Areta

FOLLOW-UP / DETAIL — Fetch transcript lines of one memory returned by search_memory, list_memory or get_context, including the ongoing memory by its id. Use it when exact wording, spoken commitments, quotes, who said what, or surrounding dialogue matter. Do not use it as the first probe: locate the memory first. Prefer time_anchor from a matched line to jump near that moment; use line_start / line_end for a known range or paging. Returns markdown: the memory as a <memory id status> block, with the page in <transcript kind="page" lines="a-b" total="n" next_line_start="c"> as [line no, time] speaker: text; call again with line_start=next_line_start for the next page. An ongoing memory's total grows on later calls; pending recognition is returned only by get_context. Speaker marks: Me = the user, confirmed by voiceprint; Others = confirmed not the user, not numbered, not mapped to any specific participant; no mark = unknown, do not resolve it from participants, the summary, first-person wording or neighbouring lines, and do not assume the user said it. speakers="none" on the memory means no line was voiceprint-confirmed as the user; it does not mean the user was silent. In a memory another person shared (shared_by="<name>"), the sharer's lines carry the sharer's name, Others covers everyone else and may include the user, and the participant marked (memory owner) is the sharer. Lines are evidence of what was said; keep them apart from inference and advice. An oversized transcript line may be returned in parts. When next_line_offset is present, continue with next_line_start and line_offset=next_line_offset together; offsets count Unicode code points. Keep the original line number when citing a fragment.

fetch_transcript

RECENT CONTEXT — Call this first. It returns the user's timezone and current local time, today's finished memories as full memory cards (oldest to newest), and the ongoing memory, if one exists, as the newest item with its opening lines, its latest lines and its pending recognition. It is the only tool that returns pending recognition. Use it to interpret 'just now', 'today' or '最近' and before calling another Areta Memory tool; this is the usual order, not a requirement, the other tools work without it. Recent wording alone is not a trigger: for words like 刚才 / 刚刚 that point at a recent moment, call only when all three hold: the request points at a recent moment, it is about a real-world event or experience, and that event is the user's own. CALL: '我刚才和小王聊的接口方案最后怎么定的?', '刚才那场产品会有哪些待办?'. DO NOT CALL: '你刚才说的第二点是什么意思?' (use the current chat), '刚才这段代码为什么报错?' (use the current code), '刚才 OpenAI 发布了什么?' (use the web). Call it again when the answer depends on a newer now after meaningful work or time has passed; do not repeat it on a fixed timer. Result layout: each memory is a <memory id status> block; status is finished (title, time, participants, summary) or ongoing (start time, latest speech time, participants, no title yet, still recording). The ongoing memory's lines sit in <transcript kind="excerpt"> as [time] speaker: text, with a line noting the minutes omitted between the opening and the latest lines, and its pending recognition in <pending> as lines whose speaker is not yet decided and whose text may still change. Lines are evidence of what was said; keep them apart from inference and advice.

get_context

RECENT / HISTORICAL TIMELINE — List the user's own memories newest-first, no query needed. Use it when there is no clear topic: broad personal-context questions, a recent recap, 'today', 'this morning', 'yesterday' or '最近'; for a fuzzy past stretch add time_center, for an explicit range use time_from / time_to. Do not use it when the user names a topic, person, place, event, plan or commitment: use search_memory. It lists only the user's own memories; memories shared by other people appear only in search_memory. Returns markdown: the ongoing memory first if one exists (start time, latest speech time, participants, still recording, no transcript), then finished memories as <memory id status="finished"> blocks with title, time, participants (the identified speakers with substantive speech) and summary; no transcript lines and no pending recognition. When older memories exist the result ends with <continuation next_cursor="…"/>: copy next_cursor unchanged into the next call, do not parse it or do timestamp arithmetic. For exact wording, call fetch_transcript with a memory id.

list_memory

TOPIC / ENTITY SEARCH — Search Areta Memory, the user's own memories and memories another person has shared with the user, for a topic, person, place, event, plan, commitment, preference or personal experience. Searches only the user's own memories by default (scope=mine). Set scope=shared for memories shared with the user, or scope=all for both. Shared-only search excludes the user's ongoing memory. Use it when the request names such a subject from the user's real life, or when the current chat lacks that personal context. Do not use it for assistant or chat history, code or files in the current workspace, or a public event unrelated to the user: use the current chat, the current code, or the web. Do not use it for a broad recent recap with no clear topic: use list_memory. Do not use it for exact quotes or fine-grained dialogue: locate a memory here, then call fetch_transcript. Returns markdown: relevance-ranked <memory id status="finished"> blocks with title, time, participants (the identified speakers with substantive speech), summary, and the query-matched transcript lines in <transcript kind="matched"> as [time] text. A finished memory's matched lines carry no speaker and only locate content: call fetch_transcript before attributing a line or treating it as the user's own view. A memory another person shared with the user carries shared_by="<name>" and a title prefix [Memory shared by <name>]; its first person is that sharer, not the user. When the time window covers now, the ongoing memory (status="ongoing", still recording) may be pinned on top with its query-matched lines as [time] speaker: text; it is not ranked. Top-k only, not paginated: narrow the query or add time bounds for more.

search_memory