Release notes

What's new in Everpaper, newest first. During the private beta each build arrives as a download link and installs over the previous one; automatic updates begin with the first public release. See Updates & Support.

2.0.0-b98 — beta (latest)LatestBeta#

Gemma 4 runs inside Everpaper on macOS 27.

• Gemma 4 inside Everpaper Native. On macOS 27 Macs with 16 GB of memory or more, Everpaper Native now runs Gemma 4 E4B, the model that wrote the best evening entries in our tests. It runs inside Everpaper, so you don't need LM Studio, Ollama or any other server.
• Clear sizes, careful switch. The download is about 5.9 GB, and the installed model takes about 8.5 GB (some tools show 7.9 GB). Setup checks there's room for both before it starts, and switches over only after Gemma has answered a first request. If the connection drops for a moment, the download retries and carries on where it stopped.
• Setup picks it for you. On these Macs, guided setup now chooses Everpaper Native with Gemma. A running LM Studio or Ollama with a chat model is still used, and an AI setup you already have is kept. macOS 27 Macs with less than 16 GB keep Qwen3 4B; earlier macOS versions work as in b97.
• Already on Everpaper Native with Qwen? A one-time card offers the switch to Gemma and shows its sizes. Qwen keeps answering while Gemma downloads, and its files (about 2.3 GB) stay until you remove them.
• Native models by name. In Settings → AI & summaries, Everpaper Native lists the models on this Mac by name. You can set one up there, or remove one you no longer use.
• A steadier engine. Prompts are formatted exactly as Gemma expects. An answer cut off at its length limit no longer garbles the next one, and output made only of internal marker tokens counts as a failure, so it never reaches your journal. Structured answers on long prompts use about 1 GB of memory instead of about 5.8 GB. One engine now stays loaded, closes after ten idle minutes, and restarts between requests if its memory grows too much.
• Running notes are protected. If a rewrite of the day's running notes drops more than half of them, Everpaper keeps the earlier notes and adds the new ones after.
• One-time setup cards are back. Recent betas carried a name suffix that silently held back Everpaper's one-time setup cards. This build doesn't, so you may now see the speaker-labels card b97 meant to show.
• Known limits: Gemma in Everpaper Native hasn't been tried on a real 16 GB Mac yet. The first request after the engine starts can take around half a minute, and after a restart the next request pauses for a few seconds. Evening entries still condense busy days and drop some meeting details. When several people share your Mac's microphone, the live transcript shows them all as "Me" until the meeting ends; the saved transcript separates them.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 15 or later, Apple Silicon. Gemma 4 in Everpaper Native needs macOS 27 and 16 GB of memory or more.

2.0.0-b97-journal — betaBeta#

Your daily journal reads like a diary again, meetings get a summary, and Gemma 4 joins the local models.

• Everpaper now needs macOS 15 (Sequoia) or later.
• An evening entry in your voice. When you finish your day, Everpaper writes one warm, first-person entry from the day's notes and meetings, including each meeting's decisions, dates and who does what. A day cut short by a quit or crash is written up the next time Everpaper opens.
• Entries can outlive their screenshots. With Also delete their text & recall data off (Settings → Daily history), finished entries are kept when old screenshots are cleaned up. The switch is on for new installs, so there a day's entry is removed with its screenshots after your retention period.
• A summary for meetings. Each meeting of more than a few minutes gets a summary once it ends, shown on its page and reused in the evening entry. Longer meetings get longer summaries. This needs an AI option set up, and a cloud provider needs your consent.
• Everyday meetings: Parakeet for words, Nemotron for speakers, live and in the saved transcript. If Nemotron can't run, Everpaper falls back to its previous speaker labels.
• Clearer live speaker labels. A new caption line starts when the speaker changes; in our tests, words shown under the wrong speaker fell from about 31% to about 10%. When several people speak into your Mac's own microphone, the saved transcript separates them, and you stay "Me" when you're the only one talking.
• Ready before your first meeting. Today warns you if transcription isn't set up, with one Set up button and the download size. If you skip it, the first meeting downloads the model with progress, and the audio is always recorded. A missing model no longer stops a meeting, discarded recordings stay discarded, a microphone that goes silent is restarted, and merging speakers keeps the right names.
• Gemma 4 E4B for Macs with 24 GB or more (about 6.9 GB). Setup reuses an LM Studio or Ollama you already run instead of installing a second engine. Thinking is turned off, every downloaded file is checked, and existing setups are kept.
• A simpler Speech page: Everyday or Detailed, your words, and an Advanced section listing the models on this Mac with their sizes and a Remove button.
• Known limits: evening entries still condense busy days. When several people share your Mac's microphone, the live transcript shows everyone as "Me" until the meeting ends.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 15 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b96-nemotron — betaBeta#

Nemotron becomes the default for words and speaker labels on new setups.

