Chief — reference

Chief is a durable memory of your work. It reads your email, calendar, documents, messages, meetings and notes, and turns them into claims — short statements of what happened — each carrying the evidence it came from and the provenance of who asserted it.

That last part is the whole design. A vector store can find text that looks like your question. Chief can tell you what the evidence established, where it came from, who said it, and when — which is the difference between a search box and a memory you can act on.

Ways in

All five read and write the same memory, under the same account.

The skill is the default way into an agent, not MCP. Every harness can run a command; only some speak MCP. And what an agent is missing is rarely the ability to call Chief — it is knowing when to, which is what the skill carries. Reach for MCP when the harness will not shell out. Run chief skill --install to set up Codex and Claude Code without repository access, then follow the beta setup check to verify the whole path.

The two layers, and why it matters

Chief has indexed memory and live sources. Choose the layer that can establish the fact you need.

Indexed memory has two reads. chief search reads the original artifacts directly — the email, the note, the document — and needs no prior claim or person lookup; start source investigation there. chief recall adds the interpreted layer: indexed claims, which record what evidence established, alongside the original passages that preserve the source words. Both are indexed and cross-source: one question can reach an email from March, a meeting note from June and a document from last week. Neither can tell you what is currently outstanding, because "outstanding" is a property of the world right now, not of the record.

Live sources are the provider's own state, queried directly. Slower, narrower, current. This is where "what meetings do I have this week" and "which emails are unanswered" get answered.

chief search "what did we decide about pricing"     # indexed original artifacts — start here
chief recall "what did we decide about pricing"     # interpreted memory: claims and passages
chief source search calendar --since 0d --until +7d # the live calendar

Open loops are a shape, not a phrase

"This thread is unresolved" is a property of a claim's structure, not of its wording, so no phrasing finds it by searching. Ask for the shape:

chief recall --shape unresolved_thread   # who is awaiting your reply
chief recall --shape commitment          # what you said you would do

Records, not verdicts

Two claims that disagree are usually not a conflict — they are the same thing at two times. "Burn is $73K" (June) and "burn is $78K" (July) are both true; the later one describes now. Chief does not adjudicate that for you at write time. It hands you both records, with dates and asserters, and you resolve at read time: recency first, then authority (something you said outranks something extracted from a service email), then the evidence itself.

If none of those settles it, that is a real answer. Two records that disagree, with their sources, beats a confident single answer that picked one.

A zero result is not proof of absence

Every answer states which layer it searched and what that layer covers. Read the footer before concluding something did not happen — the window may simply not be indexed yet. chief status shows coverage per source.

Reference

• install.md — install the CLI and sign in
• plugin.md — install the CLI-first companion skill for Codex and Claude Code
• cli.md — every command
• mcp.md — the agent protocol surface