# The local archive (/en/docs/zm/archive) Keep meeting text searchable and back up recording files before Zoom's retention removes them. Choose whether to retrieve hosted meeting information, import a transcript you already have or archive cloud audio and video. Text stored in the meeting database and media files in an archive folder are separate copies; a recording backup is not automatically searchable in the database. ## Choose what to keep [#choose-what-to-keep] | Source you have | Use | Check afterwards | | --- | --- | --- | | Hosted meetings available to your app | `zm pull --lookback-days 7` | List meetings, open one transcript and inspect missing parts | | Downloaded VTT transcript files | `zm import ./downloaded-transcripts` | Compare the imported date/title with the original | | Cloud recording assets | Recording list, download or sync | Verify the destination manifest and checksums | | Archived VTT files to add to search | Archive import | Check its receipt, then search the stored text | For the API paths, connect the app with the [permissions for that task](/llms.mdx/docs/zm/zoom-app/content.md#add-archive-or-management-access). File import and local archive inspection do not need browser login. Start with one meeting before setting up repeated retrieval or a daily backup. ## Where meetings are stored [#where-meetings-are-stored] Login, cloud recording downloads and remote meeting commands work independently of the meeting database. Import, pull, folder watching and meeting reads use the published shared SQLite adapter. Store connections open lazily and close before successful results are emitted. Meeting data persists in the shared WireCat database; `MESSAGING_STORE` selects an explicit database path. The shared model, exports and ingestion loop come from `@wirecat/cli-messaging/meetings`. The shared store uses `wirecat.db`. Opening its default path deletes the legacy unencrypted `messages.db` and its `-wal` and `-shm` sidecars, without migrating them. The legacy files are kept when `MESSAGING_STORE` is set, an explicit store path is supplied, or another process holds the legacy database open. A later default-path open retries deletion after that process closes. Select a [local profile](/llms.mdx/docs/zm/sessions/content.md#select-a-local-profile) before importing or pulling. ## Pull hosted meetings [#pull-hosted-meetings] With the API source selected, pull a recent period and inspect the result: ```sh zm config set source api zm pull --lookback-days 7 zm meetings list --json zm meetings transcript --json ``` Use an ID from the list. `zm pull` retrieves hosted occurrences with available participants, transcripts and AI summaries. It does not create missing transcripts. Check Zoom's [generation and retention settings](/llms.mdx/docs/zm/zoom-app/content.md#before-your-first-zm-pull) if the text is absent. Use `pull --lookback-days 7` to revisit the last seven days, or an older explicit `pull --since 2026-10-01` for earlier occurrences whose transcript appeared later. The cursor alone does not guarantee that late transcripts are revisited. Serialize pulls for the same account to avoid concurrent event creation. A failed occurrence leaves the cursor unchanged so the next pull can retry it. ## Downloaded transcripts [#downloaded-transcripts] For a meeting you only attended, obtain a VTT file from the host or download it if your Zoom permissions allow. Attendance does not guarantee transcript access. Put the file directly in your chosen folder and rename it with the meeting date/time/title, for example `2026-10-01_1000_example-planning.vtt`. Renaming does not change its transcript contents. ```sh zm config set account.profile personal zm config set files.timezone Europe/Madrid zm import ./downloaded-transcripts zm meetings list --json ``` The files source reads ordinary `.vtt` files directly inside a folder, without recursion. Names use `YYYY-MM-DD_HHMM_topic-slug.vtt`. Occurrence IDs derive from timestamp and normalized topic; these files carry no Zoom occurrence UUID. Participants remain unknown: speaker names are preserved as text, without guessing a person's identity. Check the imported date, title and transcript before attaching project or person records. File times default to UTC. Set `files.timezone` to the IANA zone used when the filenames were created, for example `Europe/Madrid`. Ambiguous or nonexistent daylight-saving times are refused. Conflicting contents for the same normalized filename key are an error; correct the conflicting files before retrying. File discovery fails explicitly on invalid names rather than silently importing a partial folder. ## Watch for new meetings [#watch-for-new-meetings] For repeated API ingestion, use `zm watch --source api --lookback-days 7 --jsonl`. A folder argument selects file watching; otherwise the configured source applies. File watching can use `files.folder`. API watching accepts `--since` and revisits zero through 31 days, and requires online access. `--since` sets