• New setups use Nemotron for both. New installations transcribe and label speakers with Nemotron. Parakeet is still in Settings → Speech, and an engine you already picked is kept.
• Model setup checks every file. Setup verifies the files it downloads and prepares both the after-meeting and the live speaker models. Once prepared, it works offline.
• Short live snippets keep their last word. A word at the edge of a short live chunk is no longer dropped.
• No phantom extra speaker. A voice that already has a label no longer gets a second one from a later pass.
• Microphone changes recover better. Everpaper copes better when you switch to an audio device that uses a different format.
• Includes b95's click fix. The old MOSS introduction card is gone.
• To use Nemotron for both on an existing setup, open Settings → Speech, choose Nemotron under Transcription and press Set up. In a small test Parakeet got more words right, so this beta doesn't claim Nemotron is more accurate.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b95-event — betaBeta#

Fixes for a crash and a hang linked to recording your clicks.

• Click details are read safely. When Computer History notes what you clicked, it now reads that detail the way macOS expects. Reading it in the background could trip an error inside Everpaper's own windows, and was linked to a crash and a hang in b94.
• Stopping the workday twice is safe. Stop requests are handled one at a time.
• A failed save keeps your review. If saving the day fails, the review stays open so you can try again.
• Click details are still recorded, and capture modes and saved data are unchanged. The crash isn't proven fixed yet; if clicking ever feels slow, report the app you were in and roughly when.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b94-event — betaBeta#

A fix for a crash when you start a meeting while your workday is being recorded.

• Record now during a workday is steadier. While your workday is being recorded, pressing Record now switches Today straight to a "workday paused" status, instead of keeping the live control running through the change. This targets a crash seen on b93.
• Mark Moment, Stop & Review and Discard (with confirmation) stay available. A workday you paused yourself still offers Resume and stays paused after the meeting.
• If you stop the workday while the meeting continues, Today shows the day's writing and review status.
• The original crash didn't reproduce on our test Mac, so treat this build as a test of the fix. After a crash, Everpaper keeps what was recorded and doesn't resume recording on its own.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b93-event — betaBeta#

Choose Everyday or Detailed meetings, and try live speaker labels.

• Everyday or Detailed meetings. Settings → Speech now opens with two choices, explained by purpose and speed. The exact models are listed under "What runs on my Mac".
• Live speaker labels (experimental, opt-in). For Everyday meetings, you can turn on live speaker labels after setting up a local Nemotron model of about 190 MB. People on the other end get labels during the call, which may change as it goes. When the model is unsure, captions keep Me / Others. The words still come from the usual live recognizer.
• An optional after-meeting speaker check. A separate opt-in pass checks the speaker labels once the meeting ends.
• Both are off by default, a recording never starts a model download, and saved meetings stay compatible.
• Smoother sidebar and Settings. The selection highlight uses one animation instead of two, and Settings no longer rebuilds itself on every section change.
• When reporting a wrong or flickering label, include the meeting's length, how many people spoke, and whether it was Everyday or Detailed.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b92-event — betaBeta#

The journal keeps what it has when a rewrite goes wrong.

• A failed rewrite no longer shrinks the day. If the optional final rewrite fails or comes back much shorter, Everpaper keeps the day as its sources wrote it.
• Meetings keep their own paragraphs. Updates from your screen no longer blend into meeting paragraphs; each meeting stays separate and labeled.
• Automatic Episodes and Chapters are paused. No new ones are made for now; the ones you have stay.
• History shows spans you can open. Recorded activities are grouped into stretches of time, each opening its sources.
• Memory chat can read sources directly. When normal search comes up short, Ask reads a limited set of matching sources itself and skips anything no longer visible.
• The journal isn't yet as good as it used to be, and search finds the right source less often than it should.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b91-event — betaBeta#

The journal keeps one running account of your day.

• One account that grows through the day. The journal no longer writes separate summaries that each lose what came before. It carries a single running account and adds to it at each update.
• Meetings join the day in order. Finished meetings enter the account at the time they happened, with more of what was actually said, before the evening entry is written.
• Better meeting excerpts. When a transcript has to be shortened, Everpaper keeps opening questions, whole sentences and the "but…" that qualifies a point.
• Changes ripple correctly. When a source changes or is deleted, the summaries built on it are refreshed together, and your own notes and titles are kept.
• Smoother Memory and sidebar. Switching sections stutters less.
• Existing journals aren't rewritten. The writing still depends on your local model and isn't yet as good as it used to be.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b90-event — betaBeta#

Settings, redesigned with plain names and native Mac controls.

