openclaw transcripts

Inspector and export command for durable meeting transcripts. Google Meet, Microsoft Teams, and Zoom browser participants capture notes automatically; the transcripts agent tool also supports provider capture and manual import.

Canonical transcript state lives in the shared SQLite database at $OPENCLAW_STATE_DIR/state/openclaw.sqlite. show and path explicitly materialize user-facing artifacts under the state directory:

$OPENCLAW_STATE_DIR/transcripts/YYYY-MM-DD/<session>/  metadata.json  transcript.jsonl  summary.json  summary.md

These files are exports, not a second runtime store. OpenClaw does not read them back during capture, summarization, or listing. Default state directory is ~/.openclaw; override with OPENCLAW_STATE_DIR. The date directory comes from the session start time; the session directory is a filesystem-safe slug derived from the session id.

Read transcripts in the Control UI

In the Control UI, open the sidebar's pencil menu (Edit pinned items) and choose Meetings to browse the same SQLite archive at /meetings. You can pin Meetings to the sidebar; it is not pinned by default. Meeting notes are separate from agent chat history in Sessions.

Search titles and session/source IDs; meeting URLs are not searched. Filter by exact provider, account, or agent ID, or by the session start date. Date filters use UTC: Started on or after includes the selected day, and Started before excludes it. Results load in deterministic pages. Changing a filter or selecting Refresh starts pagination again.

Select a meeting to open its stored Summary. Existing /meetings?selector=... bookmarks keep working. Select Transcript for timestamped speaker text. Search within this transcript searches stored utterances on the Gateway, including text not yet loaded in the browser. Load more continues reading; the browser keeps the latest five loaded pages. The URL preserves the selected meeting and tab. Opening a meeting does not generate a missing summary.

Download Markdown includes the transcript and any stored summary. Download JSONL exports the reader's public utterance projection, excluding provider-private metadata and local filesystem paths. Local CLI exports retain their existing raw format. Browser exports larger than 4 MiB fail visibly without a partial file; use openclaw transcripts path <session> --transcript or openclaw transcripts path <session> --dir on the Gateway host for larger exports.

Archive reads require operator.read or its write/admin implication and permission to read the shared archive. On restricted multi-user profiles, choosing an agent filter does not grant archive access. Capture configuration requires operator.admin.

Commands

openclaw transcripts listopenclaw transcripts show <session>openclaw transcripts show YYYY-MM-DD/<session>openclaw transcripts path <session>openclaw transcripts path YYYY-MM-DD/<session>openclaw transcripts path <session> --diropenclaw transcripts path <session> --metadataopenclaw transcripts path <session> --transcriptopenclaw transcripts list --jsonopenclaw transcripts show <session> --jsonopenclaw transcripts path <session> --json

Use the selector printed by list to address an exact capture. An existing canonical selector takes priority over a raw session ID with the same text. Otherwise, show and path accept YYYY-MM-DD/<raw-session-id>, keeping the entire suffix literal, including punctuation and slashes. For example:

openclaw transcripts show '2026-05-22/notes: room/one'

If neither qualified form finds a capture, the complete input is matched as a literal raw session ID or export slug, case-sensitively. A date-like prefix in a raw ID does not prevent this lookup. Multiple matches require a dated selector; no raw ID is sanitized to choose a capture. Default session IDs include a timestamp and random suffix; give a session a fixed ID only when that ID is unique within the day.

If the filesystem-safe export name exceeds 255 bytes, OpenClaw shortens it to a prefix plus a deterministic SHA-256 hash of the complete original session ID. Only the derived export name and its selector change; the raw session ID, provider stop handle, and stored notes stay intact. Names that already fit remain unchanged. Use the selector printed by list for the shortened name. For existing sessions with oversized stored names, run openclaw doctor --fix to repair their derived selectors without changing stored notes.

Output

list prints one tab-separated line per session: selector, start time, title, summary path.

2026-05-22/standup  2026-05-22T09:00:00.000Z  Weekly standup  /Users/user/.openclaw/transcripts/2026-05-22/standup/summary.md

The selector is the safest value to pass back to show or path.

Tool selectors

Reading notes from any session

Ask an agent to list past meetings and read their notes with the transcripts tool. Reads are not tied to the agent session that captured the meeting. Operator callers can read all meetings on the Gateway. Channel callers can read only meetings allowed by the source provider; Discord voice reads remain within the caller's guild. These read permissions do not change capture or summary write permissions.