the initial lower bound; after a successful cursor write, later cycles use the cursor with the selected overlap. Failed cycles retain the initial bound until a cursor write succeeds. The overlap also subtracts from the initial `--since` bound. `--interval` controls the delay between completed cycles. Profile, source and file settings are captured, along with the OAuth client, before watching begins; changing configuration requires restarting the watch. Cycles never overlap and close their store connection before emitting a receipt. Stop with Ctrl-C. To watch a transcript folder: ```sh zm watch ./downloaded-transcripts --interval 10s --jsonl ``` Watching rescans the folder for each cycle, so corrected older files are revisited. It imports serially, closes the cycle's store connection before emitting its receipt and waits for output to drain before starting another cycle. Intervals range from one second to one day. This is a foreground process; stop it with Ctrl-C. Headless watching requires `--jsonl`; a stream cannot use `--json`. Select a local profile before watching, just as for a one-time import. Cancellation allows up to one second to flush a fully completed receipt if it arrives during store close. A blocked output pipe stops that flush; no later cycle starts. ## Archive cloud recordings [#archive-cloud-recordings] `zm` archives available cloud recording video, audio, transcripts and chat files. Find an occurrence and download its available files to an explicit destination: ```sh zm recordings list --from 2026-10-01 --to 2026-10-10 --json zm recordings show --json zm recordings download --destination ./zoom-archive --json zm recordings archive verify ./zoom-archive --offline --json ``` Replace `` with an occurrence UUID returned by the recording list, not a local meeting ID or the numeric Zoom meeting number. The download receipt identifies saved and unavailable assets. Verification checks the local files against their manifest; it does not prove Zoom had every expected recording. If there are no recordings, check the date range, host account, cloud recording availability and scopes before widening the backup. For recording history across months, use: ```sh zm recordings sync --from 2026-01-01 --to 2026-10-10 --destination ./zoom-archive --jsonl ``` Sync accepts explicit ordered dates spanning at most ten calendar years. It splits them into monthly requests, deduplicates occurrence UUIDs and downloads serially. Each JSONL receipt describes one completed occurrence; earlier receipts and completed files remain if a later request fails. Repeat the same range or an overlapping recent range to discover new assets. With a selected local profile, sync saves durable progress and retains pending recordings. Use `--recheck-from YYYY-MM-DD` to rediscover recent dates for late or changed assets, and `--checkpoint` to select its checkpoint file. `--max-pages` and `--max-entries` bound discovery. Headless sync requires `--jsonl`. Checkpoints acknowledge earlier successful archives; they do not check current disk integrity or prove that all provider recordings were available. Run `recordings archive verify` to check local files. Use `--no-resume` to rescan the full requested range without durable progress. Recording lists accept an ordered date range of at most 31 days. Downloads stream to disk rather than loading a video into memory. The default per-file limit is 20 GiB; change it with `--max-bytes`. Use `--file-ids` to select comma-separated asset ids. Only completed, downloadable assets are saved; the receipt reports unavailable assets separately. Zoom permissions, recording retention and account features determine what can be downloaded. Assets from a meeting you attended are not necessarily available through your cloud recording API. Each occurrence has a manifest containing metadata, file sizes and SHA-256 checksums. Signed download URLs and OAuth tokens are excluded. Repeating a download verifies existing manifest files and skips matching assets. Existing files are never overwritten; a mismatch requires a new destination. With a selected local profile, interrupted downloads retain verified partial bytes and resume only when a strong validator and the returned range agree. Changed representations restart safely. Explicit `--resume` requires a profile; `--no-resume` and downloads without a profile use ephemeral partials. Completed manifest entries remain available if a later transfer fails. ## Locks and recovery [#locks-and-recovery] Archive commands inspect and verify local data without logging in or opening a database. Concurrent writers to the same destination are refused. After a process crash, inspect `recordings archive locks` and preview `recordings archive recover`. Add `--yes` to remove recognized locks owned by a dead process on this machine; active, foreign and unrecognized locks are preserved. Recovery removes stale lock markers only. Download continuation can recover a completed promotion only when its owned