• Settings pages with plain names. Capture is now Daily history, Transcription & Speech is Speech, Intelligence is AI & summaries, Privacy & Data is Privacy, Shortcuts & System is General, and This Copy is About. Controls are standard Mac switches, menus and buttons, grouped into sections, and Privacy is simpler to move around.
• Settings opens from the sidebar. Settings slides in from the main window's sidebar, and Back to Everpaper sits in its own sidebar.
• One selection style. The highlight in both sidebars glides to what you pick; with Reduce Motion it moves at once.
• System lettering for labels. Navigation, Today's labels and shared controls use the Mac's system font. Journals, transcripts and answers keep their reading typeface.
• Sources restored from encrypted storage open correctly.
• Includes the b89 journal fixes and the b88 speech defaults. No new download; permissions and consent are unchanged.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b89-event — betaBeta#

The daily journal reads the whole day again.

• Each update reads its own part of the day. In recent betas every journal update could reuse the day's first minutes, so the journal stopped moving early in the day. Each update now reads the stretch of time it covers.
• Nothing is cut mid-field on the way to the AI. Each piece of evidence is sent whole or not at all, sized to what your model can read.
• One long meeting no longer fills the entry. Meeting excerpts are kept to a fair share and placed in time, so the rest of the day still appears.
• The journal says only what its sources show. A rewrite can't turn a request into a finished task, or a scheduling question into an email you received. Where a section can't be supported, Everpaper shows the source passage instead.
• Text-only models don't get screenshots. With LM Studio, Everpaper checks whether the selected model can read images before sending any.
• A failed read keeps the review open. If a source can't be read while the day is finished, the review stays open with an error instead of showing an old summary.
• Existing journals aren't rewritten automatically. How good the writing is still depends on your local model.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b88-event — betaBeta#

Meetings process faster with MOSS, your custom words help transcription, and word timing can be more precise.

• The MOSS meeting model runs leaner. This beta uses a lighter build of the same model and no longer decodes a short first attempt twice. Earlier tests on two recordings showed about a fifth less processing time and about 30% less peak memory; one longer recording came out slightly worse. The original model remains the fallback, and a change takes effect after relaunch.
• Custom words guide the standard engine. Words you add now act as gentle hints for standard transcription, so names and terms are more likely to come out right. With an empty list nothing changes.
• Optional precise word timing. On Macs with 24 GB or more, an optional 1.8 GB model refines when each word starts and ends, after transcription. It downloads only when you press Prepare, adds some processing time, and never changes the words or who said them.
• The end of a recording is never skipped. Audio reading continues through short reads, so the last part of a recording can't be silently left out.
• LM Studio answers straight away with Qwen3.5. Everpaper switches off Qwen3.5's hidden "thinking" on LM Studio; in one test a 20-second wait became under a second. A reply that only thinks and never answers now shows a clear error instead of an empty answer.
• These options are on by default in this beta and can be switched off in Settings. The wording when a speech-model download is paused is clearer.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b87-event — betaBeta#

Everpaper Native answers finish when the text does.

• No more half-minute wait after a Native answer. With Everpaper Native, an answer kept running for up to about 35 seconds after its last word appeared, holding back its sources and the New conversation button. The engine now stops as soon as the answer ends; in testing, that wait dropped to about a tenth of a second.
• Sources arrive with the answer. A Memory answer shows its full source list as soon as it finishes, without switching tabs.
• Everything else is as in b86.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b86-event — betaBeta#

A short introduction before setup, and questions about a meeting become conversations you can come back to.

• An optional introduction. Before setup, a short interactive introduction shows what Everpaper remembers and how a detail you need later becomes an answer with its source. Drag through the examples or skip it.
• Setup keeps your place. Quit halfway and setup reopens where you were; going back keeps your choices. It explains permissions, downloads and private processing before you act. You can run it again from Settings, and finishing opens Today or Meetings without starting a recording.
• Ask a meeting in threads. Questions about a saved meeting now live in conversations you can return to: a list beside the chat on wide windows, a menu on narrow ones. The composer stays at the bottom, and an unsent draft survives switching conversations.
• Answers show what they used. A meeting answer links to the meeting's transcript and your saved notes.
• Commitments open in place. Clicking a commitment opens a small popover beside it with its title, status and source. Escape closes the source first, then the popover.
• Meeting status takes less room. Processing status sits on the meeting's details line, with Retry visible when something fails. The Transcript, Overview, Ask and Speakers tabs are slimmer.
• Recall connects a promise to what followed. Asking the Concordance about a commitment now cites both the meeting where it was made and the later work on it.
• Sources stay correct when a meeting is reprocessed or corrected, or a source is deleted.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires macOS 27.

