Every agent you open starts from nothing. It cannot know what you decided last week, what you already tried, or who is waiting on you — so you explain it again, every session, in every chat. Chief is the memory and tools that they share: your mail, calendar, documents, and meetings. All as durable memory that your agents can both query and write back to.
Real session — one name replaced for privacy.
❯I'm a fresh agent with no context. What did Adil decide about the marketing hero, and what should I not touch?
…looked for local memory and a repository, and found neither
chief recall { "query": "marketing site hero section decision", "limit": 10 }
10 claims · layer: claim graph
chief show { "id": "c/a34dngk425hz" }
You prefer options #1, #3, and #4 as the best candidates to show on the Hero. valid from 2026-08-29
chief show { "id": "c/qeuodkg4yrff" }
You are undecided about how many items should be shown on the Hero. valid from 2026-08-29
Adil picked hero options #1, #3, and #4. If more than one shows, those are the three. For #3, use Rio Vantage — he said he can get approval. Don't touch: · The item count — explicitly undecided (c/qeuodkg4yrff). Don't pick a number for him. · #3's Rio Vantage usage — approval isn't in hand yet, so don't publish it. No local memory files or repo here; this comes from Chief.
A fresh agent in an empty directory — no repository, no prior session, no context window to inherit.
One project account, composed from the same evidence for everyone doing the work.
Chief ships a CLI and speaks the Model Context Protocol. Both read and write the same memory, under the same account, as the web app.
The installer and the MCP server are public and shipped — the commands below run on any Mac or Linux machine right now, and chief mcp will start and list its tools before you have signed in to anything. What is still gated is the account: Chief is in early access and accounts are issued by invitation, so chief login needs one. Ask for an invitation and the rest of this works the day it arrives.
curl -fsSL https://runchief.ai/install.sh | sh
chief login
chief recall "what did we decide about pricing"A prebuilt binary for macOS and Linux, checksum-verified before it replaces anything. No clone, no toolchain, no build. chief login opens your browser and takes a one-time code back on a loopback port — there is no token to paste and no config file to hand-edit.
claude mcp add chief -- chief mcp{
"mcpServers": {
"chief": { "command": "chief", "args": ["mcp"] }
}
}The second form is for Codex, Cursor, Zed, or anything else that reads an MCP config file. The server uses the credential chief login already left on the machine, so there is no API key and nothing to configure per harness.
Ten tools over MCP — recall, who, show, remember, source_search, status, mail_draft, mail_reply, mail_archive and mail_send — which delivers immediately and cannot be undone, and says so in its own description where the agent will read it. They are the CLI's own commands and return the CLI's own output, because that output was designed for an agent to read: bounded, with stable ids that are valid arguments to the next call and errors carrying a code the harness can branch on. The initialize response carries the instructions an agent needs to work the surface out for itself.
Chief preserves source material, compresses it into a current project narrative, and keeps every consequential statement traceable to evidence the model actually opened.
Agent sessions land in the memory beside the human context connectors ingest around them, and both become evidence for the same account.
The model decides what matters, what changed, and what to omit, producing a coherent account instead of a wall of metrics.
Act on one named finding—draft an update, create a task, or start project work—with the finding and evidence attached.
Chief tells you what the evidence established, where it came from, who asserted it, and when. That is a different product from a search index with a chat box in front of it, and the difference shows up on the days the answer matters.
The passages nearest your question in embedding space. Similarity is a guess about relevance, and a confident one.
Claims — short statements of what happened — each carrying the evidence it came from. Every consequential statement is traceable to evidence the model actually opened, not to what a ranker would surface now.
A citation is attached afterwards, if at all. Whether a human asserted it or a billing robot did is not part of the record.
Every claim records who asserted it and when. Something you said outranks something extracted from a service email, and the answer shows you which one you are reading.
Two conflicting passages become one confident answer. You never learn there were two.
Records that disagree are usually the same thing at two times. Chief keeps both, with their dates and asserters, and resolves at read time: recency, then authority, then the evidence itself. When none of those settles it, it says so.
An empty result and an unindexed corpus look identical.
Every answer states which layer it searched and what that layer covers, and chief status reports coverage per source. A zero result over an uncovered window means something different from a zero result over a covered one.
One index, asked everything. Ask what is still outstanding and it answers from the past.
Two layers, deliberately. The claim graph is past-tense — what the evidence established. Live sources are the provider's own state, queried directly. Asking the wrong one is how you get a confidently wrong answer.
Read-only. Whatever you worked out in the chat is gone when the window closes.
Conclusions go back in, attributed. An agent records its own with chief remember; your verbatim words are recorded as yours, and carry more weight the next time records disagree.
Chief reads a source only where you have already granted it access, one source at a time, and it reads nothing you have not connected. It does not add an invitation flow or a second permissions model. Connected today: Gmail, Google Calendar, and Google Drive and Docs over OAuth; Obsidian through the Mac app or the Linux desktop runtime; Granola and iMessage through the Mac app.
A memory your agents query is only worth having if you can check it and live with where it keeps things. This is the guarantee behind the claim above, answered on the page rather than only in a policy.
The memory itself — your claims, their evidence, and the source material behind them — sits in a database on hardware Chief Labs runs, not on a hosted database service. Local sources are read on your own machine by the desktop app, which you install and can quit: it opens your Obsidian vault, your Granola transcripts and your iMessage history locally, and then sends their contents to that same database, like every other source. Reading is local, and so is the heaviest thinking. Storage is not. Nothing about a local source stays only on your machine, and you should decide about Chief knowing that.
Increasingly, yours. The heaviest pass — the one that reads a thread and writes the claims from it — now runs on your own machine through your own model subscription, so that content reaches a model through your account rather than through a rack somebody else owns. Be clear about the size of that: it is one pass, not all of them. Lighter passes, embedding and reranking still go to third-party providers — today Anthropic, OpenAI, Google, and an open-weights model served from a US host — bounded by our contracts with them. Chief trains no models on your data, and data obtained through Google Workspace APIs is never used to develop or train generalised models: the Limited Use requirements of the Google API Services User Data Policy. That provider list is the thing here most likely to change, and the policy is where it stays current.
Reading meaning out of someone else's email is inference, and inference is sometimes wrong. Chief does not hide that. Every claim shows the evidence it came from and who asserted it, and you can open the source and see for yourself — which is the point of keeping the source at all.
One command records the correction as your own words and retires the claim it replaces, in the same transaction. Nothing needs an operator and nothing is edited in place: the old record stays true about the day it was made, and the new one wins from then on.
chief remember "..." --from-user --supersedes c/3kf9q2chief logout revokes the session on the server before it clears the credential on your machine, and a Google grant can be withdrawn from your Google Account at any time, which stops Chief reading anything new. Settings will export everything Chief holds about you in one request — not a summary, the rows — and will delete your account and its data outright. The table set for both is read from the database catalogue rather than a hand-written list, so a table added next month is covered without anyone remembering, and the deletion runs in one transaction that recounts every table before it commits.
Early access
Request access to Chief's evidence-backed project accounts.