journal, verified partial and final file all match; unrelated untracked files remain protected. ## Daily backups [#daily-backups] To plan a daily backup, choose an absolute destination and timer timezone: ```sh zm recordings schedule create --name daily --destination /home/AliceExample/zoom-archive --from 2026-01-01 --at 03:15 --timezone Europe/Madrid --json zm recordings schedule list --offline --json zm recordings schedule run --jsonl zm recordings schedule disable --yes --json ``` Creation previews by default. Add `--yes` to install a Linux systemd user timer; `--dry-run` overrides it. Automatic installation currently supports Linux only. The timer follows its IANA timezone, while recording discovery uses UTC dates and defaults to seven days of recent rechecks. Systemd user services must be available for the job to run; previews also work on other platforms. Jobs capture the local profile, public client ID, destination and runtime/config paths, without credentials. Running a saved job requires its captured profile and client to remain selected. Saved status describes the last acknowledged operation; list/show do not query live timer state. Disabling stops the timer and preserves job definitions and archived files. ## Read and import archived transcripts [#read-and-import-archived-transcripts] Cloud archive assets use hashed filenames and retain their identity in manifests. Read or search archived transcripts directly: ```sh zm recordings archive transcript ./zoom-archive --offline --json zm recordings archive search ./zoom-archive "example plan" --limit 100 --offline --json ``` Both commands verify each WebVTT file they read against manifest sizes and SHA-256 digests before returning parsed cues. Search stops after finding one match beyond its limit; later files remain unread. They need no login, configuration or database. Use `--uuid` or comma-separated `--file-ids` to select assets, and `--max-bytes` to change the default 8 MiB per-file read limit. Search matches literal cue text after Unicode normalization and lowercasing; speaker names are not searched. Its default limit is 100, with a maximum of 1000. `hasMore` means another matching cue exists; increase the limit or narrow the occurrence, file or query selection to read more. Results have no continuation cursor. Finite output remains bounded by `--max-output-bytes`, and a failed verification emits no successful result. Use a separate command to import these assets into the selected local profile: ```sh zm recordings archive import ./zoom-archive --offline --jsonl ``` Import verifies every selected VTT asset before opening the store. The default limits are 8 MiB per file and 64 MiB in total; use `--max-bytes` and `--max-total-bytes` to change them. Select an occurrence with `--uuid`, or assets with comma-separated `--file-ids`. Each occurrence is appended atomically by its raw Zoom UUID, preserving other transcripts, summaries, participants and the pull cursor. Corrected content retains earlier revisions; identical asset hashes are idempotent. An occurrence with no known start time can receive transcripts only if it already exists. Receipts report acknowledged or unavailable occurrences. When the exact receipt capability is available, they also report meeting creation and inserted, replayed and superseded transcript counts from the same transaction. Older adapters keep the acknowledgement without guessed counts. Earlier acknowledged occurrences can remain committed if a later append fails. ## Transcribe local audio [#transcribe-local-audio] Create an independent local transcript from an audio file for an existing stored meeting: ```sh zm meetings transcribe 1 ./AliceExample.wav --model parakeet-v3 --json zm meetings transcribe 1 ./AliceExample.wav --model parakeet-v3 --yes --json ``` Preview verifies the input and describes the chosen model without loading it. `--yes` recognizes with an already installed local model and atomically appends the completed transcript; `--dry-run` overrides it. No model is automatically installed. Input currently accepts PCM16 WAV, mono or stereo, at 16 or 48 kHz; video files are not decoded. `--models-dir` selects an absolute installed model directory. Media, duration, segment count and transcript text have explicit bounds through `--max-bytes`, `--max-duration`, `--max-segments` and `--max-text-bytes`. Recognition uses fixed windows of up to 30 seconds, so timestamps describe those windows. Speakers remain unknown. Provider transcripts and historical revisions are preserved; the applied receipt gives transaction counts and provenance. Recognition needs an installed supported speech model; this guide does not assume one is present. Convert unsupported audio/video to the stated WAV format with a tool you already use before transcribing. After an applied receipt, open `meetings transcript` and check the words and timestamp windows before treating the new text as evidence. Continue with [search and citations](/llms.mdx/docs/zm/search/content.md).