Every team we talk to has the same ritual. Someone asks "where did we land on that?" - and the answer is twenty minutes of scrolling through old notes, searching Slack, and finally pinging the one person who was in the room. The knowledge exists. It was said out loud, in a meeting, probably captured somewhere. But captured isn't the same as available. That gap is what our new release closes, and closing it properly turned out to require rebuilding how Earmark thinks.

Why a search box wasn't enough

Our first instinct was the obvious one: semantic search over transcripts. Embed the question, fetch the closest snippets, summarize. We built it - and it failed in ways users notice immediately. Real questions aren't lookups. "What did we decide about pricing, and did anyone push back?" spans several meetings, needs different levels of depth in each, and has a right answer. A pile of similar-sounding snippets isn't that answer.

So we built an agent that investigates the way a diligent chief of staff would: discover the candidate meetings, shortlist the ones that matter, read them at the right depth, and only then answer. Most questions resolve from the artifacts Earmark already wrote during your meetings - summaries, notes, pins. The raw transcript is reserved for when exact wording or who-said-what actually matters. That discipline is why cross-meeting answers arrive in seconds: the agent skims broadly, reads selectively, and deep-reads almost never.

Why every answer shows its work.

An AI that answers questions about your business is only useful if you can act on the answer without double-checking it - and the moment it invents one decision that was never made, you'll double-check everything forever. Trust doesn't survive a single hallucination. So we made honesty structural rather than aspirational. Citations are constructed from the evidence the agent actually read - it cannot reference something it never saw. Every answer ends with a coverage note: what was searched, what was read, what's missing. And when the evidence isn't there, Earmark names the gap instead of inferring through it. No invented owners. No invented decisions. No invented deadlines. The same rule runs through everything Earmark writes: every claim in an artifact must be traceable to what was actually said.

Why multiplayer had to be enforced, not suggested

Here's the thing about shared memory: it only works if the boundaries are real. Teams won't pour their most important conversations into a system where "who can see what" is fuzzy. So a Project isn't a label - it's a scope enforced in code on every single read. Ask a question from a project and the agent cannot see outside it. Invite a collaborator and they get the same quality of answer as the person who was in the room, because their question runs the same investigation over the same evidence with the same receipts. That's the promise behind "the same fidelity as if you'd attended" - it's not a summary of a summary, it's the full machinery working on your behalf.

And team memory can't have a single point of failure. Once a meeting is added to a project, it stays with the team - even if the person who captured it changes roles or moves on. Institutional knowledge shouldn't leave in someone's backpack.

Why we test it like we mean it

A system this dynamic can degrade invisibly - a small improvement for one kind of question quietly breaks another. Your memory is not the place to find that out in production. So we benchmark Earmark's memory against a hand-built corpus of meetings where every answer is known in advance - including the questions whose correct answer is "that was never discussed." Every change is scored against that baseline before it ships. If a change would make answers worse, it doesn't merge. You'll never see this system, but you'll feel it: the answers stay trustworthy as the product evolves.

The point of all of it

None of this machinery is the feature. The feature is what it makes possible: you no longer have to be in the meeting to stay in the know. What one person captures, everyone can use. And every conversation your team has makes the next answer, the next draft, and the next decision faster.

Meetings that don't just end - they compound.