MEMORY NOTCH BLOG

Claude meeting notes from transcripts you keep

Claude writes solid minutes if you give it the transcript: paste it into a chat or point Claude Code at a folder. Either way the text leaves your Mac.

Published September 28, 2026 · The Memory Notch team

Can Claude write meeting notes?

Yes, but only from text you give it. Claude meeting notes start with a transcript: paste one into a chat, or open a folder of them in Claude Code, and ask for minutes. Anthropic opened its Claude Marketplace on September 23, 2026. Two days later it published a guide to building plugins. Neither names a meeting-notes tool, so you still bring the transcript.

Paste one transcript into a chat with a minutes prompt. There is nothing to install.
Or run Claude Code in the folder that holds your transcripts, with a short instruction file that sets the format.
Either way, the text goes to Anthropic’s servers, even when the file sits on your Mac.
Wrong words or missing speakers in the transcript carry straight into the minutes.

	Claude chat	Claude Code on a folder
What Claude reads	The transcript you paste or attach	Files in the folder you open
Setup	None	Claude Code, plus a `CLAUDE.md` or skill
Suits	One meeting	Many meetings, repeat formats
Sent to Anthropic	What you paste	Every file Claude opens

What changed this week, and what it means for your notes

On September 23, 2026, Anthropic opened Claude Marketplace. It lists "more than 2,000" connectors and plugins from other companies. A connector links Claude to an online service so it can fetch data from there.

Connectors are built on the Model Context Protocol (MCP), an open standard. On September 25, Anthropic’s post on building plugins said a plugin packages connectors, skills, or both. It said Claude supports the latest version of the standard, "commonly referred to as MCP 2.0", and promised one place to find plugins across its apps "over the coming weeks".

Neither post names a meeting recorder or note taker (checked September 28, 2026). If your meetings are recorded on your own Mac, there is no service to connect, and you don’t need one, because plain text is enough. A skill, the other half of a plugin, is a text file you can write yourself. The example below is 17 lines.

What a good transcript looks like to Claude

Minutes come out better when each line of the transcript says who spoke and when. Give the model a wall of words with no speakers and it has to guess who agreed to what.

Memory Notch, the Mac app we make, saves each recording as a folder of plain files. Here is one, with fictional sample data. Every block in the .txt transcript starts with a timestamp and a speaker number:

~/memory/notes/2026-08-15-1432-kitchen-extension-quote/
  2026-08-15-1432-kitchen-extension-quote.md    metadata
  2026-08-15-1432-kitchen-extension-quote.txt   transcript
  2026-08-15-1432-kitchen-extension-quote.m4a   audio

[0:04] [speaker 1] So where did we land on the kitchen extension quote from Elias?

[0:09] [speaker 2] He came in at 34,000 for six weeks, but that excludes the steel.

Speakers are numbered, not named. The prompts below keep the numbers unless you say who "speaker 1" is, which beats a wrong name. A transcript from any other tool works the same way, as long as it carries speakers and times.

Route one: paste a transcript into a Claude chat

Start a new chat, paste the transcript (or attach the .txt file), then paste the prompt below under it. An ordinary chat reads only what you give it. Typing a path such as ~/memory does not let it open your files.

You are writing meeting minutes from the transcript above.
Speakers: speaker 1 = [name], speaker 2 = [name]. Leave any others as numbers.

Write, in this order:
1. One-line purpose of the meeting.
2. Decisions: only things someone clearly agreed to. Cite the [m:ss] timestamp.
3. Action items as a table: task | owner | due | timestamp.
   If no owner or date was said out loud, write "not stated". Do not guess.
4. Open questions nobody answered.
5. Numbers, prices and dates mentioned, each with its timestamp.

Rules: use only the transcript. If a word looks mis-transcribed, flag it
instead of correcting it silently. Keep it under 300 words.

The "not stated" rule does the most work. Without it, a model tends to invent a sensible owner and a Friday deadline, and the minutes read better than the meeting went. The timestamps let you check any line against the audio in seconds.

Pasting suits a single meeting. Anthropic’s guide to uploading files covers attachments. Each paste is a snapshot: Claude does not see the next recording unless you paste it too.

Route two: point Claude Code at a folder of transcripts

Claude Code can read a folder on your Mac directly. That suits weekly minutes, a month of client calls, or a question like "what did we decide about pricing". Our setup guide covers choosing the folder and keeping permissions read-only, so this post skips it.

What you add is a CLAUDE.md file in the folder you open. It loads when a session starts in that folder, according to Anthropic’s memory documentation. A minimal one for ~/memory:

# Meeting notes folder

@README.md

- Transcripts are notes/<id>/<id>.txt. The .md beside each one is metadata.
- Never edit, rename or delete anything in notes/.
- Write minutes to minutes/<id>.md, one file per meeting.
- Keep speaker numbers unless I name them in the conversation.
- Action items: task | owner | due | timestamp. "not stated" beats a guess.
- Cite an [m:ss] timestamp for every decision.