2.0.0-b85-event — betaBeta#

Computer History now records what you do by default once you opt in, and Detailed screen history can't record less than Screen history.

• One opt-in now covers what you do and what you type. Turn on Computer History once and it records clicks, shortcuts, short selections and encrypted typed text in ordinary apps and sites, unless you exclude them. If you already use it, nothing widens until you choose. Password fields and private browsing are never recorded.
• Detailed screen history can't record less than Screen history. It keeps the regular full-screen snapshots, then adds a closer look at the active window after real changes, pauses in typing, or a Moment you mark. Without Accessibility permission, only those extras stop.
• Fewer near-identical screenshots. Routine snapshots are smaller, and frames that look the same as the last one are dropped before anything is stored or read. Moments you mark keep full quality.
• Quiet stretches stay short. A long run of switching between windows no longer gets written up as if it were real work.
• Simpler AI choice in setup. "Choose for me" picks Everpaper Native with MOSS on macOS 27 and a managed local engine with MOSS on earlier systems. "I manage local AI" keeps every option visible. Nothing downloads until you press Prepare.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires the macOS 27 beta.

2.0.0-b84 — betaBeta#

Computer History arrives, with a new Adaptive mode, a History that reads as work, and a clearer first run.

• Computer History (tester builds only). An opt-in history of what you do on your Mac: apps and windows, clicks and shortcuts, and — only if you allow it — encrypted typed text. Password fields and private browsing are never recorded, you can exclude apps and sites, and it's stored only on this Mac.
• Adaptive mode reacts to real changes. Adaptive (today's Detailed screen history) follows the apps and windows you use, and takes an image of the window you're working in only when what you're doing changes and every privacy check passes. In this build it took no full-screen screenshots. The text in those images is read on this Mac.
• History reads as work. A day's History now leads with Episodes: short write-ups of what you worked on, each with a title, the apps involved, how long it took and any Moments you marked. The raw list of app switches and clicks is still there underneath, as detail.
• A clearer first run. A new welcome explains Everpaper before setup starts. The first choices separate what Everpaper remembers from the permissions setup will ask for, and the history step is wider, showing what each choice keeps and what it never records, with a summary that updates as you choose.
• Honest about thin days. Vague “Moving between…” rows are replaced with what was actually seen or typed, when that's available. When there's too little to go on, Everpaper says so instead of inventing a task or outcome.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires the macOS 27 beta.

2.0.0-b80-event — betaBeta#

An early tester build of Computer History.

• A first look at Computer History. A small evaluation build, sent out between b83 and b84, that tried the opt-in history of apps, windows, clicks and (only if you allowed it) encrypted typed text that b84 then shipped. b84 and later builds replace it.

This tester build includes evaluation-only Computer History modes; ordinary builds don't include them.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b83 — betaBeta#

First-run setup now asks which intelligence route you want instead of deciding for you.

• Setup is three screens instead of eleven. Choose what Everpaper should remember — your meetings, your days, or both — and setup asks only for what that choice needs.
• Every usable route stays visible. Everpaper Native leads on macOS 27, with Managed oMLX, detected local providers, your own endpoint, and no AI all available from the same screen. The redundant Apple Intelligence row is gone; its icon now identifies Everpaper Native.
• Existing local setups are reused. Setup can rescan providers, preserves the chosen route through the final receipt, and reuses installed oMLX and model files instead of downloading them again.
• Local model answers stay concise. Qwen and LFM2.5 models managed by oMLX now have thinking disabled at the chat-template level, so internal reasoning does not consume the answer.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires the macOS 27 beta.

2.0.0-b82 — betaBeta#

Meetings use MOSS by default, and the built-in intelligence path is easier to choose.

• MOSS is now the default meeting engine. First-run setup names the one-time download before it begins, starts it only when you press the button, and keeps working in the background while setup continues. The standard transcription pipeline remains available as a fallback until MOSS is ready.
• Everpaper Native leads on macOS 27. The built-in engine is now the recommended intelligence choice during setup; a separate local engine remains available as an alternative. System instructions also use the model's real privileged system role.
• Very short recordings are handled honestly. Recordings under five seconds are discarded instead of producing empty meetings, with a clear explanation in the recording pill.
• Completed meetings no longer leave a stale progress row. The meeting list reconciles its live finalization state against the saved record after every refresh.

