Skip to contentSkip to main content
← Back to all posts

AI Meeting Notes That Remember Decisions: A Source-Linked Workflow

Turn meeting transcripts into trustworthy AI memory with source-linked decisions, verified action items, human review, and a practical 30-minute workflow.

Share by email

A meeting ends with a friendly AI summary: “Launch next week, update onboarding, and follow up with sales.” Two weeks later, nobody can tell whether launch was approved, who owns onboarding, or which objection sales raised. The summary sounds useful but has lost the evidence needed to act.

Reliable AI meeting notes separate what someone said, what the team decided, what each person committed to do, and what remains unresolved. They also point back to the original record. This guide shows how to turn a transcript or written meeting note into durable, reviewable knowledge without treating an AI interpretation as an approved decision.

1. Preserve the source before generating the summary

Start with a transcript or written notes stored under a stable title and date. Keep the original even if you produce a cleaner decision record. A summary is a derived artifact, not a replacement for the underlying conversation.

Before recording, obtain the necessary participant consent and comply with applicable recording laws and workplace policies. If you cannot record, take structured notes with attribution instead. Minimize personal or confidential detail that the follow-up does not need.

A useful source header contains:

  • Meeting date, title, project, and location or call reference
  • Participants, including anyone whose attendance is uncertain
  • Recording or processing permission status
  • Transcript version, storage location, and access classification
  • Whether timing markers came from the source or are unavailable

Do not assign precise timestamps that were never captured. A line number can still provide useful provenance when exact timing is absent.

2. Extract five types of information, not one generic summary

Treat each candidate as a different kind of record. The categories determine what an assistant may safely do next.

RecordExample from a hypothetical launch meetingRequired verification
Observation“Three pilot users could not find the import control.”Link to the stated source
Decision“Approved: move import to the first screen.”Confirm decision-maker and effective scope
Commitment“Maya will update the design.”Verify that Maya accepted the work
Due date“Review the design on Tuesday.”Resolve the actual date and timezone
Open question“Does the change affect mobile?”Leave unresolved until evidence arrives

An AI model may infer that Maya should own a task because she is the designer. That is not the same as an explicit commitment. Mark inferred owners as unconfirmed rather than converting them into assignments.

Likewise, “We should launch next week” is a proposal unless the group clearly authorized it. Preserve the distinction between discussion, approval, and execution.

3. Make every consequential item traceable

For each proposed decision or task, capture an exact pointer to the source. A pointer might be a timestamp, transcript line, note heading, or versioned document reference.

Here is a reusable Markdown template for the reviewed output:

# Meeting memory — Import experience review
Date: YYYY-MM-DD
Project: Example product
Source: [[Import experience transcript]]
Review status: Human reviewed

## Decisions
- Decision: Move import to the first screen
  Status: Approved, pending design verification
  Authority: Product owner — confirm identity
  Source: [[Import experience transcript#line-18]]
  Effective scope: Onboarding flow only

## Commitments
- [ ] Prepare revised screen
  Owner: Unconfirmed
  Due: Not established
  Source: [[Import experience transcript#line-24]]
  Status: Proposed

## Open questions
- Does the revised flow work on mobile?
  Owner: Unassigned
  Source: [[Import experience transcript#line-31]]

The example is illustrative, not a real customer transcript. Adjust the source syntax to what your knowledge system actually supports, and test that the link resolves to the intended passage.

An effective decision record is deliberately modest. If a quote does not establish approval, write “proposed.” If the meeting did not establish a deadline, write “not established.” A blank is safer than a fabricated fact.

For deeper background, read why AI knowledge provenance matters.

4. Run a 30-minute meeting-to-memory review

Minutes 0–5: Identify the authoritative source. Check the date, project, participant permissions, transcript completeness, and whether a newer correction exists. Label incomplete recordings and ambiguous speaker names.

Minutes 5–12: Extract candidates. Separate explicit decisions, commitments, owners, dates, and unanswered questions. Capture source pointers. Do not combine unrelated statements into a single “consensus” claim.

Minutes 12–20: Verify each candidate. Reopen the cited line or timestamp. Ask: did the person actually agree? Was the decision approved or merely suggested? Does a due date have a calendar date? Are two speakers being conflated? Remove or qualify unsupported claims.

Minutes 20–25: Review the proposed write. Inspect exactly which note will be created and whether it would overwrite an existing file. Confirm the target vault and the intended audience. A read-only preview should precede any durable write.

