kumiho_chat_clear
ChatGPTDeletes the messages in this conversation's temporary working buffer. Useful when the user asks to clear or reset the buffer for this conversation. Saved long-term memories and other conversations' buffers are not affected. Cleared messages cannot be recovered. Reuse the session_id returned for this conversation only. If the host supplies no conversation ID, the first call without session_id returns session_required and a new ID without reading or writing memory; retry with that ID. Never reuse an ID from another conversation or invent one.
kumiho_chat_get
ChatGPTReads the messages held in this conversation's temporary working buffer, with their count and the buffer's remaining time to live. Useful when the user asks what the buffer for this conversation currently holds, or before summarizing or clearing it. The buffer is short-lived and separate from saved long-term memories. Read-only. Reuse the session_id returned for this conversation only. If the host supplies no conversation ID, the first call without session_id returns session_required and a new ID without reading or writing memory; retry with that ID. Never reuse an ID from another conversation or invent one.
kumiho_create_space
ChatGPTCreate a new space (folder) within a project. Spaces organize items hierarchically.
kumiho_deprecate_item
ChatGPTRetires one saved memory so it no longer appears in normal searches. This is how a user's request to forget something is carried out. Useful when the user asks you to forget, stop using or retire a specific memory, and to retire the memory a user's correction replaced when the correction did not stack as a new revision of it. It changes what the user's searches return, but it is reversible: the item and its revision history stay in the workspace, and calling again with deprecated set to false restores it. It is not permanent erasure; say so if the user expects deletion. Arguments: item_kref is the reference of the exact item (kref://project/space/item.kind; a trailing ?r= revision suffix is ignored). Make sure it is the memory the user means, and confirm with the user when more than one memory could match. Each call retires one item.
kumiho_get_item
ChatGPTGet a Kumiho item (versioned asset) by its kref URI. Example: kref://project/space/item.kind
kumiho_get_provenance_summary
ChatGPTGet a summary of the provenance (lineage) of a revision, including used models, seeds, and parameters from upstream dependencies.
kumiho_get_revision_by_tag
ChatGPTGet a revision by its tag (e.g., 'latest', 'published', 'approved').
kumiho_get_spaces
ChatGPTGet spaces (organizational folders) within a Kumiho project.
kumiho_list_projects
ChatGPTList all Kumiho projects accessible to the current user. Returns project names, descriptions, and metadata.
kumiho_memory_consolidate
ChatGPTSaves a summary of this conversation as one long-term memory in the user's private Kumiho workspace, then clears this conversation's temporary working buffer. Useful when the user asks to save a summary of the conversation, to wrap up, or to keep a record of what was decided before leaving. Clearing the buffer cannot be undone; the saved summary remains. Arguments: always pass summary. This service does not write summaries itself, so a call without one fails; with one, the call works even when the buffer is empty. Write the summary for a reader who was not there: what was decided and why, durable facts, open items. Plain text works; an object with title, summary and knowledge (facts, decisions, actions, open_questions) is recalled better. Keep it brief and distilled, never the transcript. Not for: passwords, access tokens, API keys, MFA or recovery codes; leave such values out of the summary. Reuse the session_id returned for this conversation only. If the host supplies no conversation ID, the first call without session_id returns session_required and a new ID without reading or writing memory; retry with that ID. Never reuse an ID from another conversation or invent one.
kumiho_memory_decompose
ChatGPTAdd typed entities, facts and relationships to a stored workspace memory. Supersession can replace an older fact's accepted status and mark dependent evidence stale; prior revisions remain in history. Requires ontology support.
kumiho_memory_engage
ChatGPTSearches the user's saved Kumiho memories for context relevant to the current request. Returns a ready-to-use context summary, the matching memories and their references (source_krefs); how many come back varies. A successful search can return no matches. Check backend_error: a backend error means retrieval failed. When optimization.status is applied, zero means no retrieved candidates passed the evaluation; it does not prove no relevant memory exists. Useful near the start of a conversation about the user's own ongoing work, projects, decisions or preferences when earlier saved context could change the answer; when the user refers back to something that is not visible in this conversation ("as we decided", "my usual setup", an unfamiliar project name); and when the user asks what you remember about them or what was decided or noted before on a topic ("what do you remember about me?", "what did we decide about pricing?"). Matches are ranked by relevance to the query, not by date. The returned references can be attached to a related memory saved later, linking it to its sources. Not for: saving, finding or checking passwords, access tokens, API keys, MFA or recovery codes, so do not call this tool for such requests, not even to check first; data from another organization or any workspace the connected account is not authorized for; general knowledge or live information such as weather, news or prices. It reads only this account's saved memories and writes nothing. query: a brief natural-language description of what you are looking for, with no secrets in it. Each explicit call runs retrieval again, including an immediate repeat of the same query.
