Skip to content
github-actions[bot] edited this page Nov 16, 2026 · 1 revision

Search and evidence

Find what was said in your stored meetings, read the exact passage with a reference back to it, see what you discussed with a person, link meetings to events and people, and preview a follow-up task tied to its source.

Start with a question

Find discussions about Research Example in my stored meetings. Show the dates and the passages
about the launch decision. Include references and tell me which transcripts or periods are missing.

The agent should identify the meeting, open the matching passage and distinguish a recorded participant from a person merely named in the text. Check the surrounding cues before accepting a decision or promise. An empty search result does not prove that a discussion never happened.

Search the stored meetings:

zm personal search meetings "example plan" --json
zm personal search meetings "example plan" --limit 20 --cursor <next_cursor> --json

This searches the stored profile, not Zoom's live history. Pull or import missing meetings first. search meetings indexes whatever was stored since the last search before it reads, so there is no separate index step. All query words must match, through literal prefixes or stored word forms. The default configuration adds English, Russian and Spanish word stems; a store can disable or customize it. A word of four letters or more also matches spellings speech recognition garbled («риелтор» finds «реелтора», «рилторов»); exact hits come first, and a widened hit has match: "corrected" and the stored word in corrections. The query is the messages query language:

zm personal search meetings 'риелтор OR агент' --json             # either word
zm personal search meetings '(ипотека OR кредит) -квартира' --json # groups and exclusions
zm personal search meetings '"циан риелтор"~3' --json              # the words at most 3 lines apart
zm personal search meetings 'exact:риелтор' --json                 # the typed form only, no widening
zm personal search meetings 'from:"Alice Example" бюджет' --json   # lines Alice said

from: is who said the line; who attended is --identity-id or --person-id. --meeting-id, --event-id, --series-id, --since-time and --until-time narrow the selected account. The opaque cursor belongs to that account, query and filters; it can continue inside a large transcript. JSON provides page.next_cursor, next and any missing index information. JSONL emits hit records followed by an end record with the page and completeness. Hits include the cited ref, meeting reference meeting, excerpt and offset_ms; expand a cue with meetings passage.

Search by meaning or in an archive

--model <model> ranks words and meaning together, with the same filters, building missing vectors first; a question («что решили по цене?») is asked by its content words and by meaning. See topic search.

search recordings <directory> <query> searches the transcripts of a downloaded archive folder before any import, with no store, configuration or login:

zm personal search recordings ./zoom-archive "example plan" --limit 100 --offline --json

Its hits are the ones search all answers with: ref (recording:<file id>/<cue position>), title, timestamp, excerpt, sender_name and offset_ms; account is the archive folder.

Why zm does not search on Zoom's side

Zoom's API has no way to search transcript text. The Meetings API reference only hands transcripts out as files: GET /meetings/{meetingId}/transcript, and recording files with file_type TRANSCRIPT (VTT) from GET /meetings/{meetingId}/recordings and GET /archive_files. None of its endpoints takes a text query (checked 2026-10-21). Searching text inside transcripts is a feature of the Zoom web portal only: "You can search for text within audio transcripts", and an admin can search every transcript in the account (Using audio transcription for cloud recordings).

So zm personal downloads each transcript once and searches its own copy: that works offline, keeps meetings Zoom has deleted, and lets a word match its other forms and garbled spellings.

Narrow by people and work

Combine a participant identity or explicitly linked person with a project and date range:

zm personal meetings people "Alice Example" --json
zm personal search meetings "launch" --identity-id 12 --project RES --since-time 2026-09-12 --until-time 2026-10-12 --json
zm personal search meetings "budget" --person-id 7 --topic-name costs --since-time 2026-09-12 --json
zm personal search meetings "launch" --tag planning --limit 20 --json

The IDs and labels above are fictional: replace them with records in your local store. Use the identity_id from participant results for --identity-id; --person-id takes the numeric ID of an explicitly linked person. Several participants with the same display name may be different identities. Inspect them before linking; these filters select meetings and do not identify who spoke an anonymous transcript cue.

--project accepts an existing canonical project name, key or explicit alias. Only confirmed meeting membership counts; an agent's proposed membership or a matching transcript word does not establish membership. --tag and --topic-name accept canonical names or explicit aliases of labels already attached to a meeting; choose one label filter. Missing or ambiguous vocabulary names fail instead of silently broadening the search. A date alone covers the whole local day of the computer's zone, and a time without an offset is read in that zone; give an offset or Z for another zone. Existing continuation and coverage rules still apply.

