Skip to content

Talk to Your Knowledge Base

Once your agent is connected, everything below is a conversation — no commands, no file paths to remember. Each scenario shows what you say, what the agent does, and what a well-behaved agent looks like (the conduct is built into the tools and skills, so you can hold it to this).

"Set up a knowledge base for me"

You: Set up a knowledge base called research at ~/kb/research.

A good agent: restates the resolved path — "I'll create /Users/you/kb/research and register it as research. Go ahead?" — and only then calls init_kb. It reports back the registered name and path.

Creating a directory is your decision; an agent that initializes without restating the path first is out of contract. The new KB is immediately queryable, and it opens in Obsidian later with zero migration.

"I already have one — register it"

You: Register /Users/you/vaults/work-kb as work.

A good agent: restates the path, calls register_kb, and confirms: "Registered work → /Users/you/vaults/work-kb. It's immediately queryable."

Registration writes the address book only — nothing inside your vault is touched. Two refusals you may see, both correct:

  • The folder isn't a Knowlery workspace yet. The agent should relay both fixes: an empty folder can be created with init_kb; a folder with existing notes needs one CLI command first (knowlery init works on brownfield folders without touching your notes), then register.
  • The name is taken. The agent should surface the conflict and ask — never silently pick another name.

"Remember this"

You: Save our conclusion about the caching strategy to work.

A good agent: calls capture and echoes the path — "Captured to inbox/2026-07-09-143012-caching-strategy.md. It's uncooked until the next /cook."

Captures land in inbox/ — never in your compiled knowledge. They are immediately findable by search and show up as uncooked notes in the staleness report, so nothing captured is ever lost, just not yet compiled. An agent should never capture silently in the background: only what you asked to save, always with the path echoed back.

"What do I know about…?"

You: What did we decide about the rollout? Check work.

The agent calls query and answers from compiled pages, with the source paths cited.

You: Not sure which KB it's in — search everything.

The agent queries with kb: "*" — results merged across every registered KB, each hit labeled with its KB.

And the answer you should want to hear sometimes:

The agent: Your knowledge base has no confident match for this.

That is the retrieval engine's abstention — a real answer, not a failure. A good agent relays it instead of padding it with guesses. If the question matters, the material that should answer it needs cooking.

"How's my knowledge base doing?"

You: Give work a checkup.

The agent runs health (workspace integrity) and stale (what's waiting to be compiled) and summarizes: "Healthy. 2 compiled pages have changed sources; 5 captures are waiting to be cooked."

You: Fix what's fixable.

The agent runs sync (if health reported missing skills or outdated files) and reports exactly which files were updated. For the stale pages and uncooked captures, it offers a /cook session — compilation stays a reviewed, human-in-the-loop act, so it won't cook without you.

stale output is a work list, not an alarm — it tells you what to re-cook when you next tend the vault. Many notes are legitimately never compiled.

"Compile my notes"

You: Cook the new material in work.

The agent loads the cook skill, reads the staleness report to scope the session (inbox captures first — they were captured precisely to be compiled), distills notes into compiled pages with sources: citations, and keeps SCHEMA.md's taxonomy in sync.

This is the one place agent judgment is supposed to be exercised — which is why it's a skill guiding the agent, not a tool it fires blindly.

The loop, end to end

A week with a knowledge base looks like:

  1. Conversations produce captures ("remember this") — zero friction.
  2. Questions get answers with citations — or honest abstentions.
  3. When abstentions accumulate around a topic, that's the signal to cook — the captures and notes become compiled pages.
  4. An occasional checkup keeps the workspace healthy.

Everything above also works from the CLI (knowlery query, knowlery stale, …) — same engine, same answers. See CLI Workflows for the terminal-first version, and Agents & MCP for the full tool reference and conduct notes.

Released under the MIT License.