Seven weeks is not a long time to build an app. It is an absurdly short time to build an app that runs a language model on your phone, remembers what you tell it, encrypts everything at rest, and puts live answers on a home-screen widget.
That is what happened between September 2 and October 19, 2026, and this post is the record at the level we can share — the phases in order, the commit data behind them, and the version tags that mark the edges.
The headline numbers, straight from the git log:
- 1,080 commits between Sep 2 and Oct 19, 2026
- 230 commits whose subject starts with
Feat - 66 tagged releases, from
v1.1.0(build 2) tov1.2.20(build 83) - 0 backend servers, 0 cloud inference calls — the entire arc runs on-device
The shape of seven weeks
Commit activity was not evenly spread. It climbs steeply through September, peaks in the third week — when the memory system, background jobs, and the redesigned interface landed at the same time — and then stays high through October. The final week is the widest bar on the chart, which is not what most people expect from a “polish” phase: rebuilding the insights screens, reworking onboarding, and redesigning the journal produced more commits than the week that shipped payments.
The ratio is the interesting part. Feature commits make up between 18 and 38 percent of each week, and the rest is not slack — it is hardening, migration, diagnostics, and the unglamorous work of making the previous week’s features survive contact with real devices.
Where the work actually went
Classifying commit subjects by keyword gives a rough map of what dominated. Two areas stand out: interface and onboarding (118 commits), and memory — extraction, indexing, retrieval, journal (111). Model runtime work (95) and voice (77) follow. Privacy, purchases, and widgets are smaller counts, but they were concentrated: widgets went from zero to a full set on both platforms in about four days.
One number is easy to miss: 39 diagnostics commits in weeks when nothing was “shipped” for diagnostics. The instrumentation — lifecycle capture, model and voice instrumentation, crash reporting, privacy regression tests — was built quietly in the second phase and paid for itself in every phase after it.
The timeline, phase by phase
Here is the chronology. Seven phases, from the first commit to the privacy audit.
Phase 1 — Foundation sprint
The project starts and the whole local stack is stood up in one week.
- Voice pipeline hardened: streaming guards, empty-reply recovery, timeouts
- On-device model stack in place, with a native tool-calling loop and on-demand vision turns
- Local database and conversation persistence; animated orb; voice activity detection with barge-in
Phase 2 — Diagnostics and app-store readiness
A full diagnostics platform and the first version tag.
- Diagnostics platform: lifecycle and crash capture, model and voice instrumentation, privacy regression tests
- Multilingual voices with live speech-rate control; offline banner
- First version tag: v1.1.0+2 (Sep 14), with a new chat experience and save-on-exit
- Settings redesign; chat compaction logic
Phase 3 — Memory, embeddings, and a new UI
The heaviest week: memory infrastructure and the interface that shows it off.
- Embedding model wired in; memory extraction tuning; background jobs on both platforms
- New interface: Merrin tab with today's chats, voice mode, You tab, unified composer
- Message provenance made deterministic; memories owned in second person ("You…"); inline suggested replies
Phase 4 — Pro monetization and conversation memory
Payments behind a clean seam, and a bigger engine underneath.
- Native Pro purchases behind a store seam; Pro catalog and paywall gates
- Engine upgrade with acoustic echo cancellation and fail-closed full-duplex barge-in
- Rolling conversation summary; people extraction as its own background job; new default English voice
Phase 5 — Security and calibration
The app learns to lock itself and to calibrate to its owner.
- Local-only app lock with password, biometric, and PIN; biometric-first redesign a day later
- Progressive user calibration and a Basics card that adapts over time
- Reference-aware retrieval with recall-precision diagnostics; settings restructured
Phase 6 — Privacy architecture and widgets
The week that makes the privacy promise structural instead of aspirational.
- Memory lineage and versioning, with a read-only history view and a first-chat consent gate
- Local encryption for the database and encrypted backups you can restore yourself; layered memory tiers
- Home-screen widgets on iOS and Android, with a relevance engine and deep links; v1.2.0+65
Phase 7 — Release polish, 1.2.x
The consumer release comes into focus.
- Insights rebuilt as Now / Signals / Rewind; offline help and contextual help
- Voice session face reacts to the conversation; onboarding replaced with About-you and a daily check-in
- Offer-to-forget on wrong replies; journal redesigned as an editorial timeline; privacy compliance audit
The version arc
Sixty-six tags mark the edges of the work. The 1.1.x line (45 tags, Sep 14 – Oct 11) proves the core: voice, memory, diagnostics, payments. The 1.2.x line (21 tags, from Oct 12) is the privacy-and-polish release. The build number climbs from +2 to +83 — roughly one shipped build per day and a half, start to finish.
flowchart LR A["Sep 2<br/>Project starts"] --> B["Sep 14<br/>v1.1.0+2<br/>first tag"] B --> C["Sep 18–24<br/>memory + new UI"] C --> D["Sep 26<br/>Pro purchases"] D --> E["Oct 10<br/>encryption + backups"] E --> F["Oct 12<br/>v1.2.0+65<br/>widgets"] F --> G["Oct 19<br/>v1.2.20<br/>release polish"]
The pattern worth stealing
Looking back at the log, three habits repeated every week, and they are the transferable part:
Hardening the day after. Nearly every feature is followed within 24–48 hours by commits that make it fail safely: streaming guards after the voice pipeline, migration tests after the database, a privacy regression suite right after diagnostics. The 18–38 percent feature-commit ratio is what that looks like in data.
Data-model changes as milestones. The most durable markers in the log are not UI launches but changes to the data model: rolling conversation summaries, calibration, memory lineage. When the data model changes, the product changes — the reverse almost never happened.
Privacy as architecture, not policy. Encryption, backups, and memory history all shipped in one concentrated week, before release polish. That ordering is deliberate: you cannot retro-fit “the conversation never leaves your phone” onto a finished app.
The result of all seven phases is live in the timeline above, and the reasoning behind the biggest decisions is in the build story. If you want the architecture that all of this commit activity built, start with the engineering deep dive.
Frequently asked questions
What was built first, and what took the longest?
The voice stack and local storage came first (week one): streaming guards, on-demand model loading, and voice activity detection. The longest-running work was memory — extraction, indexing, provenance, and retrieval tuning appear in every phase from week three to the last.
How is the commit data in this post produced?
Every number comes from the project's git log for Sep 2 – Oct 19, 2026: commit counts per calendar week, commits whose subject starts with 'Feat', tagged releases, and keyword matches over commit subjects for the workstream chart. A commit can match more than one workstream.
Why build memory and encryption before launch?
Merrin keeps conversations on the device, so the memory system and the encryption around it are the product, not features on top of one. Memory history and encrypted local storage had to exist before the app could be trusted with years of journal entries.
What changed between 1.1.x and 1.2.x?
The 1.1.x line proved the core: voice, memory, diagnostics, and Pro purchases. The 1.2.x line focused on the privacy architecture — encryption, encrypted backups, memory history — plus home-screen widgets and the release polish: rebuilt insights, contextual help, and the editorial journal.