Minutes 25–30: Save and read back. Create a linked meeting-memory note only after approval. Reopen the saved note, follow at least two source links, and confirm that tasks and decisions retained their status. Make corrections as revisions, rather than quietly changing what the source said.

For automation, this is a practical instance of a propose–review–apply workflow.

5. Understand what Phasoric supports today

Phasoric's current Meeting Memory implementation accepts a selected meeting note or a locally captured transcript, extracts recognizable structured statements, and shows candidates for review before saving them. The code recognizes explicit labels such as “Decision:”, “Action:”, “Owner:”, “Due:”, and open questions; it does not establish that every natural-language promise can be extracted reliably.

After confirmation, the local workflow writes a linked Markdown memory through a guarded Forge preview/application path. Reviewed decision candidates can become Strata records. Commitments are represented in a Forge follow-up proposal, while open questions can become Pulse signals. The application also includes local capture recovery and an optional calendar-event link.

These are implemented workflow paths in the application, not guarantees of flawless transcription, automatic identification of real speakers, or background execution of follow-ups. Speaker labels and timestamps must exist in the source to be recognized. A Forge proposal is not proof that an external task, email, or calendar change has been executed.

The distinction between local and hosted access also matters. An AI client connected through Phasoric's hosted MCP interface can read only explicitly mapped, authorized knowledge. Local-only meeting notes do not automatically become accessible to a remote agent. Do not upload confidential transcripts simply to make them searchable.

6. Turn the reviewed record into persistent AI context

The next assistant should not receive only a prose recap. Give it a narrowly scoped knowledge record that can answer:

  • What was decided? The approved decision, authority, date, and scope
  • Why? The cited source and alternatives or constraints actually discussed
  • What is pending? Commitments with verified owners and dates
  • What is unresolved? Questions and missing evidence
  • What changed later? A successor decision or amendment, not a silent rewrite

Suppose a follow-up meeting reverses the import decision. Keep the first approved decision as historical context, mark it superseded, and link to the later authority and evidence. An assistant asked about today's plan should return the current decision; an assistant auditing the project should still be able to reconstruct the change. See the conflicting-knowledge review guide.

For cross-model workflows, the important artifact is the reviewed, permissioned record—not a copied transcript or the first model's private chat history. This lets another authorized assistant retrieve the decision without pretending it witnessed the meeting.

7. Five acceptance tests before trusting the workflow

Run these tests against a small meeting you are permitted to use:

  1. Proposal versus approval: Include “We could ship Friday” without a decision. The resulting record must not say shipping was approved.
  2. Unknown owner: Mention an action with no assignee. The task must retain an unconfirmed owner rather than guess.
  3. Broken provenance: Move or rename the transcript. The review should reveal a missing reference instead of silently citing a different file.
  4. Contradictory dates: Include two proposed deadlines. The record must flag ambiguity until someone resolves it.
  5. Write boundary: Deny a proposed note write or select a different vault. No unapproved mutation should occur.

Also inspect one saved result manually. Automated checks do not substitute for validating the actual source, permissions, and human interpretation.

Make the next meeting more useful than the last

The quality of AI meeting memory is not measured by how polished the summary sounds. It is measured by whether someone can later determine what was approved, who accepted work, which questions remain open, and where the supporting evidence lives.

Start with one meeting, confirm the source and consent, review five kinds of candidates, and preserve the result as linked Markdown. Durable context emerges from a sequence of accountable decisions—not from asking an assistant to remember everything.

Frequently asked questions

Can AI meeting notes automatically assign action items? They can propose candidates, but explicit assignments should be checked against what participants actually accepted. Treat inferred owners and dates as unconfirmed.

Does Phasoric automatically know who spoke in a recording? No universal speaker-diarization guarantee is established. Its Meeting Memory extraction uses speaker labels already present in a transcript; precise timestamps depend on the recorded source.

Can an MCP-connected assistant read my local meeting notes? Not automatically. Hosted MCP requires a supported, explicitly mapped and authorized source. Local-only browser knowledge is separate from hosted agent retrieval.

Frequently asked questions

Can AI meeting notes automatically assign action items?

They can propose candidates, but owners and deadlines must be verified against explicit commitments. Keep inferred assignments unconfirmed until someone accepts them.

Does Phasoric automatically identify who spoke in a recording?

No universal speaker-diarization guarantee is established. Meeting Memory recognizes speaker labels and timestamps already present in the source; it cannot invent exact speaker identity.

Can an MCP-connected assistant read my local meeting notes?

Not automatically. Hosted MCP needs an explicitly mapped and authorized source. Local-only browser data stays separate unless a supported connection and permission are configured.