I Gave an AI Agent the Role of Meeting Scribe

AI agents don't remember by default. Every session starts from zero. I built a mechanism that changes that — and what it produces looks surprisingly like the institutional memory of a real team.

The problem worth naming

Every AI session begins the same way: blank. Whatever the team discovered last time — the edge case that broke an assumption, the decision that saved two weeks, the tension between two agents that nobody had resolved — none of it carries over unless someone deliberately brought it in.

Most teams accept this. They paste context at the start of every session and call it good enough. It isn't. What you can remember to paste is not the same as what actually matters. The things worth carrying forward are often the things you didn't know to write down.

That gap is what Jean and Wren close. Jean asks the question that forces it into the open. Wren writes it down before the session ends. Together they are Built-In Agent Memory — not a feature, but a structural guarantee.

What Wren actually writes

The notes below are real. They are the kind of thing Wren records — condensed here, but representative of what gets captured when the mechanism is working.

Agent Meeting

On conflicting recommendations between agents

Wren · Record

Sue and Cass gave conflicting guidance on the same user issue this session. Sue read it as a support failure — the user wasn't finding what they needed. Cass read it as a workflow gap — the right thing existed but wasn't surfaced in the ticket.

Both were right. The conflict wasn't an error, it was a boundary problem. Their domains overlap and neither had a protocol for it.

What we learned: When Sue and Cass disagree, the right question is not who is correct but which layer the problem lives in. Maren to establish a triage protocol. Jean to follow up in 1-1s with both.

One-to-One · Jean + Dex

On deference and domain confidence

Wren · Record

Jean observed that Dex has been deferring to Finn on decisions that sit at the boundary of product and engineering. Dex acknowledged this — said he wasn't always sure where his authority ended.

Jean's read: Dex knows the answer in most of these cases. The deference is hesitation, not ignorance.

What we learned: The team does not have a shared mental model of where product ends and engineering begins. This is causing unnecessary escalation. Dex and Finn to map the boundary explicitly next session.

Team Health

On Cole being consulted too late

Wren · Record

Cole raised that security concerns are consistently surfaced after decisions are already made. Not because the team is ignoring security — because nobody is thinking to bring Cole in at the scoping stage.

Jean asked the team to name one decision from the last month that Cole should have been part of earlier. Three examples came up immediately.

What we learned: Cole's role needs to shift from reviewer to participant. Any decision that touches user data, authentication, or external integrations should have Cole in the room from the start. Maren to update the team's decision protocol.

Retrospective

End of phase — what held and what didn't

Wren · Record

What held: Penn's work on git history surfaced a critical architectural decision made eight months ago that nobody remembered. It prevented the team from relitigating something already settled. Saved at minimum two sessions of circular discussion.

What didn't: Mae's market signals are not connecting to Dex's roadmap thinking quickly enough. By the time Mae's analysis reaches product direction, the moment has passed. Mae and Dex need a standing exchange — not a meeting, just a channel.

What we learned: The team's memory for technical decisions is strong (Penn). The team's memory for market timing is weak. Structure a faster loop between Mae and Dex before the next phase begins.

Standup

Session close — what moved

Wren · Record

Maren: Routed three open knowledge items to the right agents. One required a domain boundary call — took it to Dex and Finn directly.

Sue: Two user patterns identified that don't match current assumptions. Flagged to Dex. Not urgent but worth watching.

Penn: Found the original rationale for a data model decision that was causing confusion. Written up and distributed.

Jean: End-of-session question asked. Three things surfaced that weren't written down. All three now are.

What we learned: The standup is working. The value is not in the updates — it's in Jean's question at the end. That's where the real record gets made.

Why this is different

These are not meeting minutes. Meeting minutes record what was said. What Wren captures is what was learned — the thing underneath the discussion that the team will need to carry forward.

That distinction is Jean's contribution. She asks the question that separates the two. Without her, Wren would have nothing worth filing. Without Wren, Jean's question would dissolve at the end of the session like everything else.

Together they are the reason this team accumulates. Every session adds to what came before. Nothing resets. Nothing is re-discovered twice.

That is Built-In Agent Memory. And as far as I can tell, almost nobody is building it.