The @README.md line imports the README.md Memory Notch writes at the top of ~/memory, which already explains the layout. Memory Notch’s library lists recordings from notes/, so a minutes/ folder beside it stays out of the app’s way. Keep Claude out of notes/ for a concrete reason: the .md files carry metadata the app parses, and a bad edit there breaks that note in the library.

A meeting-minutes skill for Claude Code

A skill is a SKILL.md file that Claude Code loads when its description matches your request, or when you type its name. Personal skills live in ~/.claude/skills/<name>/SKILL.md and work in every folder, according to Anthropic’s skills documentation. Save this as ~/.claude/skills/meeting-minutes/SKILL.md:

---
name: meeting-minutes
description: Writes meeting minutes from a transcript file. Use when the user asks for minutes, a recap, decisions or action items from a meeting or recording.
---

Read the transcript named in $ARGUMENTS (a .txt file, or a folder containing one).
If none is named, list the five newest folders in notes/ and ask which one.

Write minutes with these headings: Purpose, Decisions, Action items,
Open questions, Numbers mentioned.
- Cite the [m:ss] timestamp for each decision and action item.
- Owner or due date not said aloud: write "not stated".
- Keep speaker numbers unless the user names them.
- Flag words that look mis-transcribed; do not fix them silently.

Save to minutes/<recording id>.md, show the result, and never modify the transcript.

Then, in a session started in ~/memory, type /meeting-minutes notes/2026-08-15-1432-kitchen-extension-quote, or just ask "minutes for the kitchen call". The skill and the CLAUDE.md overlap on purpose: the CLAUDE.md guards the folder, and the skill carries the format with you to any folder.

What leaves your Mac, and what to check

Claude is a cloud service, so everything it reads goes to Anthropic for processing. That holds for a pasted transcript and for a file Claude Code opens from a folder. Choosing "Local" in Claude Code decides where your files and commands live, while the model itself still runs on Anthropic’s servers. Your recording and transcript can stay on your Mac; the minutes Claude writes are produced in the cloud.

Before you send a client call or an HR conversation:

Check your privacy settings. Anthropic says it uses chats to improve Claude only if you allow it, and never uses incognito chats for training (help center, updated May 22, 2026).
Check your employer’s rules. Some workplaces do not allow meeting content in any AI tool.
Send one meeting, not the whole library, when one is all the task needs.
Check the minutes against the audio. Transcripts contain errors; the timestamps let you listen back.
Get consent first. Recording locally does not replace the permission of the people you record, and sending their words to an AI service is a second step they may care about.

Where Memory Notch fits

Memory Notch records your microphone and your Mac’s system audio from the notch, with no bot joining the call. It transcribes and labels speakers on the Mac, and writes the files shown above. It does not write summaries or action items itself; the routes above are how you get them from Claude. Speakers are numbered, not named. Overlapping voices can be difficult to separate, and system audio needs macOS Screen Recording permission.

The app makes three network requests, and none carries a recording. It fetches a signed list of revoked licences, checks for updates once a day (you can turn this off), and downloads an update when you choose one (privacy policy). It costs US$79 once, with a free five-minute recording trial and no account (get it). For how capture without a bot works, read bot-free AI note takers. For the on-device side, see local AI note takers for Mac.

Frequently asked questions

Can Claude take meeting notes during a live call?

Not from these routes. Claude works on text you give it after the fact, so you record and transcribe first, then paste the transcript or open its folder in Claude Code. Claude Marketplace (September 23, 2026) did not name a meeting-recording tool.

Is there a Claude meeting notes skill?

You can write one. A skill is a SKILL.md file with a name, a description and instructions, saved in ~/.claude/skills/<name>/ for Claude Code. The example in this post writes minutes with timestamps and marks unassigned owners as "not stated".

Does Claude keep my meeting transcripts private?

Claude processes what you send it on Anthropic’s servers, so the transcript leaves your Mac. Anthropic says consumer chats train its models only if you allow it in your privacy settings. Check that setting, and your employer’s policy, before sending sensitive meetings.

Do I need an MCP connector to use my own transcripts with Claude?

No. Connectors fetch data from online services. Transcripts saved as text files on your Mac can be pasted into a chat or read by Claude Code from a folder, with no connector, plugin or server.

Why does Claude invent owners for action items?

A model fills gaps to make the minutes read smoothly. Tell it to write "not stated" when no owner or date was said aloud. Ask for a timestamp on every action item too, so you can check it against the recording.

Sources

Download Memory Notch · More from the blog