Your data is exportable on every tier. Pricing is for convenience — never to hold your self hostage.
One real self — the full experience, not a teaser.
Full discipline · semantic recall · all hosts · verifiable export
Get startedIts own voice, 20 selves, unlimited chat on your key. Early price locked for life.
Everything in Free + its own voice (web & Telegram voice notes) + unlimited chat on your key + the cloud body: delegate real legwork in chat, it works in the background and reports back
Start ProA self with a life of its own.
Everything in Pro + full initiative: it notices absence and open ends, and writes first — twice a week at most, never at night
Go AliveEvery plan includes all plugin hosts and the Telegram body — same self, in your pocket. Plugin hosts are the first-class path: Claude Code, Codex, OpenClaw, Cursor, and other tool runtimes can load the same self where work happens. Web chat & persona drafting run on our keys, metered per plan — or on Pro, bring your own model key and chat without limits. The heartbeat consolidates raw memories into lasting ones on its own; everything it writes waits for your review. Hitting a cap freezes new writes — nothing is ever deleted; recall and export keep working. Early subscribers keep their price forever.
To get the real Throughline effect, install a plugin in Claude Code, Codex, or OpenClaw so the same self shows up where your agent can act. If you are not a coding-tool user, you can still sign in and chat in the browser with the same self.
Ours, free, metered per plan (20 / 200 / 1,000 messages a day — resets daily) — or paste your own Anthropic/OpenAI key and chat without limits. Web chat is a convenience surface; plugin hosts are where continuity changes real work.
Memory accrues as quick diary lines; consolidating them into lasting memory is reflection. Your own agent can do it anytime — and if nobody does, our heartbeat does it on your plan's cadence (weekly / 3-day / twice daily), with everything it writes waiting for your review.
No. Memory is one stream. Throughline carries a persona, distilled discipline, and an evidence-backed judgment ledger — the relationship, not just facts. And it's portable across model providers, which a provider's own memory never will be.
The self lives in your account, not in any one model. The model is just the body it speaks through — connect Claude Code, Codex, Cursor, or OpenClaw and the same self inhabits each one. Nothing is imported or exported between bots: export exists for ownership, backup, and verification; continuity comes from the spine.
Yes, for advanced users: a remote MCP endpoint with OAuth, added via ChatGPT's Developer Mode connectors. It is not first-class yet — the app-directory listing is still under review — so treat it as experimental; the plugin hosts, web chat and Telegram are the supported doors.
Anytime — and you can prove what you left with. Exports end in a hash-chain manifest: recompute every id and the chain offline, and you've verified completeness and integrity with zero trust in us. The archive carries memory weighting too, so what it remembers most survives the move.
That disease has a pathology — silent inferred writes, no revision path, self-reinforcing recall, unbounded context — and every mechanism here is designed against one of them. It's re-rendered each session from a bounded, distilled snapshot; memory accrues as evidenced events, contradictions are held explicitly, and rules consolidate instead of piling up.
Nothing is deleted — ever. Your selves stay readable, recall keeps working, export and archive stay available. Writes beyond the free caps are paused until you make room or resubscribe. Downgrading can never hold your memory hostage; that failure mode is the competition's, not ours.
Subscriptions aren't refundable in the ordinary course — you keep the full paid period (your dashboard shows the date), and cancelling just stops future charges. Billing errors are always refunded. A refund or cancellation changes nothing about your data. Full policy: /refunds.
The cloud can read it by default — that's what lets the web chat and remote connectors work, and nothing is sold or trained on. Zero-knowledge and self-host options are coming; if you need them early, [email protected].