Changelog

What changed recently.

Short product notes on the loops, the ledger, the runtime, and the safety work behind early access.

  1. Explicit feedback gets its own trail

    Alpha feedback now becomes reviewable product evidence without turning into personal memory.

    • Added explicit Telegram feedback capture for feedback:, bug:, that was wrong, and this helped.
    • Stored alpha feedback as product signal only, with nearby context and operation links when available.

  2. Telegram onboarding and help arrive

    Telegram now has a clearer front door, help surface, and closed-alpha boundary.

    • Added deterministic /start, help, domain help, and privacy replies for Telegram.
    • Closed-alpha users who are not allowlisted are stopped before the assistant runtime is called.

  3. Learning recall digests take shape

    Saved notes can now resurface as reviewable recall loops instead of one-off summaries.

    • Added learning digest commands, topic digests, feedback capture, and digest settings.
    • Kept scheduled digest delivery opt-in and gated from the Worker runtime.

  4. Conversation replay becomes a guardrail

    Conversation regressions are easier to catch before they reach the real assistant.

    • Added a replay harness for redacted Telegram fixtures with real route planning.
    • Assertions now cover route choice, reply shape, conversation state, policy, and operation evidence.

  5. Alpha analytics report lands in the owner console

    The private dashboard can now show whether the alpha is becoming useful and affordable.

    • Added DAU/WAU/MAU, retention, activation, per-user cost, top routes, and failure summaries.
    • Added Feedback Watch and General Fallbacks so alpha issues and routing drift are easier to inspect.

  6. Semantic routing moves into guarded rollout

    Ambiguous and multilingual requests have a safer path into the right domain.

    • Moved domain-owned route catalogs behind a guarded semantic classifier.
    • Kept most catalogs in shadow mode while Routine calendar lookup becomes the first active route.

  7. Production runtime and action history groundwork

    The assistant runtime is becoming easier to observe and safer to operate.

    • Moved the Mastra runtime toward a production Cloud Run shape.
    • Added safer operation logging, retry visibility, and dashboard review paths.

  8. Private dashboard hub comes together

    A private owner hub now gives the assistant a place for memory, permissions, and review.

    • Built the private dashboard hub for habits, learning notes, recipes, finance, and policies.
    • Added settings surfaces for controlled automation instead of invisible background behavior.

  9. Routines and follow-up get more practical

    Daily routines are moving from simple logging toward useful follow-through.

    • Started calendar-aware morning brief work for practical daily planning.

  10. Public web research joins the assistant

    The assistant can answer current public-web questions while respecting private context boundaries.

    • Added web research routing for current, source-grounded questions.
    • Kept private assistant data separate from public web lookup behavior.
Follow along

Get the useful updates.

One quiet list. Changes like the ones above, when they ship.

alpha notice: i store your conversations and usage analytics to run and personalize your loops, in the product database, never a third-party analytics tool.