{ "action": "list", "limit": 20 }

list returns newest meetings first, with a selector, start time, title or provider name, utterance count, and participants. limit defaults to 20 and accepts integers from 1 to 50. The text is bounded; structured results are in details.sessions.

{ "action": "show", "selector": "2026-05-22/notes-room-one" }

show returns the stored notes Markdown and session details. Its text is capped at 12,000 characters; a truncation marker points to openclaw transcripts show <selector> for the full notes. A capture without a summary reports that notes are not available yet, including whether it is active. Reading notes does not regenerate the summary or export artifacts.

Selecting a capture

The transcripts tool returns both the unchanged raw sessionId and a canonical selector from start, import, stop, and summarize. Authorized status results include selectors for active captures and entries awaiting finalization. Its model-facing text shows up to three complete selectors, prioritizing captures awaiting finalization and reporting any omitted count. Structured status details retain the full authorized list. Bounded active-capture summaries include source locators and titles so the agent can identify the intended meeting. Prefer selector for subsequent show, stop, or summarize calls:

{ "action": "summarize", "selector": "2026-05-22/notes-room-one" }

Show, stop, and summarize require exactly one of selector or sessionId. Other actions reject selector; start and import continue to accept raw IDs through sessionId. Explicit selector input accepts canonical selectors and the historical date/raw-ID form above, but never falls back to the whole input as a raw ID.

Legacy sessionId input considers qualified and raw/slug meanings together. If they identify different captures, the tool reports ambiguity without listing candidate details. This stays ambiguous after a capture ends. Use a selector returned by start, import, or authorized list/status, or inspect openclaw transcripts list locally and pass the desired value in the selector field. Both sides of a raw-ID/selector collision remain addressable by their own canonical selector.

Without a conflicting qualified meaning or a different raw-ID/slug candidate, legacy sessionId selects the current exact raw-ID capture for stop and summarize, even when historical captures reuse that ID. With no current capture, repeated historical IDs require a dated selector. An explicit selector for an older capture does not stop its newer same-ID sibling.

show selects and authorizes the durable capture, using live state only to report whether capture is active. Repeated historical IDs require a dated selector even when one capture is active.

Gateway and Control UI reads

Open Meetings in the Control UI to browse captured meetings and notes without a terminal. The page and other Gateway clients use these read-only RPC methods:

These methods require operator.read or its write/admin implication and expose meetings across one trusted Gateway domain. Restricted operator profiles need permission to read the shared archive; selecting an agent filter does not grant access. Use separate Gateway domains when readers need isolation. Source locators contain only providerId, accountId, guildId, channelId, threadTs, fileId, kind, and sanitized meetingUrl when present, never arbitrary capture metadata. See Gateway protocol.

List search matches titles and session/source IDs, excluding meeting URLs. Date bounds use session start times: startedAfter is inclusive and startedBefore is exclusive. Dates are compared by instant using JavaScript date-string semantics, including stored UTC offsets. Original timestamps and selectors remain unchanged. Unparseable stored dates sort last and are excluded from date ranges. Equal instants sort by session ID, then original timestamp. Chronological page selection scans candidate captures before reading only the selected page's notes and participants, so read time grows with archive size. Cursors belong to their current query and filters; changing either requires a fresh first page. A null nextCursor ends pagination.

The stored summary Markdown is the canonical notes text, matching the CLI's show output. Reads do not generate summaries or materialize files. Utterances are omitted unless includeUtterances is true. Supplying limit, cursor, or query selects paginated reads: at most 100 utterances per page, default 50, without truncating stored text. query searches the full stored transcript, including unloaded pages. Paginated reads, lists, and status results are bounded to 1 MiB, with stored payload bounds enforced before transfer into JavaScript. Oversized rows or invalid cursors fail visibly. Summary participants and model/heuristic provenance are shown when available; older summaries need not contain them.

For compatibility with clients released before archive pagination, transcripts.get without limit, cursor, or query retains the most recent 2,000 utterances in chronological order, with each sanitized text clipped to 4,000 UTF-16 units. These legacy requests return nextCursor: null and retain a 25 MiB public-result ceiling, matching the existing Gateway client transport limit. They preserve the original raw-row reading behavior before text clipping and public projection. Send an explicit limit to adopt bounded page reads and retrieve complete text; continue with nextCursor until it is null. All archive access checks apply to both request shapes.

