Hemory

Dismiss a voice that is not identified and not worth tracking (noise, TV, a passer-by), like 'ignore' in the app; or restore it with dismissed=false. Applies to that voice across all history. Identified people cannot be dismissed. Arguments: recording_id + speaker_label (as in label_speaker); dismissed (default true). Only use this tool when the user asked you to dismiss or restore a speaker. Result: '# ok: ...' or '# error: <code> — <reason>'. Codes: recording_not_found, speaker_label_not_found, no_voiceprint, already_identified.

dismiss_speaker

Label who a speaker is, exactly like tapping a speaker name on the app's replay transcript: the decision applies to that voice across ALL history, not just one line. Arguments: - recording_id + speaker_label: from a search_memory line (the local label in the speaker column, e.g. 'speaker_2' or 'Alice:speaker_2' -> speaker_2) or from a list_unidentified_speakers sample line. - Exactly one of speaker_id (an existing person from list_speakers) or new_speaker_name (creates the person). - confirm: default false, see below. Behavior: - A voice that is not identified yet is labeled immediately. - A voice already identified as someone else is NOT changed: the tool returns '# confirm_required: ...' with the impact. Tell the user, and call again with confirm=true only after the user explicitly agrees in this conversation. - Only use this tool when the user asked you to identify or label speakers. Result: '# ok: ...' (done), '# confirm_required: ...' (nothing changed) or '# error: <code> — <reason>' (nothing changed). Codes: invalid_arguments, recording_not_found, speaker_label_not_found (call search_memory again for fresh labels), no_voiceprint (this speaker cannot be labeled), speaker_dismissed (call dismiss_speaker with dismissed=false first), speaker_not_found, speaker_name_conflict (use the returned speaker_id instead).

label_speaker

List what one speaker said across recordings, starting from one of their lines, to gather more evidence before deciding who a voice is. Usage: - recording_id + speaker_label: the recording_id and local label of any line (from search_memory, list_unidentified_speakers, or this tool). - If that voice is already identified as a person, all of that person's voices are included; otherwise only that voice. - limit: recordings per page (default 10, max 20). Paging: pass the header's cursor back as `after`. Output (text): # speaker=<name> (<speaker_id>)|unidentified[:dismissed] | voices=<N> | recordings=<N> | more=<bool> [| cursor=<tok>] ## <recording_id> | started=<local time> | lines=<N> <local time>\t<recording_id>\t<local_label>\tctx=<ISO>~<ISO>\t<text> (up to 3 longest lines per recording, in time order) - Recordings are newest first; lines = this speaker's lines in that recording. - Local labels differ between recordings, and one person may have two labels in one recording: always use the label printed on the line. - label_speaker / dismiss_speaker on a line changes only that line's voice, not every voice of the person. - ctx is a ready-made window around the line: pass the two values as started_at / ended_at to search_memory (keywords=[]) to read the surrounding conversation. - Errors: recording_not_found, speaker_label_not_found, no_voiceprint.

list_speaker_lines

List the people in this vault's speaker library, most recently heard first. Usage: limit (default 50, max 100); paging via the header's cursor passed back as `after`. Output (text): # total=<N> | more=<bool> [| cursor=<tok>] <speaker_id>\t<name>\tdisplay=<display name|->\tvoices=<N>\trecordings=<N>\tlast=<local date|->\tnotes=<notes|-> - name is unique in the library (case-insensitive); display is an optional shown name. - voices = voiceprint clusters bound to the person. notes = free-form context about the person (role, relationship, how others address them); it is the last field and line breaks are shown as ' / '. - Check this list before creating a person so the same person is not added twice.

list_speakers

List voices in this vault that are not identified yet (the same list as the app's 'unidentified speakers'), most talkative first, optionally within a time window. Usage: - started_at / ended_at: ISO 8601 with timezone (Z or +HH:MM). local_day: YYYY-MM-DD in the user's timezone, used only when started_at/ended_at are omitted. Omit all for all time. - limit: voices per page (default 10, max 20). Paging: pass the header's cursor back as `after`. Output (text): # window=<...|all> | total=<N> | more=<bool> [| cursor=<tok>] ## unidentified <n> | lines=<N> | recordings=<N> | last=<local time> <local time>\t<recording_id>\t<local_label>\tctx=<ISO>~<ISO>\t<text> (up to 3 sample lines per voice) - lines / recordings / last are counted inside the window. - All sample lines under one '## unidentified' block are the SAME voice. The recording_id + local_label of any of them addresses that voice in label_speaker / dismiss_speaker, and labeling it changes all of its lines in all history. - ctx is a ready-made window around the line: pass the two values as started_at / ended_at to search_memory (keywords=[]) to read the surrounding conversation before deciding who is speaking. - For more of a voice's lines in other recordings, pass any sample line's recording_id + local_label to list_speaker_lines.