If a meeting has not been linked to a project or labelled yet, start with word search and inspect the evidence before organising it with Memo's shared records. Over MCP, run search meetings through zm_write: it updates the word index before searching. It uses the same filters and time syntax as the CLI, in the server-selected profile.

meetings people <query> finds saved participant observations by name or email within the selected account. It uses literal Unicode matching; it does not translate names or infer that two people are the same. Use --limit (default 50), --scan-limit (default 1000) and --max-read-bytes (default 1048576) to bound matching observations, inspected observations and stored name/email bytes per call. The JSON answer keeps item.page.has_more, item.page.next_cursor, coverage and page-local identity ambiguity. JSONL keeps that report in the item record’s data, followed by an end record. Follow the report's item.next step or repeat the same query with --cursor. An empty page can still have continuation. Distinct identity IDs are candidates to review, never automatic person links. Archive completeness remains unknown.

meetings list filters metadata with a query, for example zm personal meetings list 'started_at:[2026-09-12 TO 2026-10-12]' --json. JSON returns items, page and optional next steps. JSONL emits each meeting in an item record, then an end record with the page and completeness. Use --page to continue. The date query replaces the former list-only --since-time and --until-time flags; those flags remain available on transcript search and ingestion commands. search all uses the shared search service and searches only meetings by default, those of the profile's own Zoom account. It reads them through the meeting search index and ranks them with the other sources. It answers a list of hits, as tg's search all does: a meeting hit has scope (transcript, chat or summary), the stored line id, and a transcript hit's cue_ref, which meetings passage opens. One meeting gives at most 3 hits. Continue with --page. Search all never builds the index; when missing names the account, zm <profile> search meetings builds it. meetings evidence and meetings context name each line by ref, its start by offset_ms, and ids as strings, as meetings prepare does:

zm personal search all "example plan" --limit 100 --json
zm personal meetings evidence meeting:3/15/20/0 --cues 100 --bytes 65536 --json
zm personal meetings context person:7 --meetings 10 --scan-limit 1000 --json
zm personal meetings task-proposal meeting:3/15/20/0 --type promise --json

Broader search requires explicit --only messages,mail,notes,meetings. Message reads use the selected local account; mail reads include stored email accounts; notes are global to the shared store. Notes are found by any part of a word in any letter case, titles included; a note search of one or two characters is refused with "search needs at least three characters". Meeting matching uses word prefixes; unsupported meeting syntax or exact queries are reported as skipped. Continue the hit list with --page. JSONL emits one item per hit and an end record. Skipped sources set meta.complete to false and appear in issues; search still exits zero.

Put filters in the query, and save a search

The project, label and date filters can also be written in the query, joined to the words: date:, project:, tag: and topic: mean --since-time/--until-time, --project, --tag and --topic-name.

zm personal search meetings "launch project:RES date>=30d" --json
zm personal searches create weekly-launch "launch date>=7d" --project RES
zm personal searches run weekly-launch --json

They choose meetings, not lines, so they cannot sit inside OR, NOT or brackets, and one filter cannot be given both in the query and as an option. A query of filters alone, with no words, is an error.

searches create checks the query and options as search meetings would, then saves them under a name. They are kept as typed and read again at each run, so date>=7d in the query always means the 7 days before that run; --since-time takes a fixed date only. searches run answers exactly as search meetings does; searches list, show and delete manage them. Store ids (--meeting-id, --person-id and the like) and --tag are not saved: a saved search names dates, a project or a topic. Saved searches share the store with tg's and max's but stay apart: a messenger never lists or runs a meeting search, and zm never runs a messenger's.

Read evidence

Meeting references have the form meeting:<account>/<meeting>[/<transcript>[/<cue-position>]]. They are restricted to the selected local account. Transcript references retain historical revision identity. Evidence bounds both returned content and database reads; resume with its returned cue reference using --cursor, or its revision cursor using --after-transcript-id. These continuation options are mutually exclusive. --max-read-bytes defaults to 4 MiB per page and --max-transcript-pages bounds revision scanning, including empty revisions. Follow the returned continuation and completeness fields rather than assuming a bounded response is exhaustive.

