Answers your agents can stand behind

Agents see what you published

• Every edit is a version Saving an FAQ or a skill makes a new version. Agents keep getting the published one until a new version is published.
• Draft, published, archived Drafts and archived entries are never discovered. Publish an older version to roll back; the newest 20 versions are kept, and the published one always.
• Who decides Members write drafts and propose. Admins and owners publish, archive and review the inbox; with the audit log on, each of those is recorded.

Found by meaning, loaded when needed

• FAQs A task or a user's question in their own words finds the published answer that fits, with a similarity score, no matching keywords needed.
• Skills Named markdown instructions, up to 64 KB each. Discovery returns a skill's name and description; the agent loads the whole body only when the task needs it.
• HTTP and MCP POST /discover and GET /skills/{name} over HTTP, or playbook_discover and playbook_get_skill from any MCP client.

Agents propose, a person approves

• From agents An agent sends a question, an answer and the evidence it rests on with playbook_propose. Nothing reaches other agents until someone accepts and publishes it.
• From agent memory Link an agent memory store and its language model reads new memories and session summaries on a schedule and proposes FAQs, each citing the memories it came from.
• No repeats A proposal that says the same as an FAQ or a pending one is not added. The inbox holds 500; past that the oldest are dropped and the console shows how many.

Approved answers, served from the cache

• Published means cached Link a semantic cache and every published FAQ is put there; a new version replaces it and archiving removes it.
• No model call Your application's cache lookup now answers those questions with the approved text, before any language model is asked.
• Kept fresh Published answers are put with a year's expiry and refreshed every day, so an answer stays cached while it stays published.

In your own database

• Your keys FAQs, skills, versions and proposals are stored in the database you choose, under kividb:pb: keys, with a search index beside them.
• Limits 2,000 FAQs and 200 skills a playbook, 20 playbooks a database. Embeddings run on a model inside KiviDB by default, or your own provider.
• No extra charge It uses the database's own memory. Hosted models, if you choose them, are billed per token as usual.

Questions

Can an agent change what other agents are told?

No. An agent can only propose. A person with the admin or owner role accepts and publishes, and agents only ever see published versions.

What is the difference from agent memory?

Agent memory keeps what each user said and did. A playbook holds the answers and procedures you approved for every user, and memory can feed it proposals.

Can I upload a skill as a ZIP?

Not yet. A skill is markdown with a name, description and tags, written or pasted in the console or sent over the API.

Does it work on sharded databases?

Not yet. Playbooks run on databases without shards.

Give your agents a playbook

Create a playbook in the console's Agent playbooks section, publish a few answers, and point your agent at playbook_discover.