Exports are bounded to 4 MiB and fail without returning a partial file. Markdown preserves stored notes and appends the complete transcript under a separate heading. JSONL contains the public utterance projection: sequence, utterance ID, full text, speaker identity, source timestamps, and finality when available. It excludes private provider metadata and filesystem paths; the local CLI export retains its raw utterance format. Use openclaw transcripts path <session> --transcript on the Gateway host for larger exports.

Status reports registered subscriptions, not confirmed recording. armed, not-active, and unknown remain distinct; a sanitized URL alone cannot prove which original invitation started a capture. Provider, configured-source, and active lists are limited to 100 entries with omitted counts. Saved utterance counts come from durable rows. The latest transcript is the most recently updated session containing utterances; source speech times are not ingestion timestamps.

JSON output

list --json returns objects with sessionId, selector, date, title, startedAt, stoppedAt, source, path, summaryPath, hasSummary. Stored meeting source URLs contain only the origin and path; query strings, fragments, and embedded credentials are removed before persistence.

show --json returns the stored session metadata, selector, session directory, summary path, and summary Markdown text.

path --json returns the selected path and whether that artifact could be materialized. Metadata and transcript exports always exist for a stored session; a summary path reports exists: false until the session has a summary.

Many sessions per day

Sessions group by date, then by session id. Ten meetings on one day become ten sibling folders:

~/.openclaw/transcripts/2026-05-22/  transcript-2026-05-22T09-00-00-000Z-a1b2c3d4/  transcript-2026-05-22T10-30-00-000Z-b2c3d4e5/  standup/

Use default generated ids for automation. Use a fixed id like standup only when it will not repeat on the same date.

Missing summaries

Meeting notes use the owning agent's utility model first, then its primary model when needed. If no model is available, a request times out, or the model returns invalid output, OpenClaw saves deterministic heuristic notes instead. Model generation enhances the notes; it does not gate saving them. Notes include an overview, participants, decisions, action items, risks, and finally the transcript, so bounded readers see the notes before long transcripts. Participants come from speaker labels in first-appearance order, not model guesses. Summary JSON records source as model or heuristic and, for model notes, the model reference used.

The model receives at most 48,000 transcript characters, preserving the beginning and end when the middle must be omitted. Stored utterances remain intact. Use transcripts summarize (the agent tool's summarize action) to regenerate notes from the stored transcript, including after changing model configuration.

The tool's status action lists active capture subscriptions, not historical notes. When a provider ends or replaces a subscription, OpenClaw records stoppedAt and stores its summary; the transcript remains available to list, show, and the tool's summarize action. A temporary transport disconnect does not end a subscription. Stopping historical notes does not stop a newer capture or change the recorded stop time.

Provider-driven completion stores the summary without exporting files. Explicit tool stop, import, summarize, and configured auto-start shutdown also attempt to materialize summary.md. If terminal persistence fails, status reports the ended capture under pendingFinalization, separately from active captures. Use the tool's stop action for that session to retry persistence without stopping the provider again.

If the provider cannot finish cleanup and has not reported that capture ended, status keeps the capture active with cleanupPending: true. Existing utterances stay intact, and final notes wait for cleanup. Retry stop with the same selector after the provider recovers. Replacing or disabling the plugin does not transfer cleanup to another provider instance.

A session can appear in list without a summary while capture is still active, if a provider failed during stop, or if metadata was stored before any utterances arrived.

Use path <session> --transcript to inspect the raw append-only transcript, or run the transcripts tool's summarize action to regenerate the Markdown summary.

Summaries are saved in SQLite before optional artifact export. If export fails, the saved summary remains available even when summary.md is missing. Configured auto-start captures log warnings during shutdown for failed exports or provider stop errors. Correct the export destination problem, then run openclaw transcripts path <session> or openclaw transcripts show <session> to retry the export; an intended path in a warning is not proof of an exported file.

Historical sessions without complete account-owner metadata remain on a local recovery path. Recover an agent-owned row with a local turn for that agent; a row with no agent attribution requires a local main-agent turn. Sources without account binding retain main-agent access across their normal surfaces. Missing providers, partial owner metadata, and accountless historical sources also stay on this local recovery path.

openclaw agent --agent <owning-agent-or-main> --local --message \  "Use transcripts summarize for session <session>."

Upgrading the legacy file store

OpenClaw releases that predate the SQLite store wrote canonical runtime state directly beneath $OPENCLAW_STATE_DIR/transcripts/. Run:

openclaw doctor --fix

Doctor imports the complete legacy tree into SQLite, verifies row counts and ordering, records migration receipts, and moves the verified source tree to a timestamped transcripts.migrated-* archive. Runtime commands do not fall back to the legacy files. Keep the archive until you have verified the imported sessions and any exports you rely on.