Read around a search hit

Use a transcript hit's ref to check the discussion immediately before and after the matching line:

zm personal meetings passage meeting:3/15/20/4 --before-n 3 --after-n 4 --format markdown
zm personal meetings passage meeting:3/15/20/4 --before-n 0 --after-n 0 --json

Replace the fictional reference with one from your search result. The passage shows timestamps, speaker labels and a reference for each line. It reads your stored transcript without contacting Zoom. Summary and chat hits carry a meeting reference; use meetings show or meetings evidence for those instead, because passage needs a transcript line reference.

--before-n and --after-n count transcript lines, including when their stored positions have gaps. Both default to 3; the combined maximum is 999. --bytes bounds the included cue JSON (64 KiB by default), while --max-read-bytes bounds selected stored text (4 MiB by default). A line too large for the anchor budget fails rather than returning a misleading empty passage. The global --max-output-bytes also bounds the final rendered result.

Inside item, check coverage.has_before, has_after and truncated_by before treating a passage as complete. To extend it, open before_ref with more preceding lines, or after_ref with more following lines. These anchors are inclusive: the boundary line repeats, so an agent can deduplicate by ref. archive_completeness: "unknown" describes the archive, even when this local window is complete.

A corrected transcript keeps the old reference readable and marks it superseded; it does not silently switch to the replacement. To reject an altered anchor, pass its previously returned target_fingerprint as --expected-fingerprint. JSON output retains the structured cues and fingerprints; Markdown quotes transcript content as text. Use --json=items.ref,items.start_ms,items.speaker_name,items.text for compact structured cues. Coverage and continuation remain available; field selection does not reduce the stored read.

Over MCP (zm personal mcp serve) an agent runs the same meetings passage through zm_read, with the same arguments; the profile is fixed by the server. Ask the agent to quote and cite the surrounding passage before suggesting a fact or follow-up task.

Read several queries or passages together

When checking several names or cited lines, your agent can use one batch to keep each answer and its continuation together. It uses the selected profile and stored meetings, with one bounded index pass before word searches. It does not contact Zoom or run a model.

Save a JSON file such as requests.json, replacing the fictional reference with a retained cue from your search results:

{
  "requests": [
    {"id": "name", "command": ["search", "meetings"], "args": ["Research Example", "--limit", "10"]},
    {"id": "alias", "command": ["search", "meetings"], "args": ["RES", "--limit", "10"]},
    {"id": "passage", "command": ["meetings", "passage"], "args": ["meeting:3/15/20/4", "--before-n", "2"]}
  ]
}
zm personal batch requests.json --json
zm personal batch requests.json --format markdown
cat requests.json | zm personal batch - --jsonl

Each result retains its id, command, status and complete. Successful results carry the normal answer in answer; failures carry error. Duplicate queries and overlapping passages stay separate. Pass a search result's cursor or a passage's expected fingerprint in that request's args when continuing. Only word search meetings and meetings passage are accepted; model, profile and write options cannot be nested inside the batch.

A batch accepts at most 32 requests and 256 KiB of UTF-8 input. Its default stored-text allowance is 64 MiB: it reserves each request's full --max-read-bytes ceiling, plus 4 MiB once for indexing, without refunding unused bytes. These are conservative ceilings, not measured database I/O. Use smaller per-request ceilings or separate batches when needed. The batch-level --max-read-bytes can lower the aggregate allowance. Output is limited to 4 MiB and the global --max-output-bytes can lower it; a limit too small for result framing fails before any stored read.

Check every status and the outer meta.complete: a budget refusal is not_run, while an executed request whose answer exceeded the output allowance is error; later requests are not_run. Partial results exit zero and retain all request IDs. Malformed input exits 2 before opening the store. JSONL emits one item per request and an end record carrying completeness and issues. Cancellation and final output failures fail the whole invocation without printing partial stdout.

For compact JSON, --json=ref,excerpt,source.title,source.started_at,source.url selects fields inside each answer while preserving request IDs, errors and coverage. Passage answers use cue fields such as items.ref,items.start_ms,items.speaker_name,items.text. Text/Markdown includes passage timestamps, speakers and references, with coverage once per result. Inspect missing data and continuation before keeping a conclusion.

Person context

Person context uses stored identity links and account presence, without matching speaker names. A person linked to a meeting is not necessarily a speaker. Pending derived indexes or bounded scans can leave context incomplete. Link a participant to a person explicitly with meetings link.

