M Blog is introducing a new AI agent skill called meeting-ingestion. This skill processes meeting recordings from various sources – AI notetakers, voice memos, video transcripts, or even pasted notes – and transforms them into structured brain pages. It includes attendee enrichment, entity linking, and timeline entries, making meeting data more useful for anyone who needs to quickly refer to past discussions.
How meeting-ingestion Works
The core function of this skill is to take disparate meeting data and integrate it into a coherent record. The process begins by normalizing the input into a standard transcript record, regardless of its original format. A key capability is detecting and splitting multi-meeting recordings into their distinct sessions, ensuring each meeting is treated individually. To maintain data integrity, the skill deduplicates content across multiple recorders if the same meeting was captured by several tools.
Speaker identities are resolved by evidence within the transcript, never through guessing. Once speakers are identified, the skill creates a dedicated meeting page in the brain. It then verifies claims made in the meeting against existing brain truth, adding a layer of accuracy. All attendees and relevant entities mentioned in the meeting are enriched with information from the knowledge base. Before finalizing, the skill runs a mandatory verification checklist, ensuring all steps are complete and accurate. A critical rule is that one brain page is created per real meeting; any recorder summaries are treated as claims to verify against the transcript data, not as standalone records.
Supported
Inputs and Triggers
meeting-ingestion is designed to handle a broad array of input sources. This includes output from various AI notetakers, personal voice memos, video transcripts, and even plain pasted notes. The goal is to make it simple to get meeting data into the system, no matter where it originated.
To initiate the skill, specific triggers are recognized: 'meeting transcript', 'process this meeting', 'meeting notes', and 'ingest this recording'. These phrases prompt the agent to begin its processing pipeline. For example, if a team just concluded a virtual meeting and an AI notetaker generated a transcript, a user could simply provide that transcript to the agent with the command 'process this meeting'. The skill would then take over, normalizing the text, identifying participants, linking entities, and creating a structured brain page for that specific discussion.
It is important to note that the skill is not intended for raw unstructured transcript files; for those, the 'capture' skill is more appropriate. Also, summary-only payloads are rejected because the system requires the full transcript to perform its verification and enrichment processes accurately.
The
Toolchain and Core Principles
Underpinning the meeting-ingestion skill is a set of robust tools that enable its detailed processing. These include search and query for retrieving existing information, get_page and put_page for interacting with brain pages, add_link for creating relationships between entities, and add_timeline_entry along with get_timeline for managing chronological events. This suite of tools allows the agent to construct rich, interconnected meeting records.
The skill operates on strict principles to ensure data quality. Speaker identification relies purely on concrete evidence within the recording or associated metadata, preventing erroneous assignments. Every claim made in the transcript is evaluated against known truths in the brain, which helps maintain the integrity of the knowledge base. The mandatory verification checklist serves as a final quality gate, confirming that all required data points have been processed and cross-referenced correctly. This focus on verification and evidence-based processing ensures that the structured meeting pages are reliable records.
FAQ
Q: What kind of input does meeting-ingestion accept? A: It accepts meeting recordings from various sources, including AI notetakers, voice memos, video transcripts, and pasted notes.
Q: Can I use it for just a meeting summary? A: No, summary-only payloads are rejected. The skill requires actual transcript data to perform its verification and enrichment processes.
Q: How does it handle multiple recordings for the same meeting? A: The skill deduplicates across recorders and ensures that only one structured brain page is created for each real meeting.
This skill provides a structured approach to managing meeting data, making the information more accessible and verifiable. It helps ensure that meeting outcomes are accurately reflected and connected within your knowledge base.
