What's changed.

For the last week the AI’s three takes (Witness, Mirror, Reframe) lived behind a picker interface: you read one, switched tabs to compare, picked the one that fit, and the other two were discarded. The picker was a clever solution to the “which AI voice do you want?” problem — but on reflection it was still asking you to choose between AI voices, which is the gimmick the product set out to avoid.

New shape: when you save an entry, the digest renders all the takes together as labeled sections. Witness sits next to Reframe. Mirror joins when there’s a real pattern from your past entries to point at. There’s no pick action, no “remember my choice” button, no decision to make. Holding several readings of the same moment in tension is the reflective work, not a UI to navigate.

Under each take, a quiet “Ask about this take →” link opens a small inline conversation about that piece of your writing. Type a question, send, and the AI replies right under the section. The conversation persists alongside the entry — come back weeks later and you’ll see the exploration you ran the first time around. This is the part of the product where genuine learning happens: you pull on a phrase that catches you, dig into it, and the trail stays.

The Settings toggle to control the picker is gone too. The chat surface, separately, was simplified to a normal back-and-forth (no in-chat picker either) — chat is conversation, the multi-take view lives on the digest.

The AI that responds to your journal entries was retuned. Instead of three takes that varied mainly in warmth (Warmer / More direct / Reframe), the picker now offers three takes that do genuinely different things:

• Witness— names what’s underneath in one sentence. No question, no advice, no next-step. The act of being seen is the whole interaction.
• Mirror— points at a pattern across your past entries (only shown when there’s a real connection to point to; if not, you’ll see two takes instead of three).
• Reframe— surfaces the interpretive leap you made and offers a kinder reading of the same situation.

Behind the scenes, the primary text-generation provider was swapped to a model that produces noticeably warmer, more literary output for journaling. The new provider is hosted in mainland China; the failover providers are US-based. If you’d rather your data not leave the US/EU and you’re on a paid plan, write to us and we’ll pin you to the failover chain. The full posture is on the privacy page.

Each entry’s Summary / Insight / Next-step is now drafted as three takes; you pick the one that fits and over time the picker quietly leans toward your style. Previously this only existed in chat.

The Companion / Scholar / Minimalist / Coach picker was retired. It was a gimmick — in practice the four personalities were variations on the same instinct to paraphrase the surface and ask a soft question. The new voice was designed around two demands: real empathy (naming what’s underneath, not the surface words) and concrete direction (a small grounded step, never a wellness-app cliché).

The AI now also reads from your own past entries to ground its responses in who you actually are. We never train on your content; the entries are only included in-context for that one call. Entries you mark private are excluded from this entirely.

When a sign-in lands from a device or location we haven’t seen before, you’ll get a one-time code by email before the session is granted. If you prefer the extra step every time regardless of device, there’s a toggle in Settings → Sign-in security. You can also see and revoke individual known devices from there.

Entries can be appended to. When you save something that resembles an existing page of yours, you’ll get the option to continue that page instead of starting a new one — so a single thread can grow over weeks without cluttering your timeline. The original digest stays preserved on the parent page even after re-digesting.

A short scrollable tour at /tour walks through how the writing surface, the picker, and the continuation flow fit together.

Deleting an entry now soft-deletes it for seven days. You can restore it from the Recycle page during that window; after seven days it’s permanently removed. There is no user-facing hard-delete — the grace period is the policy.

Every entry now has a Private switch in its Saved view. When it’s on, that entry is exempt from all AI features — no digest, no chat reference, no inclusion in weekly summaries, no embedding. The exemption is enforced at the API boundary, not just the UI.