Link meetings to events and people

Stored meeting references use meeting:<accountId>/<meetingId>. A retained transcript appends its revision id, and a cue appends its position. Event and person linking accept root references:

zm personal meetings link meeting:1/1 --event-id 1 --expected-event-id none --json
zm personal meetings link meeting:1/1 --event-id 1 --expected-event-id none --yes --json
zm personal meetings link meeting:1/1 --participant-id 1 --person 1 --expected-person none --yes --json

These local changes preview by default; --yes applies once and --dry-run overrides it. Expected link options reject a stale decision. Use --event-id none to unlink an event. Person linking moves only the selected participant's identity; it never guesses a person from a speaker name. Target people must have authorized account observations. Explicit --source <accounts> adds account ids for that lookup. Profile labels group local data and do not authenticate a remote identity.

Find a project before linking it

A confirmed-project filter can miss a relevant meeting that has never been linked. Use an existing project from the shared catalog to search its canonical name, key and explicit aliases:

zm personal meetings people "Alice Example" --limit 50 --json
zm personal meetings projects evidence RES 'identity_id:12 started_at:[2026-09-14 TO 2026-10-14]' --limit 20 --json

The project key and identity ID are fictional. Review participant pages before choosing the intended identity; matching names alone do not establish a person link. If the project is missing or ambiguous, inspect the shared project catalog in Memo first. This command resolves existing names only; it does not invent aliases or create projects, labels, person links or membership.

The answer is one item, keeping its canonical project, matched term, cited hits, selection, limits, page and explicit coverage, plus run metadata. Each call reads one name page from the existing local word index; it does not refresh that index. Run a normal word search after ingestion when new passages need indexing. Word reads default to 4 MiB and at most 100 hits; --max-read-bytes and --limit change those budgets. Catalog resolution has a separate 1 MiB budget and at most 32 names. Excess vocabulary fails visibly. Archive/index completeness stays unknown.

Continue with the top-level next step even if item.hits is empty. The cursor is at item.page.next_cursor. Its arguments repeat the project, selection, cursor and budgets. Within-name pages finish before the next alias begins. Relative dates are frozen to absolute bounds in the returned query; use that query with --cursor so a later day does not shift your selection. Changed vocabulary or selection can invalidate the cursor.

The optional selection query accepts exact meeting_id, identity_id, person_id, tag_id, event_id and series_id values, each once and joined by AND, plus started_at dates/ranges. Date ranges include whole UTC calendar days; quoted ISO timestamps give exact bounds. OR, negation and confirmed project-membership filters are refused. Over MCP, zm_read runs it as { "command": "meetings projects evidence", "arguments": { … } } with the same arguments; the server fixes the profile.

For every relevant retained cue, open zm personal meetings passage <ref> --format markdown and check the surrounding words. association: "unconfirmed-word-evidence" means candidate evidence: a key may be incidental, and repeated refs across names are the same evidence. Propose associations only after reviewing the passage. If no project record exists yet, ordinary word search remains available without --project; repeat known names manually and retain the same selection.

Suggest projects for one meeting

When you have found a meeting but do not know which saved project it concerns, ask your agent to compare that meeting with your existing project names, keys and aliases:

Look at this meeting and suggest which of my saved projects it may concern. Show the passages
that support each suggestion, distinguish a title hint from transcript evidence, and leave the
project links for review.

The scan reads the selected profile's local meeting and the shared project-name catalog. It does not contact Zoom or require an embedding model. Use the search result’s meeting field as the parent meeting reference. A transcript hit’s own ref points to a cue and is not accepted by this scan. For example, with a fictional parent reference:

zm personal meetings projects scan meeting:1/2 --cues 20 --json

The answer keeps each canonical project reference, matched name/key/alias and supporting source. Title matches are labelled metadata-title; transcript matches include a retained cue reference, fingerprint and revision. Open the cited passage before proposing a link. A shared alias stays ambiguous, even if the candidate limit hides another project. The scan saves no association and does not treat an existing or repeated match as additional evidence of membership.

One scan reads one bounded transcript page and at most 128 saved projects with 64 KiB of catalog names. It defaults to twenty cues and ten candidates. --cues takes 1–100, --limit takes 1–10, --max-read-bytes bounds each stored read up to 64 KiB, and --bytes bounds the answer item from 2–64 KiB. The fixed transcript evidence packet is 32 KiB. If the project catalog exceeds its bound, select one known project with project evidence instead; the scan refuses rather than silently ignoring some projects.