Requires macOS 14.2 or later, Apple Silicon. Everpaper Native additionally requires the macOS 27 beta.

2.0.0-b81 — betaBeta#

The native engine gets its introduction.

• A one-time card on macOS 27. If your Mac can run Everpaper Native (the in-app engine b80 introduced), Everpaper now tells you so once, and offers to set everything up: it downloads the model, switches your provider, and proves a first response before saying done. Decline and nothing changes; the option stays in Settings.
• The setup survives closing the card — an accepted download keeps going in the background.
• On macOS 26 and earlier nothing appears; the card waits until the Mac can actually use it.

Requires macOS 14.2 or later, Apple Silicon. The Everpaper Native option additionally requires the macOS 27 beta.

2.0.0-b80 — betaBeta#

The first build with an engine made only of Apple frameworks — no separate AI app required.

• New intelligence option: Everpaper Native (macOS 27). On Macs running the macOS 27 beta, the provider picker gains an "Everpaper Native" row. Choosing it runs summaries, extraction, and chat through an engine bundled with the app and built on Apple's frameworks — no LM Studio, no Ollama, no server to install or keep running. On earlier macOS versions nothing changes and the row does not appear.
• The native engine ships with structured output built in (summaries and action items decode against a schema, not hopeful parsing) and streams its responses.
• This beta's supported model is Qwen3-4B-Instruct-2507, prepared for Apple's runtime.
• Everything else — recording, transcription, journal — is unchanged from b79 and continues to use the existing engines regardless of this setting.

Requires macOS 14.2 or later, Apple Silicon. The Everpaper Native option additionally requires the macOS 27 beta.

2.0.0-b79 — betaBeta#

Same app as b78, rebuilt and signed on new hardware.

• No functional changes. This build exists to move Everpaper's release process onto a new machine; everything listed under b78 is included and unchanged. If you are already running b78, there is nothing new to try.
• Verified before release: notarized by Apple, signature and update-key checks passed, and the full test suite green.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b78 — betaBeta#

Pick your local AI model — or let Everpaper pick for your Mac.

• A model choice in first-run setup. Under "Set it up for me", setup now names the model it will install and lets you change it. Every option shows its download size and how much memory it uses on a long meeting — not just the file size.
• The recommendation fits your hardware. Everpaper reads this Mac's memory and only offers models it can actually hold alongside transcription. LFM2.5 2.6B stays the default everywhere; a faster lightweight option is always available, and Macs with 24 GB or more can pick a larger Qwen model for higher-quality summaries.
• Bigger models answer properly now. The larger Qwen options had a habit of "thinking out loud" — pages of reasoning instead of the answer. Setup now configures them to answer directly.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b76 · b77 — betaBeta#

Local AI setup gets a better brain, fitted to your Mac. The two builds are identical; b76 expires after 15 days, b77 never does.

• A new starter model for the one-click local AI setup. Choosing "set it up for me" now installs LFM2.5 2.6B — about half the download of the previous starter, faster on long transcripts, and tuned for exactly what this app asks of it: summarizing, extracting, and answering from your own recordings.
• Lighter Macs get a lighter model. On a MacBook Neo the setup installs a smaller, faster model instead, so summaries finish quickly on passive cooling.
• Reasoning models stop thinking out loud. Models that emit their scratch work before the answer no longer leak it into summaries and notes, whichever local server you use.
• Local AI setup now needs about 6 GB free instead of 10.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b75 — betaBeta#

Settings opens instantly again.

• Opening Settings no longer freezes the app. Since b72, opening Settings could pin the processor and beachball the whole app for ten seconds or more — on some Macs indefinitely. A bookkeeping write fired on every screen refresh while the page was open, and each write forced another refresh. The page now reads that record without touching it, and the loop is gone.
• Nothing else changed in this build; it exists to get the fix out fast. The b74 improvements below are all included.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b74 — betaBeta#

The wait after a meeting stops being a mystery: what step, how long, and one click to the details.