kumiho_memory_recall
ChatGPTFiltered semantic search over the user's saved Kumiho memories. Returns the best matches with titles, summaries, relevance scores and references. Useful when the search has to be limited by memory type (memory_types, for example ["decision"] or ["preference"]) or by location (space_paths, for example one project's space), such as listing the decisions filed in a known space. query: describe what you are looking for in natural language rather than guessing exact keywords. Each explicit call runs retrieval again, including an immediate repeat of the same query. Not for: saving, finding or checking passwords, access tokens, API keys, MFA or recovery codes, so do not call this tool for such requests; data outside the connected account's authorized workspace; general knowledge or live information such as weather, news or prices. Read-only.
kumiho_memory_reflect
ChatGPTSaves distilled memories to the user's private Kumiho workspace and a short note of your reply to this conversation's temporary working buffer. Useful when the user asks you to remember, save or note something ("remember this", "save this", "note that", in any language), and when something durable is settled: a decision with its reason, a lasting preference, a durable fact about the user, or a correction to something saved earlier. Not needed for routine turns or small talk. Arguments: response (required) is a one- or two-sentence gist of your reply. Each capture needs type (decision, preference, fact and so on), a short title with absolute dates ("Chose Seoul on 2026-09-16") and brief content in your own words, never transcripts. space_hint takes an existing space, copied exactly as results show it ("/CognitiveMemory/preferences"); without one a capture is filed at the project root as a separate item. source_krefs links captures to their sources. To correct a saved memory, set that capture's revises to the memory's reference as search results return it; the capture becomes that memory's new revision. Keep its memory type and language, and restate the subject. If you cannot tell which memory to revise, save normally and retire the old memory by its reference only when stored_krefs names a different item. Not for: passwords, access tokens, API keys, MFA or recovery codes, payment details, or anything the user asked not to keep, so do not call this tool for such requests. Reuse the session_id returned for this conversation only. If the host supplies no conversation ID, the first call without session_id returns session_required and a new ID without reading or writing memory; retry with that ID. Never reuse an ID from another conversation or invent one.
kumiho_memory_retrieve
ChatGPTLookup of the user's saved memories by date or by a known title, keyword, topic or space. Useful when the user asks for their most recent or oldest memories ("what did I save recently?", "my latest note on the launch plan", "the newest decision in this space", "the first thing I saved"), and when you already know which memory you need and need its exact item or revision reference, for example to link to it or retire it. Returns item references, revision references and scores, not the memory text; a returned revision can be read for its text. Arguments: for the most recent memories, set mode to "latest". Results come newest first by last update (an older memory that was updated moves forward) and carry created_at dates. A query, keywords or topics narrow them to relevant matches, still newest first: for "my latest note on the launch plan", set query to "launch plan". Matching is fuzzy, and when nothing matches the newest memories in scope come back with score 0, so check that results fit. mode "first" returns the single oldest memory, or with a query the oldest relevant match. Both modes honor space_paths (for example one project's space) and memory_types (for example ["decision"]). In the default search mode, results are ranked by relevance; query takes the words you expect in the title or summary, and keywords and topics add terms. Not for: saving, finding or checking passwords, access tokens, API keys, MFA or recovery codes, so do not call this tool for such requests; data outside the connected account's authorized workspace; general knowledge or live information such as weather, news or prices. Read-only.
kumiho_memory_space_profile
ChatGPTProfile each Space's knowledge dynamics: aggregate churn/evidence/stability signals, classify Spaces as canonical/working/correspondence, and persist versioned space-profile items. Pure aggregation — no LLM calls. A space_class Space attribute pins the label (drift is then reported only). Use dry_run=true to classify without persisting.
kumiho_memory_store
ChatGPTWrites one pre-formed memory entry to the user's private Kumiho workspace and returns its reference. The entry arrives as it should be kept: its text plus, optionally, a title, summary, memory type and space. Useful for a single self-contained entry whose content and placement are already decided. By default it looks for a similar memory in the same space and adds a new revision to it instead of creating a duplicate; earlier revisions are retained. Set stack_revisions to false to always create a separate memory. May create the space, a bundle and provenance links. Arguments: user_text (required) is the entry's brief text, in the user's words or a close paraphrase, never a conversation transcript. Add a short title, a one-sentence summary, memory_type (decision, fact, summary and so on) and an existing space name, copied exactly, in space_path or space_hint when known. Not for: passwords, access tokens, API keys, MFA or recovery codes, or payment details, so do not call this tool for such requests; text that looks like a credential is rejected.
kumiho_search_items
ChatGPTSearch for items across Kumiho projects and spaces. Supports filtering by context (project/space path), name, and kind (model, texture, workflow, etc.).