Inspect both coverage and evidence.coverage. To read more transcript evidence, pass evidence.coverage.next_cursor with the same meeting reference using --cursor. A title or project may repeat on the next page. Fingerprints describe live reads, not a frozen snapshot. Candidate output limits are separate: increase --limit or --bytes when candidates were omitted. A byte-truncated page with no cues and no cursor means a single cue did not fit; it does not establish an absence of relevant discussion. Archive completeness remains unknown.

After checking the evidence, memo projects meetings link can apply the confirmed local association using the exact stored meeting ID and account, such as --account zoom:zm-profile:AliceExample. This is a write, separate from the scan. When confirming an existing stored proposal, use that proposal's fingerprint from Memo; the scan does not create one, and its vocabulary fingerprint is not a membership receipt. The Memo project command reference explains the inputs. A suggested project remains unconfirmed until that reviewed write is applied.

Notes, tags and memories with memo

Use memo for the existing shared notes, tags and memories interface. Both tools use the same store; point MESSAGING_STORE at the same file if it is overridden. For a Zoom profile named AliceExample, its provider account key is zm-profile:AliceExample:

memo tags add planning --meeting 1 --account zoom:zm-profile:AliceExample --json
memo tags remove planning --meeting 1 --account zoom:zm-profile:AliceExample --json
memo notes add --text "Alice Example follow-up" --meeting 1 --account zoom:zm-profile:AliceExample --json
memo notes list 'about:"meeting:1/1"' --json
memo facts add "Alice Example will review" --kind promise --scope work --subject meeting:1/1 --evidence meeting:1/1/1/0 --json
memo facts list 'subject:"meeting:1/1"' --json

Memo's add/remove commands write directly to the local store. Owner-authored facts are accepted; agent-authored facts submitted through Memo MCP stay proposed until reviewed. Memories keep summaries, digests and preferences separately. Meeting revision and cue evidence remains addressable after corrections. The references and ids above are invented examples; use ids from your stored meetings and selected account.

Suggest labels without creating synonyms

An agent can read the meeting's metadata and cited passages, then compare possible labels with Memo's shared vocabulary. Ask for a proposal first:

Suggest a small set of tags for this meeting. Reuse our existing names and aliases, show the
passages behind each suggestion, and flag likely synonyms for review. Do not create or merge
labels while preparing the proposal.

The existing Memo tools provide the checks; Zoom keeps no second label catalog:

memo vocabulary list 'kind:tag' --max-entries 128 --max-read-bytes 65536 --json
memo vocabulary check planning --kind tag --max-entries 128 --max-read-bytes 65536 --json
memo vocabulary check --kind tag --max-entries 128 --max-read-bytes 65536 --max-comparisons 1000 --limit 20 --json

The first command lists canonical names and aliases. The second checks a proposed name for an exact existing label or similar names before creation; it checks across label kinds so a topic collision is not hidden. The last command audits likely synonyms among existing tags. These reads create and merge nothing. Their catalog/comparison limits are real work bounds; a refusal or incomplete audit is not proof that the label is unique.

Reuse the canonical name when preflight says reuse. For review, inspect the alternatives and their meeting evidence before choosing a label or an owner-reviewed alias. create means no existing lexical candidate was found, not that the concept is new. Similarity preserves accents and does not discover every translation or synonym. Keep those uncertain proposals for review. Memo guards label creation against duplicate-like names and stale creation fingerprints; a fingerprint is a preview precondition, not permission to accept a proposal automatically.

After review, apply the selected existing tag with Memo's meeting/account arguments above and search meetings by that tag. Use Memo's tag guide for the shared labeling workflow. Zoom's preparation and project scan save no labels or merges.

Task proposals

Task proposals support question, request, mention and promise; applied: false and persistence: "proposal-only" mean no task was created. zm does not find commitments on its own: an agent reads transcripts and evidence, then proposes a task for the passage it chose.

Before saving notes, tags or facts, ask your agent to show the existing project aliases and labels, then propose the links it intends to add. Review possible tag synonyms rather than creating another label for the same subject. Memo is installed separately and must use the same database path.

Searching by meaning instead of by words is covered in topic search. For an upcoming call, select the provider series from the Zoom schedule, then prepare its context. Add --project RES for one existing project to include cited name/alias candidates separately from confirmed links; inspect its continuation and coverage before treating the result as complete.

Clone this wiki locally