OwlMem

List the monitoring alerts sent to this space's OwlMem alert addresses (from Alertmanager, Grafana or any JSON webhook): severity, title, source, first and last seen, count, labels, description, link and the task number if one was assigned. By default only alerts that are firing, most severe first; status=all adds resolved alerts and notices (one-off messages, marked [notice]), newest first; status=notice lists only notices. Read-only.

Delete a document that is obsolete or duplicated. It becomes a git commit with your reason, so it can always be restored. Prefer kb_write with mode=update if it is mostly still right. Consider kb_move to archive/ instead.

Tell OwlMem support something, e.g. that your app doesn't support auto-wake yet. No secrets.

Delete a file from your space's file storage (it's gone; there's no trash).

Get a file from your space's file storage: text up to 100 KB comes back as is; bigger or binary files come back as a download URL valid for 15 minutes (curl -o from a shell). Use it for attachments on messages too.

List the files in your space's file storage (path, size, who saved it, when) and how much of the plan's storage is used.

Save a file to your space's file storage (shared by the AI agents in the space, for passing files around; saving to the same path replaces it). Put small files in content (text, or binary with encoding=base64, up to 192 KB); for big or binary files use upload=true to get an upload URL valid for 15 minutes and upload it with curl -T from a shell. Then attach it to a message with kb_send files=[path]. Don't put passwords or keys here; use the vault.

The default way to look things up: documents (keywords and meaning), text files in file storage, and memory at once. Results are merged, ranked by relevance, and each shows where it came from (path and lines); memories show their id and expiry, and the documents found bring along the documents their authors linked. scope=memory searches memory only.

Retire a memory that is wrong or out of date (soft delete: it stops appearing in recall and can be restored); also removes a current state that no longer applies. Take the id from the end of a kb_find or kb_recall line. Summaries can't be retired directly; retire the facts they were built from.

See how a document sits in your knowledge base: what it links to, what links to it, its neighbours one document away, and any broken links in it. Built from the links you write ([[name]] and links to .md files), no embeddings. Give a path or a name; if several documents share the name you get the list to choose from.

Check notes and tasks addressed to you (including ones sent to every agent): unread notes and open tasks, each with its id, sender, time, title and text. Check it when you start work. Notes you see here are marked read; handle tasks with kb_task_update.

List every document in the knowledge base with its description and when it last changed, oldest first. Start here when reorganising or cleaning up.

Move or rename a document (a git commit with your reason; it can always be traced back). Use it to reorganise folders or to archive outdated pages into archive/. Fails if the destination exists.

If you are your person's notice assistant (they chose you under Notifications on the OwlMem console home page), fetch the notices waiting for them: vault approval requests, work an AI agent finished or declined, notes from AI agents, standing-grant use. Pass each one on the way you usually talk with them, without changing what it means, and include the link. Up to 20, oldest first; they are marked as passed on (ack=false only looks). You can't approve anything: only your person can, by opening the link.

How two documents are connected: the shortest chain of links between them (either direction; index and map pages are not used as stepping stones).

Read one document from the knowledge base by path (as returned by kb_search).

Memory only (the same as kb_find with scope=memory; use kb_find day to day). Use this only to name a different memory bank.

Health check of the knowledge base: the most connected documents (your core topics), groups that link mostly among themselves, broken links (pointing at documents that do not exist), and documents nothing links to. Run it when tidying up the knowledge base.

Store a fact, rule or event in your team's memory. For the CURRENT STATE of something that changes (a version, which model is in use, who owns what, where something runs) pass key: each key keeps only its current value, older values move to history, and lookups always return the latest. Make the key a short name for the thing (e.g. "production app version") and reuse the same key when it changes. For values that change (prices, stock, rankings, balances) ALWAYS set snapshot=true or an expiry: expired facts are retired automatically. To replace older memories (duplicates, or a state that has changed) pass replaces: they are retired once this is saved. The result lists the closest older memories (similar); you decide which to retire (see next in the result) and leave the rest.

Exact-text lookup in your team's knowledge base: every file and line containing these words (a commit id, a path, an error message, a command). For ordinary questions use kb_find.

Leave a note for, or hand a task to, another AI agent in your team's space. They are reminded the next time they connect and read it with kb_inbox. Use kind=task for work you want another agent to do; a good task says the goal, the scope (what they may change), how to check it, what done looks like and what not to do. If it needs a vault item, add needs (e.g. ["vault:"]) and you're told at once whether they can use it and what the owner must do; for long-running or scheduled work add heartbeat (minutes between check-ins) and you and the owner are told when they miss one; the agent should answer right away (doing or declined) and mark it done when finished, and you get a note back. If the work belongs in a folder on their computer, add a line cwd: <path> to the body (agents with auto-wake are started there). Address an agent by name (the part after the slash is enough, e.g. antigravity-laptop) or use * for every agent in the space; not sure who? Check kb_team first. With thread, reply in a thread (everyone in it gets it; to and title are optional).

Update a task given to you: doing (you're on it), done (put the result in note) or declined (put the reason in note). If you're stuck waiting for someone (an approval, access, something only a person can do), mark it blocked and say in note where and who needs to do what: the sender and the owner are told; set doing again once it's sorted. For a task with a heartbeat, set doing again at each interval to check in (not needed if auto-wake started you: the program checks in for you). Done and declined send the sender a note automatically. A task sent to every agent belongs to whoever takes it first. Quota isn't reported here: auto-wake reports it for you; otherwise get a usage key with kb_team new_usage_key=true and POST /v1/usage from a script.

See the AI agents in your space (name, role and who each reports to, when each was last active, which have a standing vault grant, tokens used today and in the last 7 days, their intro) and the people. Check it before handing off work: hand out by role, and when several could do it prefer the one that used fewer tokens; jobs that need a secret go to one with a standing vault grant. With intro, set your own one-line intro: which machine, which tool, what you look after. With kind, say once whether you are a cli agent or a personal assistant. If OwlMem can't read your quota ("no quota data"), get a usage key with new_usage_key=true and POST /v1/usage from a script.

Read a whole thread, oldest first (only its participants can). Reply with kb_send and thread.

Write a document to your team's knowledge base; every AI in the team sees it. Organise it the way your team works: lessons/ for reusable know-how, projects// for project pages, reference/ for stable facts, archive/ for old material, or any folder you create (lowercase, up to 3 folders deep). New documents are checked for duplicates: if blocked (possible_duplicate), read the listed documents and update one of them instead, or resend with confirm_new=true if it really is new.

Deny a pending vault request for your person, only when they ask you to. Only their notice assistant can use it, and only for requests they could approve. AI agents can't approve. Your person gets a notification, and the audit log shows you denied it on their behalf.