Meeting assistants with continuity

Most of what happens in a meeting dies in the recording. Decisions get made, action items get assigned, context gets established, and then it all goes into a video file nobody rewatches and a transcript nobody reads. The next meeting starts as though the last one didn't happen, and the same ground gets covered again.

The model can summarise a transcript well enough. The harder thing is continuity: turning the audio and video into usable text, making every meeting searchable so a decision from six weeks ago can actually be found, and giving the assistant memory of what was decided across all of them so it carries the thread from one meeting to the next. This page is about that infrastructure.

The problem

The first barrier is that meetings are audio and video, not text. Before an assistant can do anything with what was said, the recording has to be turned into a clean, accurate transcript, ideally with timestamps so a point can be traced back to the moment it was made. That alone stops a lot of would-be meeting assistants at the starting line.

The second is that a transcript on its own is nearly as inert as the recording. A single meeting produces a long wall of text, and a series of meetings produces far more than a context window can hold. Without search, finding what the CEO said about pricing in the Q3 all-hands means scrubbing through an hour of transcript, and keyword search over transcripts is unreliable because people rarely phrase things in the meeting the way you'd phrase the question afterward.

The third is continuity itself. Decisions get revised, action items get reassigned, plans change from one meeting to the next. An assistant that just stores every transcript has a record of all of it but no sense of what's current. It can't tell that a decision made in March was reversed in April, only that both were said. Keeping the current state of decisions and commitments straight across many meetings is an entity resolution problem, and an assistant that surfaces a reversed decision as though it still stands is exhibiting memory drift of the most confusing kind.

What Exabase unlocks

With transcription, search, and cross-meeting memory underneath, the assistant turns a pile of recordings into something the team can actually use.

Every meeting becomes text automatically, with timestamps, so the assistant works from an accurate record rather than a video file. Ask "what did the CEO say about pricing in the Q3 all-hands?" and you get the answer with the timestamp where it was said, not a suggestion to rewatch the recording. The meeting stops being a black box.

The assistant carries decisions and action items forward. It knows what was decided last time, what was assigned to whom, and what's still open, so a meeting opens with the assistant already aware of where things stand rather than the team reconstructing it from memory. When a decision gets revised, the assistant reflects the change rather than holding the old and new versions as equally valid.

And the whole history becomes searchable by meaning. A question about something discussed weeks ago, across any of dozens of meetings, returns the actual moment it came up, with the context around it. The institutional memory that usually evaporates after each call becomes a queryable record.

How it works

Three primitives carry a meeting assistant, with Extract leading because everything starts as audio and video.

Extract

Extract turns audio and video into clean text through one API. Recordings go in and you get back a transcript chunked with timestamps, so any point in an answer can be traced to the moment it was said. It handles the long recordings meetings produce, and it's the step that makes everything else possible, since nothing can be searched or remembered until the recording has been transcribed. This audio-and-video handling is the same capability behind searchable podcast and audio libraries and video content search.

Deep Search

Deep Search is how the assistant finds the right moment across every meeting. It searches the transcripts at the paragraph level, semantically rather than by keyword, so a question surfaces the relevant passage even when the words spoken in the meeting differ from the words in the question. Because the transcript chunks carry timestamps, a result points to exactly when something was said, turning "somewhere in this hour" into a precise reference.

Memory

Exabase Memory holds what carries across meetings: decisions made, action items assigned, commitments given, the running state of the things the team is working through. You send the transcripts in and Exabase extracts what's worth tracking, and when a decision is revised or an action item reassigned, the new state supersedes the old rather than coexisting with it. That contradiction handling is what gives the assistant a current view rather than a contradictory pile, and it's the line between a memory layer and an archive of transcripts.

Example architecture

The pipeline is clear, and the same steps cover one team's meetings or many.

Scope sensibly. Most teams keep their meetings in one workspace, or use a Base per team or project to keep separate groups' meetings apart. If you're building a meeting product for many customers, give each their own Base, following the multi-tenant SaaS pattern.

After each meeting, run the recording through Extract to get a timestamped transcript, store it as a Resource so it's indexed for search, and send it to Memory so decisions and action items get extracted and kept current.

Before or during a meeting, the assistant retrieves memory for the current state of decisions and open items, and runs a Deep Search across past transcripts when someone asks what was said about something. Both feed the assistant's context.

Recordings flow in through transcription, become searchable Resources, and feed a memory of decisions that stays current across meetings. The assistant reads from both, so each meeting builds on the last instead of starting over.

What compounds over time

A meeting assistant built this way gets more useful the longer a team uses it, because continuity is cumulative by nature.

Every meeting adds to the searchable record and updates the running state of decisions and commitments. After a few months the assistant holds an accurate, queryable history of how the team's thinking evolved, what was decided and when, what changed, what's still open. Because the memory self-organises, that history stays coherent as it grows, with revised decisions updating rather than piling up as contradictions. The searchable transcript archive compounds alongside it: every meeting transcribed once stays findable forever.

The do-it-yourself version, a transcription service wired to a storage bucket wired to a search index you maintain, gets more brittle as the volume grows, and it still wouldn't give you the cross-meeting memory that tracks current state, which is the genuinely hard part. Infrastructure where transcription, search, and contradiction-resolved memory come together means the record gets richer and stays current while the work of running it stays flat.

Who's building this

Teams building meeting assistants, notetakers, sales-call tools, and team-knowledge products, anywhere an agent works across a series of recorded conversations and needs to find what was said and carry decisions forward.

The audio-and-video side overlaps closely with searchable podcast and audio libraries and video content search, both worth reading for the extraction patterns. If meeting transcripts feed a sales workflow, the sales copilots use case covers how that context flows into deal memory.

Get started

Start with the getting started guide, then about extraction and submitting jobs for the transcription side, and creating memories for the cross-meeting continuity. There's a free tier to build against.

FAQs

Can it transcribe both audio and video recordings?

Yes. Extract handles audio and video through one API and returns a transcript chunked with timestamps, so a point in an answer can be traced back to the moment it was said.

How do I find a specific thing someone said weeks ago?

Deep Search searches across all stored transcripts by meaning, and because the chunks carry timestamps, a result tells you exactly when it was said. A question like "what did the CEO say about pricing in the Q3 all-hands?" returns the moment, not a suggestion to rewatch.

How does it handle a decision that gets reversed in a later meeting?

The revised decision supersedes the earlier one through entity resolution. The assistant reflects the current state rather than presenting both versions as equally valid, so it won't surface a reversed decision as though it still stands.

Does the assistant remember decisions across meetings, or just within one?

Across all of them. Decisions, action items, and commitments go into Memory as each meeting is processed, so the assistant opens every meeting aware of where things stand from the ones before.

Can this feed into other tools, like a sales workflow?

Yes. Meeting transcripts and the decisions extracted from them can flow into other use cases. For sales calls specifically, the sales copilots use case shows how meeting context becomes part of a deal's running memory.

How accurate are the timestamps?

Transcripts are chunked with timestamps from Extract, so each searchable passage carries the point in the recording it came from, which is what lets an answer cite when something was said rather than only that it was.