list_unidentified_speakers

Maintain the speaker library. - action=create: name (required), display_name, notes. Adds a person immediately. - action=update: speaker_id plus any of name, display_name, notes. Omitted fields stay unchanged; an empty string clears display_name or notes. A notes-only update applies immediately; changing name or display_name renames the person everywhere and needs confirmation. - action=merge: source_speaker_id, target_speaker_id. Moves every voice of the source person to the target, appends the source notes to the target, then deletes the source. Needs confirmation. - action=delete: speaker_id. Deletes the person; their voices return to unidentified. Needs confirmation. Confirmation: with confirm=false (default) an action that needs confirmation changes nothing and returns '# confirm_required: <impact>'. Tell the user, and call again with confirm=true only after the user explicitly agrees in this conversation. Use notes for context that helps identify the person later (role, relationship, how others address them). Only use this tool when the user asked you to maintain speakers or to remember something about a person. Result: '# ok: ...', '# confirm_required: ...' or '# error: <code> — <reason>'. Codes: invalid_arguments, speaker_not_found, speaker_name_conflict, same_speaker.

manage_speaker

Search this vault's section transcripts by keyword and/or UTC time window. Returns matching spoken lines in global spoken-time order (merged across recordings). Usage: - keywords: case-insensitive substring, OR-matched. EMPTY keywords = no text filter -> pull the full timeline (use with a time window and paging). - started_at / ended_at: ISO 8601, MUST include timezone (Z or +HH:MM); converted to UTC. Omit for no bound. A naive datetime is rejected. - local_day: YYYY-MM-DD in the USER'S timezone (resolved server-side). Use this for 'today'/a specific local day instead of computing UTC yourself. Ignored if started_at/ended_at given. - Paging: results are spoken_at-ascending. more=true means more transcript remains to scan in the window (NOT that more matches exist) - pass the header's cursor back as `after` to continue. Keep paging until more=false; only then is 'nothing matched' conclusive. - This is lexical + time search, NOT semantic - no paraphrase matching. Output (text, grep-style): # kw=<...|*> | window=<...|all> | scanned=<N> | more=<bool> [| cursor=<tok>] ## <recording_id> <YYYY-MM-DD> <YYYY-MM-DD HH:MM:SS (Region/City)>\t<speaker|->\t<text> (NOT ISO 8601 - no 'T' separator, no UTC offset; times are localized to the recording's capture timezone, falling back to user timezone, or '(UTC, timezone unknown)' if neither is known; date in group header matches the localized date of the first line in that group; text is the last field, may contain spaces/colons.) The header has NO total-hit count (this is a bounded streaming scan); scanned=N is how many segments this request examined. When a page has no matches, a `# note: ...` line explains whether to keep paging or stop. Speaker column: [<name>:]<local_label>[:dismissed], or '-' when the line has no speaker. - 'Alice:speaker_1' = identified as Alice; 'speaker_2' = not identified; 'speaker_3:dismissed' = a voice the user dismissed. - Parsing: drop a trailing ':dismissed'; the part after the LAST ':' is the local label and everything before it is the name (local labels never contain ':'). - A local label only means something inside its '## <recording_id>' group: speaker_1 in two different recordings is NOT the same person. recording_id + local label is what the speaker tools (label_speaker, dismiss_speaker, list_speaker_lines) take; list_speaker_lines shows that speaker's lines in other recordings. - When answering the user, refer to people by name (or 'an unidentified speaker'); never repeat local labels such as speaker_1 to the user. Examples: - Local day: keywords=[], local_day="2026-06-05" (server resolves to user's timezone automatically) - Whole day: keywords=[], started_at="2026-06-05T00:00:00Z", ended_at="2026-06-06T00:00:00Z" - Keyword: keywords=["成本","预算"] - Next page: after="<cursor from previous header>"

search_memory