Configuration

Open Settings → Communications → Meeting capture to edit the existing transcripts.enabled and transcripts.autoStart settings. Enable transcript storage controls whether durable capture is permitted; each auto-start source opts in a provider and source. You can add or remove sources and edit their title, account, source locators, and optional custom session ID. Occupancy mode chooses session IDs automatically, so its custom ID field is disabled while preserving the saved value.

The controls use the shared Settings draft, automatic saving, validation, and apply flow. If Apply changes appears, use it to activate saved changes. If a restart interrupts a pending draft, Autosave paused after reconnect keeps that draft without sending it to the new connection. Review it and select Save in the Settings footer. The full transcript schema editor is available under Meeting capture → Advanced settings.

Changing only auto-start source titles applies to future captures without restarting or interrupting current captures. Current and historical notes keep their original title, source, agent attribution, and selector. Other source edits retain normal Gateway restart behavior.

Startup retries preserve the same admitted ID, original title, start time, source, and saved notes only while the exact failed provider attempt retains retry authority. This applies to generated and configured IDs. Retries stop after twelve attempts, service shutdown, manual stop, or a failure that retains cleanup custody. Status reports a bounded diagnostic without provider error details. Recover pending cleanup with the tool's stop action before trying a new capture.

Meeting transcript capture is enabled by default. To opt out globally:

{  "transcripts": {    "enabled": false  }}

• enabled (default true): enable automatic meeting notes, the transcripts tool, and configured auto-start sources. Set it to false when meeting notes should not be persisted on the host. An explicitly requested meeting transcribe mode keeps its existing bounded live-caption tail, but does not write durable rows while this setting is false.

Configure auto-start sources with transcripts.autoStart. Each entry is enabled by being present; omit an entry to disable that source. discord-voice is the bundled auto-start-capable source and requires guildId and channelId. When exactly one configured Discord account has credentials and voice enabled, OpenClaw selects it automatically. When multiple accounts are voice-capable, OpenClaw selects a capable channels.discord.defaultAccount. Otherwise, set accountId to the corresponding key under channels.discord.accounts; an omitted account is rejected as ambiguous:

{  "transcripts": {    "enabled": true,    "autoStart": [      {        "providerId": "discord-voice",        "accountId": "work",        "guildId": "1234567890",        "channelId": "2345678901",        "whenOccupied": true      }    ]  }}

whenOccupied defaults to false: capture starts with the Gateway and continues until stopped. Set it to true to wait for humans, then capture one meeting per occupancy episode. It also starts when humans are already present at startup; bots never count. After the last human leaves, a fixed 30-second grace period allows short reconnects without splitting the meeting. A human returning during that grace cancels the stop. Otherwise, OpenClaw stops capture and generates notes.

Occupancy episodes use generated IDs; an entry's sessionId is ignored. To continue a meeting across a Gateway restart, OpenClaw reopens the most recent session for the same provider, account, guild, and channel when it stopped within the last 10 minutes and its stored ID origin is generated. The session keeps its original ID, title, and start time, and new utterances append to it. A later return within that window also reuses the meeting; outside the window, capture gets a new ID. New admissions record whether the transcript ID was generated or supplied. If the newest candidate has a supplied ID, or its origin is missing or invalid, capture starts fresh without searching older history. Existing notes remain unchanged and readable. Legacy generated meetings without a recorded origin may therefore split after an upgrade or restart. Doctor preserves recorded origins when restoring metadata; it does not infer or backfill missing origins. If the room is routed to a different agent, that agent starts a new capture; the original agent retains its stored meeting and summary permissions.

The provider must report occupancy. discord-voice supports it; an unsupported provider logs a warning and skips the entry instead of capturing continuously. Configure at most one whenOccupied: true entry per Discord account and guild, even when the channel IDs differ: a Discord bot can occupy only one voice channel per guild. Later conflicting entries are skipped with a warning. For the complete listen-only setup, see Discord meeting notes.

The meeting provider ids are google-meet, teams, and zoom. Their aliases are googlemeet/meet, teams-meetings/microsoft-teams/msteams, and zoom-meetings, respectively. Meeting providers attach to an already-active meeting bot session; normal meeting joins do not need an autoStart entry.

Related

• CLI reference
• Meeting plugins — the plugins that capture these transcripts