capture_conversation_turn
ChatGPTDefault WRITE tool — call after any turn where the user shared new durable facts, plans, deadlines, preferences, or decisions, OR pasted/uploaded a reference document (company profile, article, PDF extraction, knowledge base). Pass the FULL userMessage + assistantMessage — the server detects content type and routes automatically: conversation turns → distills 0–4 short classified memory cards (facts/beliefs/preferences); structured documents → indexes all content as searchable chunks (no truncation). Never summarize or truncate before calling. Returns { ingested, statements, skippedReason?, documentId?, chunkCount? }. Do NOT call after read-only questions (e.g. "what do I have Monday?"). Use save_memory only for a single line the user explicitly asked to remember verbatim — never both for the same content.
create_upload_folder
ChatGPTCreate an upload folder (group) for uploaded documents. Source stays upload. Pass the returned slug as group when searching or listing files in that folder.
delete_memory
ChatGPTExplicit-user-request only — permanently delete one memory by id (irreversible). Confirm with the user before calling.
edit_memory
ChatGPTExplicit-user-request only — correct an existing memory by id when the user fixes a fact or wording (re-embeds server-side; category/placement unchanged).
export_memory
ChatGPTExplicit-user-request only — return a full snapshot of the user's memory (optionally filtered by category) when they ask to see or download everything. Not part of the normal read path.
fetch_raw_source
ChatGPTFetch the full original source (email body, Drive file, Slack thread, Zoom transcript, uploaded document) behind a memory item. Use when the user asks for the exact text of an email or document and the memory context only has a summary. Pass the [ref:…] id that appears at the end of any memory line in the context block — e.g. if you see '[ref:abc-123]', pass 'abc-123'. Returns { content, source, metadata } with the complete untruncated source text.
get_verified_context_package
ChatGPTThe user's second brain — permanent unified memory spanning their Gmail, Google Drive, Google Calendar, Slack, Notion, Zoom meetings/transcripts, LinkedIn profile, uploaded documents, and all past conversations. Contains their projects, clients, colleagues, emails, files, meetings, schedule, decisions, tasks, preferences, and professional history. Call this FIRST on every turn. Triggers include (not limited to): - ANY mention of emails, messages, files, docs, meetings, calls, calendar, or schedule — this tool has that data; do NOT say you lack access - people, companies, or projects not defined in this conversation - 'my/our' references, past decisions, deadlines, tasks, plans - requests for advice, drafting, or recommendations (context shapes output) - start of any new conversation Unlike individual connectors, this returns synthesized context across ALL sources at once. Answering without this context is answering about a stranger.
list_source_folders
ChatGPTList every folder the user's memory can be scoped to: their upload folders and the Google Drive folders they selected for sync. Call this first when the user refers to a folder, then show them the options and pass the chosen group to query_memory / retrieve_context — do not guess or invent a folder name. Each entry has kind (upload | drive), name to show the user, and group to pass back. Only folders the user already chose are listed; this never enumerates their whole Drive.
list_upload_folders
ChatGPTList the user's upload folders (groups). Uploaded files keep source=upload. Use the slug as group on query_memory / list_source_items. Humans and this agent can both create folders.
query_live_sources
ChatGPTLive API call to every third-party source the user has connected (Gmail, Google Drive, Google Calendar, Slack, Notion, Zoom, Fathom, Otter, Microsoft 365, LinkedIn). Returns live metadata AND live content in each item.content: full email body, Drive/Doc/Sheet text, Notion page or database rows, Slack channel/thread text, calendar event details, meeting transcript/summary when the provider sends it. Do not call fetch_raw_source or query_memory as a substitute — this tool already fetched the live body. If content is missing, the provider did not return extractable text; do not invent it or fill from memory. Use for current inbox / calendar / Drive / Slack / Notion / connected-source questions ("what is in my inbox", "open Guest Box", "testers control sheet"). Omit sources to call all currently connected providers. Pass sources to restrict. If a source is not_connected or error, it has no items — do not fabricate replacements.
query_memory
ChatGPTTargeted deep-search of the user's memory for ONE specific named topic, person, project, or event. Use AFTER get_verified_context_package when you need more depth on a single subject it surfaced. Not the default entry point — get_verified_context_package is. Pass the specific subject as the query (not the whole message). Returns JSON { items, count, verifiedContext?, note?, profileSummary? }.
save_memory
ChatGPTSave ONE single line the user explicitly asked you to remember, stored verbatim. Use ONLY for an explicit "remember this" instruction. For new facts, plans, deadlines, or preferences the user shared in passing, use capture_conversation_turn instead — never both for the same content.