Here is a thing that happens constantly: you open Claude, you ask for help thinking through a decision, and you realize mid-conversation that you need to give it a bunch of context it doesn't have. The backstory. The constraints you've already established. The thing you discussed with your team two weeks ago that shapes the whole situation.
So you summarize it from memory, which takes five minutes and is probably missing things. Then you get a useful answer, which you then have to mentally reconcile with the stuff your AI doesn't know. It works, but it's not what working with an AI should feel like.
The actual bottleneck is that your AI has no memory of your working life. Every conversation starts from zero. And a huge portion of your working life happens in meetings.
The amnesia problem
People spend somewhere between a quarter and a third of their working week in meetings. For founders and managers it's often more. Every one of those conversations is context, decisions, commitments, and constraints. Almost none of it makes it into a form that's useful to an AI later.
Notes, if they exist at all, are sparse summaries written by someone who was trying to participate in the meeting at the same time. The institutional memory of what was actually said lives in people's heads and slowly fades.
Meanwhile, you have an AI assistant that could reason over all of it if it could just see it.
Why cloud-based solutions don't fix this
The obvious approach is to pipe your meeting audio into a cloud service that stores transcripts and makes them searchable. A few products do this. The problem isn't that it doesn't work technically. The problem is what you have to give up to make it work.
Your meeting content is some of the most sensitive information you generate. Client discussions. Salary conversations. Strategic plans. Investor calls. The stuff you say freely in a room with people you trust is not the same stuff you'd be comfortable uploading to a third-party server for indefinite retention.
When people think carefully about this, a lot of them decide the trade-off doesn't make sense. Not because they're paranoid, but because they actually looked at the terms of service and thought about what could go wrong over a five-year horizon.
The local approach
The alternative that almost no one is building is: do it entirely on the device. Transcribe locally using on-device models. Store transcripts in a local database. Never send the audio or the text anywhere.
This was impractical for most of the history of AI, because the models weren't good enough to run locally at acceptable quality. That changed. Apple Silicon can run Whisper at better-than-real-time. The M3 in the MacBook Pro that a lot of knowledge workers already own is more than fast enough.
We built MeetingVault around this premise. Every meeting you record gets transcribed on your Mac using Whisper. Audio is discarded the moment transcription is done. The transcript goes into a local SQLite database. No internet connection required for any of it.
Connecting to Claude
Local transcripts are useful on their own. You can search them, read them, skim them. But the more interesting thing is what happens when your AI can read them too.
The Model Context Protocol (MCP) is a standard that lets AI assistants like Claude Desktop connect to local data sources. You configure it once, and Claude gains access to whatever that data source exposes. Your transcript data doesn't transfer anywhere. The MCP server reads from your local database and hands the relevant context to Claude. Claude's inference still runs through Anthropic's API as normal, but the meeting content never leaves your machine via MeetingVault.
We're building a MCP server for MeetingVault. When it's installed, Claude Desktop can access your entire transcript library. You can ask it things like:
- "What did we agree on in the Q4 planning meeting?"
- "Pull every time we've discussed pricing in the last six months."
- "What commitments did I make last week that aren't on my task list?"
- "Summarize the last three client calls before I write this proposal."
Claude answers with the actual content of your meetings because it can actually see the transcripts. Not a summary you typed from memory. The real thing.
The MeetingVault stack stays entirely local. MeetingVault writes to a local database. The MCP server reads from that database. Claude Desktop connects to the MCP server over a local socket. Your transcript content never leaves your machine via MeetingVault. Claude's reasoning still goes through Anthropic's API as it normally would, but what it's reasoning over is your local data, not a copy on someone else's server. You can verify the local portion with a network monitor.
Why this matters more than it might seem
The shift people are making with AI right now is from "occasional tool" to "persistent collaborator." That shift works a lot better when the AI has actual context about your work.
Right now, every Claude conversation is a fresh start. You spend the first part of every session giving it the backstory it needs. That overhead adds up. More importantly, it means Claude can only help you with the things you actively remember to brief it on, which is not the same as everything that might be relevant.
When your meetings become part of Claude's working context, the quality of assistance changes. It can notice patterns across time. It can remind you of decisions you made and forgot about. It can help you draft follow-up emails with full knowledge of what was actually said, not what you remember being said.
That's what an AI assistant that actually knows your work looks like.
Where we are
MeetingVault is in beta. The core app works. It transcribes meetings locally on Apple Silicon Macs, produces clean transcripts, and stores everything in a local database. Beta users have been running it for a few months.
The MCP server is in active development. When it ships, it will be open source and installable via Homebrew. One config block in Claude Desktop and you're connected. No accounts, no API keys, no data leaving your machine.
If you want early access when it's ready, join the waitlist. We'll email you when the MCP server is available and when we open up broader beta access to the app.
The version of AI assistance most people are running today is powerful but amnesiac. The version we're building has memory. It's your memory. It stays yours.