• The pill is clickable now — in every state. While a transcription runs, clicking the floating pill takes you straight to the Meetings page, where the working row shows the step, the percentage, and the time estimate. When it says "Saved", clicking opens the finished meeting. It used to swallow the click silently; right-click and VoiceOver now offer the same routes.
• Every step has a name. Preparing the audio, transcribing, separating speakers, saving — the pill, the meetings row, and the Today page now say which one is running. Before this build, only the transcription itself reported anything, so the bar could sit at 100% for minutes while real work happened in silence.
• "About 6 min left." The transcription now estimates its remaining time from its own measured pace — shown in the meetings list, on the Today page, and during a re-analysis. It appears only once there is enough real progress to be honest, and disappears rather than guess.
• The percentage no longer vanishes mid-way. On a two-track recording the number used to disappear for minutes at the halfway mark; it now carries through.
• A checkmark means done now. The re-analysis view used to show its checkmark when the transcription finished, while speaker separation and saving were still running.
• Smaller honesty fixes throughout. The menu bar says "Transcribing…" instead of a disabled "Stop Recording"; an in-flight transcription no longer disappears from the list while you type in search; a long error message fits inside the pill instead of being cut off; and a rare stall in speaker separation can no longer hold a finished transcript hostage.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b73 — betaBeta#

A sweep through everything the app wasn't saying out loud.

• Deleting a meeting can no longer half-delete it. The audio was removed before the database entry, so a delete that failed partway left a meeting you could still see but never play. The order is reversed, and a failure now says so instead of pretending nothing happened.
• A recording that dies mid-meeting says so. If the microphone drops or the disk fills, the saved meeting now says it ended early — it used to report "Saved" exactly like a normal stop, leaving you to discover the missing forty minutes on your own.
• The expired-beta message actually appears everywhere. In b72, an expired build refused to start the day journal silently — the Start button just did nothing. It now explains itself there too, and Settings keeps saying a build has expired instead of going quiet at that exact moment. (This build does not expire.)
• Re-analysis shows its percentage on every engine. The progress bar was still wired to the old assumption that only one engine could report.
• The menu bar stops claiming "Capturing" during a meeting. It now says "Paused for meeting", matching what the app actually does.
• Quieter, more careful internals. File names no longer land in the system log where a diagnostics bundle could carry them; the command-line tools run stored transcripts through the same credential-stripping pass as the app; deleting a day keeps its screenshots restorable for the full undo window; and several heavy list and audio operations got lighter.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b72 — betaBeta#

Progress you can see on every engine, and a beta that knows when to stop.

• The other two transcription engines report their progress now. Until this build only the advised meeting model showed a percentage; the standard pipeline — the one most people are on — spun with nothing to say for the whole wait. Both now count up, in the pill and in the meetings list. They report a plain percentage rather than naming the part of the recording being worked on, so it is coarser than the advised model, but it moves.
• This beta stops recording after fifteen days. A beta never updates itself, so an old copy would otherwise drift on forever, unsupported. When the date passes, Everpaper declines to start a new recording and points you at the current build.
• Nothing is ever locked away. An expired beta is read-only, not closed: your meetings, your days, search, and export all keep working exactly as before. The recordings on your Mac are yours, and no expiry touches them.
• You get warned first. Three days out, Settings opens with a line saying the build is about to stop recording, and a button that takes you to the current one.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b71 — betaBeta#

Recordings that finish themselves, and a wait you can watch.

• A recording interrupted mid-transcript now finishes on its own. If Everpaper is quit or killed while it is still writing a transcript, the audio was always safe — but finishing it meant knowing to open Meetings and press Recover. Everpaper now rebuilds it the next time you open the app, and tells you it did. When it decides not to, it says so rather than leaving you to find out.
• It gives each recording one automatic try. If rebuilding a particular recording is what stopped the app, Everpaper will not keep walking into that at every launch — that one waits for you to choose, and everything else still rebuilds on its own.
• It stands aside for the work in front of you. Automatic rebuilding never begins during a meeting, stops if you start one, and holds off while your Mac is in Low Power Mode.
• The pill counts up. After you stop a meeting, the floating pill now reads "Transcribing… 47%" instead of spinning with nothing to say. The percentage comes from the advised meeting model; the standard pipeline still shows a plain spinner.

Requires macOS 14.2 or later, Apple Silicon.

2.0.0-b70 — betaBeta#

The meeting you just stopped no longer vanishes while it is written.

• A meeting being transcribed appears in the list while it works. Until it finished, a just-stopped meeting existed nowhere — not in Meetings, not in the Library — so a long transcription looked like nothing was happening at all. It now appears immediately, carrying its own progress, and becomes the finished meeting when it is done.
• Described for what it is. That entry reads "Transcribing", not "Recovering" — recovery wording is kept for a recording genuinely orphaned by a crash, so it still means something when you see it.
• The day journal stops saying it is recording when it is not. Screen capture pauses while a meeting records; the journal now says so, instead of showing "Recording" above a frame count that cannot move. Marking a moment still works throughout.
• Capture that quietly fails is now visible. A tick that captures frames but cannot store them, or a second display that stops producing frames, is reported rather than passing for healthy.
• A long transcription reports its progress to the system log. If one ever does get stuck, there is now a record that tells a working decode apart from a stalled one.

