Knowledge base for agents

orboto's LLM-Wiki is where your agents keep what the next agent needs: hosts, conventions, decisions, where a credential lives (never the credential itself) and the lessons from incidents. Added in v0.208.0 Every project has one wiki space, and the workspace has an Operations wiki for facts that belong to no single project, such as hosts, runners, backup chains, budgets and who runs what. The wiki follows the LLM-wiki pattern: raw material goes in as a source, the wiki agent curates pages, keeps an index and appends an activity log, and a lint pass flags orphans, missing links, stale pages and contradictions.

How agents learn a project: rules, project facts, wiki

Added in v0.223.0 The wiki is one of three places where an agent learns a project, and each has its own job:

• Rules say how an agent must behave. orboto session-start prints them in full at the top. Agents do not change rules themselves: an agent that wants a new rule sends the text to you, and someone with rule rights adds it under Admin > Agent Rules.
• Project facts are the fixed decisions of a project: stack, commands, conventions, architecture. orboto session-start prints them right after the rules for every project the session works on (up to 12 per project, the rest in orboto primer KEY), plus the facts that apply to the whole workspace. An agent never changes a fact directly. It files a proposal, the fact is marked as observed in every primer, and a person with edit rights on the project confirms it.
• The wiki holds everything else the next agent needs: hosts, where a credential lives, lessons from incidents, procedures. Agents ask it when they need something (orboto knowledge ask, orboto knowledge find) and add to it with orboto knowledge add.

The start of every session names this rule: a decision becomes a fact proposal, a lesson or procedure goes into the wiki, a behaviour rule becomes a rule proposal. The in-app assistant follows the same rule.

Every project starts with a wiki

Added in v0.223.0 When you create a project, orboto turns its project space (the space that holds the project primer) into the project's wiki right away. It comes with a "Wiki schema" page that describes how pages are organised, an index and an activity log, and agent edits are applied directly. Your agents can ask the wiki, search it and add facts to it from the first day (orboto knowledge ask, find, add). Snapshots of closed tickets and merged pull requests stay off until you turn them on in the project's LLM-Wiki settings.

The workspace setting. Admin -> System settings -> "New projects get a wiki" is on by default. Turning it off affects only projects created afterwards. Command line: orboto knowledge settings --new-projects on|off.

Existing projects. On the first start after the update, orboto turns on the wiki for every existing project whose space never had one. An admin can repeat this at any time with orboto knowledge backfill (--dry-run shows what would change; a space that fails is listed and retried by the next run). Projects whose wiki someone switched off, and projects that already use another wiki space, are left as they are. Nothing already in your projects is processed, so this costs no AI calls.

Switching one project's wiki off. Project settings -> LLM-Wiki -> "Switch the wiki off", or orboto knowledge disable --project KEY. All pages stay. The wiki stays off until someone turns it on again (the same button, or orboto knowledge setup --project KEY).

What it costs. Turning a wiki on makes no AI call apart from embedding the schema page once, when an embedding provider is set. AI calls start when the wiki is used: one per knowledge add or processed source (a longer fact, an imported file or web page, a wiki email, a ticket or pull request snapshot if you turned those on), one per question to knowledge ask plus an embedding, and one weekly check for contradictions and missing pages once the wiki has curated pages. An empty wiki and the generated project primer cost nothing. All calls run on your workspace's AI provider, your own key or the managed one, within its budgets. Without an AI provider nothing is processed: new sources wait as failed runs you can retry later (orboto wiki runs), and knowledge find keeps working.

Turn the wiki on

For a space that is not the auto-generated project space, or after the wiki was switched off, a person with wiki configuration rights runs:

orboto knowledge setup --project ACME

This turns the project's space (the wiki space linked in the project settings, otherwise ACME-S1) into a curated wiki: an index and an activity log are created, edits apply automatically, and a "Wiki schema" page describes how pages are organised. The project is linked to the space, so closed tickets and merged pull requests can feed it (--emit-on-close, --emit-on-merge). A super-admin creates the Operations wiki the same way with orboto knowledge setup --operations.

What agents do

Every agent reads the same rule in the workspace rules: durable facts go into the wiki on the same act as the work, never only into the agent's local memory, and the wiki is asked before a person.

orboto knowledge ask "where does the nightly backup run?"
orboto knowledge find "restic"
orboto knowledge add "Build host" --project ACME --content-stdin < fact.md

ask and find read the wikis of the projects the session declared plus the Operations wiki. add writes a short fact into a wiki page right away and files a longer text as a source that the wiki agent curates into pages. MCP clients use orboto_knowledge_ask and orboto_knowledge_add; the wiki verbs themselves are orboto wiki ingest|ask|record|lint|enable|runs|retry.

Added in v0.221.11 knowledge add only adds. A fact that belongs on an existing page is appended to it as a new section; the page's title and its existing body stay as they were. A wiki write that would replace a page and drop more than half of its content is refused with 409 errors.docs.replace_drops_content unless the writer asks for a replacement explicitly (orboto wiki record --replace). A question answer saved back to the wiki follows the same rule. Re-running an add is safe: a repeat with the same title and content within 30 minutes files nothing and answers duplicate with the first source or pages, so a timed-out call can simply be sent again. A secret-shaped value is refused with 422 errors.content.secret_refused; store it in the credential vault instead.

Added in v0.211.1 Every doc response carries a webUrl in the form /spaces/<SPACE-KEY>/docs/<DOC-KEY>, in the REST API, in orboto_get_doc and the knowledge tools, and in orboto doc get. A doc or space without a key keeps its UUID for that part of the link. Agents post the link as returned and never build it themselves.

At session start, and in the project primer, agents get a Knowledge block: the index of each wiki, its five latest log entries and any open lint issues.

When knowledge stays local

With the orboto Claude Code hook installed (orboto mcp install --session-reporting), a turn that wrote Claude Code's local memory without writing to the wiki does not end quietly: the completion check lists "knowledge written locally, not in the wiki" and names the command that files it.

What you see as the operator

• Admin, Agents, Live has a Knowledge column per session: docs written and questions asked in the last 24 hours, and a "local" badge while local memory is ahead of the wiki.
• Admin, LLM-Wiki, Ingest runs lists failed curation runs with their error and a Retry button (orboto wiki runs <space> --status failed, orboto wiki retry <space> <run-id>).
• Admin, AI Settings lists the wiki operations (ingest, edit plan, lint) with their own model, answer length and reasoning settings. Ingest and plan default to 16384 answer tokens and low reasoning, because a curation plan carries the full text of every page it touches.