Activity log
Updated 3 September 2026
What changed, and what is still broken. We keep the open problems here as well as the fixes, because a log that only lists wins is no use to anyone testing against it.
The privacy policy now spells out how your data is protectedImproved
A new "How we protect your data" section describes the actual measures in place: encryption in transit and at rest, a separate encrypted store for connected-account tokens that only a process acting for the owning account can read, per-account isolation at the database, and restricted internal access. "Retention and deletion" now gives a concrete window and states that disconnecting a Google or Microsoft account deletes its stored tokens. The AI section states plainly that data from Google APIs, including Google Workspace APIs such as Calendar, is never used to train non-personalised models.
To test: Open the privacy policy from the site footer and read the "How we protect your data" and "Retention and deletion" sections.
Homepage draft: clearer cards, a mobile-beta section, and the release-checklist link retiredImproved
On the in-progress homepage: the Remember / Understand / Act cards drop the low-contrast emoji for solid icon badges and lose the tiny question labels; the relationship-continuity pair matches height and the meeting card follows the lifecycle-card layout, with the portrait re-cropped to sit properly; the "Wherever people meet" cards get frosted-glass "Next" chips that now link to sign-in; the privacy section swaps its padlock emoji for a real icon and loses most of its empty space; and there is a new section inviting people to join the iOS / Android beta. Separately, the "Consumer release checklist" button has been removed from the top of this page (the checklist itself still lives at /activity/roadmap). None of the homepage changes are on the live site yet.
The consumer release checklist is now visible and accountableNew
The Activity Log now links to one ordered checklist covering new, existing, guest, offline and cross-device users across iOS, Android and the current consumer web. Items are checked only after the required deployment and device or browser evidence exists. Capture recording remains mobile-only; web verifies the saved encounter rather than recreating the recorder.
To test: Open Consumer release checklist at the top of this page and confirm every phase shows its completion state and overall progress.
AI, capture relationships and group follow-ups close their loose endsFixed
A fresh AI chat no longer turns a stray one-word voice result into its opening question, recent chats can be removed across signed-in devices, and relationship history keeps its last good state while refreshing. Captured names now match saved connections despite spacing or punctuation differences. Email and calendar follow-ups can reach every attendee together on mobile and web only when every recipient has a usable email, so nobody is silently left out. Capture itself remains a mobile-only experience; web receives the resulting meetings and follow-ups, not the recording flow.
To test: On mobile, start a fresh voice chat, delete a recent chat by holding it, and open a connection while history loads. From a meeting with two emailed attendees, choose Everyone for an email or calendar follow-up on mobile and web; confirm both recipients are present. Confirm web has no Capture entry point.
AI outcome learning is measurable without uploading private memoriesNew
The assistant now records privacy-minimised outcome signals for successful, recovered and failed interactions, including whether a widget appeared and whether a verified action completed. A client-side safeguard reconstructs a promised follow-up widget even when a stale or incomplete server response omits it. The new How AI improves link explains the boundary and shows the signed-in user's outcome counts. Prompts, replies, names, emails, transcripts, notes, recordings and retrieved memories are explicitly excluded from the shared learning ledger.
To test: Trigger a recovered follow-up widget and a verified approval, then open How AI improves beneath the conversation. Confirm the recovered and successful counts update after the learning migration is deployed.
AI replies can include useful tap-to-continue optionsImproved
Assistant responses can now pair readable prose with up to three concise contextual reply chips. Tapping one continues the same conversation without retyping. Pending follow-up proposals always offer Review draft and Approve through the real guarded flow, while ambiguous people keep their existing identity choices. Generic or unsupported options are explicitly excluded to avoid clutter.
To test: Ask a question with an obvious next step and confirm relevant reply chips appear. In a follow-up proposal, tap Review draft and Approve and confirm both use the existing widget and verified save path.
Pending AI cards survive long clarification conversationsFixed
The assistant was saying a draft was shown below while omitting the structured card because the original request had fallen outside a second, shorter 20-turn recovery window. Pending-action recovery now uses the full bounded saved conversation window, so the original person, task and due date continue producing the real widget until the action is approved or cancelled.
To test: In the existing Martin conversation, ask ‘show me’ again and confirm the approval widget appears rather than prose alone. Approve it and verify the exact action appears in Follow-ups.
AI approval cards have a softer visual treatmentImproved
The follow-up review card now uses a restrained translucent green outline and a diffused green shadow instead of a hard saturated stroke. Its content, spacing and confirmation behavior are unchanged.
To test: Open a follow-up proposal and confirm the card edge reads softly against the assistant background, with a subtle blurred glow rather than a strong green border.
Follow-up review widgets no longer depend on model formattingFixed
When a conversation contains the person, task and timing for a follow-up, the server now reconstructs the structured proposal if the model returns only explanatory prose. This guarantees the visible review-and-approve widget for the Martin-style missing-email flow. On iOS, the assistant dock also removes duplicated safe-area spacing only while the software keyboard is visible; its normal inactive position is unchanged.
To test: Continue the Martin conversation with ‘show me’ and confirm the review widget contains Martin, ‘Send Martin my project proposal’, and the resolved due date. Open the keyboard and confirm the dock sits close to it, then dismiss the keyboard and confirm the original resting position returns.
Conversational approval now creates and verifies the follow-upFixed
A pending AI action no longer disappears when the user says ‘show me’ or confirms with ‘yes’, ‘yea’, ‘approve’, or ‘do it’. Show keeps the review card visible; confirmation executes the same guarded save as the Approve button. The client then reloads the real follow-up endpoint and only reports success if the new open action is present. Generic no-match endings were replaced with intent-specific recovery routes that ask for the smallest useful detail and continue toward a solution.
To test: Prepare a follow-up for a new person without email, say ‘show me’, then ‘yea’. Confirm the card remains visible, the save is verified, and the new action appears in Follow-ups before success is shown.
Something behaving oddly? Tell us what you did and what happened, and it will appear here.