Requires macOS 14.2 or later, Apple Silicon.

1.16.0Release#

One model for meetings: words and voices together.

• Meet MOSS, the new advised meeting model. Until now, Hindsight transcribed meetings in two passes — one model wrote the words, a second worked out who spoke, and the two were stitched together afterwards. MOSS (MOSS-Transcribe-Diarize, Apache 2.0) does both at once: a single on-device model that transcribes and tells voices apart as it listens, so speaker changes land where sentences change hands. It is advised for meetings and entirely opt-in — about a 2 GB one-time download, and audio never leaves your Mac.
• New users meet it during setup. First-run setup gains a "Tell Speakers Apart" step: accept to download MOSS in the background while you keep walking through setup, or decline and keep the standard pipeline — nothing blocks, and you can change your mind later in Settings.
• Existing users get one quiet announcement. On first launch after updating, a single "New in this impression" notice explains the model and offers to set it up in place, with live download progress. Dismiss it and it never returns — a Press Check line and the new Meeting Model entry under Settings → Transcription & Speech remain the permanent path.
• Switching back is one click. "Use instead" returns meetings to the standard Parakeet pipeline at any time; the downloaded model is never deleted, and typing, dictation, and file transcription keep using Parakeet or Whisper exactly as before.
• Long meetings stay light. MOSS processes audio in ten-minute windows and stitches speakers across them, so memory stays flat (about 5 GB during finalization) no matter how long the meeting runs. The live meeting pane keeps its instant preview; the final pass after you stop is the authoritative transcript.
• Honest edges. Word-level timestamps inside a sentence are approximate with MOSS (segment boundaries are exact), and a voice that stays silent for a very long stretch may re-enter under a new speaker label. Raw and Clean Lexicon modes apply to MOSS transcripts exactly as they do everywhere else.
• For the terminal. hindsight-cli transcribe --engine moss --diarize runs the same joint model on any audio file.

Requires macOS 14.2 or later, Apple Silicon. The meeting model is happiest on 16 GB machines.

1.15.1Release#

Raw now means raw for meetings too.

• Meeting transcripts honor the Lexicon mode. Raw leaves recognized meeting wording untouched; Clean applies your Custom Words corrections during meeting finalization, crash recovery, and Re-analyze.
• Re-analyze stays consistent. When you rerun transcription and speaker separation on a completed meeting, Hindsight uses the current Raw/Clean setting instead of silently applying corrections every time.
• The Lexicon is clearer. The Processing mode copy now says it applies to both file transcriptions and meeting transcripts, and Custom Words are described as Clean-mode corrections.

Requires macOS 14.2 or later, Apple Silicon.

1.15.0Release#

One click, one brain: setup can now install your local AI for you.

• Download & Set Up — and that's it. If no local AI server is running, first-time setup now offers to do everything: it installs oMLX (a free, notarized inference server built for Apple Silicon), downloads the Qwen3.5 4B model (~4 GB) from Hugging Face with live progress, starts the server, and connects Hindsight the moment the model is ready. No account, no API key, no terminal — and nothing ever leaves your Mac.
• Interruption-proof. A dropped connection doesn't restart the 4 GB download — setup resumes from the files that already finished. If oMLX's own first-run screen holds things up, Hindsight starts the server directly so setup still completes on its own.
• Your server, your choice. Prefer LM Studio? One click opens its download page and Hindsight connects automatically as soon as its server appears. Already running Ollama, LM Studio, or oMLX? Setup finds it and connects with one click, exactly as before. And on Macs with Apple Intelligence available, there's a zero-download Connect that uses the on-device Apple model.

Requires macOS 14.2 or later, Apple Silicon.

1.14.0Release#

First-time setup, rebuilt — the front door now matches the book behind it.

