knows what happened in a meeting to AI that knows what the company knows.

Meetings shouldn't become notes. They should become memory that compounds.