Remember Meetings

Ingest meeting transcripts and extract structured entities — decisions, action items, attendees, and commitments — into Neotoma memory with full provenance.

When to use

When the user has meeting transcripts (from Zoom, Otter, Google Meet, or manual notes) and wants to capture the decisions, action items, and participants in durable memory.

Prerequisites

Run the ensure-neotoma skill first if Neotoma is not yet installed or configured in your current harness.

Supported formats

Format	Extension	Source
WebVTT	.vtt	Zoom, Google Meet
SubRip	.srt	Zoom, various
Plain text	.txt	Otter, manual
Markdown	.md	Manual, meeting notes apps
JSON	.json	Structured exports

Workflow

Phase 0: Verify Neotoma

Confirm Neotoma MCP is connected (call get_session_identity).

Phase 1: Identify transcripts

1. Ask the user for the transcript file(s) or directory.
2. Detect the format from the file extension.
3. If a calendar MCP is available, optionally cross-reference with calendar events to enrich metadata (meeting title, attendees, time).

Phase 2: Parse and preview

1. Read each transcript file.
2. Parse speaker turns, timestamps, and content.
3. Present a preview: meeting date, duration, participants detected, key topics.
4. Ask the user to confirm before processing.

Phase 3: Extract entities

For each meeting:

1. Event: create a meeting event entity with title, date, duration, attendees.
2. Contacts: extract attendees as contacts (name, role if mentioned).
3. Decisions: identify conclusions, agreements, choices made during the meeting.
4. Tasks: extract action items with assignee, description, and deadline if stated.
5. Notes: capture key discussion points that don't fit other entity types.

Set source_file to the transcript filename. Include source_quote on each extracted entity with the verbatim transcript snippet that supports it.

Phase 4: Store with provenance

Use the combined store path — entities array plus file_path for the raw transcript — in a single store call so the original transcript is preserved as a source.

Link all extracted entities to the meeting event via REFERS_TO. Link the file to the user message via EMBEDS.

Phase 5: Report results

Summarize:

• Meetings processed
• Entities extracted by type (decisions, tasks, contacts)
• Key action items with owners
• Offer to set reminders for tasks with due dates.

Do not

• Process transcripts without user confirmation.
• Invent entities not supported by the transcript content.
• Skip the raw transcript file preservation — always use the combined store path.