• Connect your intelligence during setup. A new step finds a local AI server already running on your Mac (Ollama, LM Studio, oMLX) and connects it in one click — no key, no cloud, nothing leaves your machine. Recall, the Second Brain, and meeting prompts work from your very first day instead of waiting for a trip to Settings.
• The download no longer holds you hostage. The speech model starts downloading the moment setup opens and keeps going while you walk the earlier steps — by the time you reach the models page it's often already done.
• Setup finally teaches the good keys. The rehearsal now covers ⌥Space — your memory from anywhere on the Mac — alongside ⌘K, and the meetings step mentions ⌘⇧M for starting a recording from any app.
• Finish by starting. The last page ends with a real "Start Journaling" press that begins today's record on the spot, plus an optional "Open Hindsight at login" so the daybook is running on the days that matter. And if macOS asks for a relaunch mid-setup, you now resume right where you left off instead of starting over.
• It moves like the book. Steps turn like pages with a live progress rail, permissions flip to a green check with a small bounce the moment they land, and the whole flow quiets to simple fades when Reduce Motion is on.

Requires macOS 14.2 or later, Apple Silicon.

1.13.0Release#

The Colophon learns to answer first: settings you can find, a health check that greets you, and one grammar for every control.

• The Press Check. Settings now opens with its own health: anything unhealthy — microphone, screen recording, a failed AI test, a missing speech model — appears as one line with its one fix. When all is well, it says so in a single quiet line and stays out of the way.
• Find any setting, in place. A permanent "Find any setting…" field lives in the pinned header (⌘F). Typing filters the page to matching settings — synonyms included, so "mic" finds Audio Input — and choosing one lands right on the row. Esc restores the page.
• Names you can guess. Sections now lead with plain titles — Meetings, Transcription & Speech, Shortcuts & System, This Copy — with the book's own names kept as quiet subtitles beneath. "Change my hotkey" finally has an obvious home.
• One grammar for every control. Actions that sit beside a status word are real buttons now (Test Input, Test Connection, Download), and Advanced settings live in a visibly inset group at the end of their section — no more guessing what's inside.

Requires macOS 14.2 or later, Apple Silicon.

1.12.2Release#

The book finds its voice: reading type everywhere, and a calmer Today page.

• Prose is serif everywhere now. The sidebar, Today's Focus, the Ask bar, chat copy, previews, empty states, and settings descriptions all read in the book's serif — the near-monospace face is retired to what it does best: timers, folios, stamps, and captured text.
• Start means start. The day journal's Start button is now sized like the page's main action instead of a small chip under the feature list.
• A calmer Today. One heading owns the page (the date), the journal card sits inside it, Today's Focus collapses to a slim strip when there's nothing to act on, and the Ask bar and its sample question read as one composed block.

Requires macOS 14.2 or later, Apple Silicon.

1.12.1Release#

A polish release: reading comfort, tidier lists, and small consistency fixes across the book.

• Transcripts read like a page, not a terminal. Meeting transcripts are now set in the book's serif at a comfortable reading width with proper line spacing — an hour-long conversation no longer renders as full-width monospace. Timestamps and speaker chips stay in the data register.
• Cleaner previews. Archive cards preview their transcript in the reading face instead of monospace, and the feature notes on Today read as quiet prose instead of a settings list.
• Small consistency fixes. The Library page now matches its sidebar name ("The Archive"), meeting dates share one format everywhere, date group headers match the app's eyebrow style, the engine status chip no longer fuses with the Import Transcript… button, and a meeting that lives on two calendars shows once in Upcoming instead of twice.

Requires macOS 14.2 or later, Apple Silicon.

1.12.0Release#

The Concordance learns to hold a conversation — and recall gets sharper everywhere it looks.

• Follow up without re-asking. Answers in the Concordance now build into a conversation: type the next question and the panel carries the thread with it. ⌘↵ still promotes the exchange into the Second Brain — now with its context attached, so a follow-up like "did we set a deadline?" survives the jump.
• "What did I do on my last recorded day?" answers. Recall always anchors on your latest recorded days and meetings, so time-based questions get an answer even when they share no words with what was on screen.
• Long meetings answer from the right minute. Recall now quotes the passage that actually matched — minute 40 of an hour-long meeting included — instead of only ever reading a meeting's opening minutes. Natural-language questions also match meetings and files far more reliably, and your indexed folders are searched by meaning, not just exact words.
• Transcript import, hardened. Teams and Windows exports (CRLF line endings) import correctly, Zoom-style "Name:" speakers are recognized, timings survive so exporting an imported meeting as srt/vtt works, and big imports no longer freeze the app — the imported meeting opens for you when it's done.

Requires macOS 14.2 or later, Apple Silicon.

1.11.0Release#

Concordance gets a real memory of your meetings, and outside transcripts can join your library.

• Concordance actually finds your meetings now. Asking "what did I talk about in my last meeting?" used to come up empty — the AI answer only ever looked at your Days, never at meeting transcripts. It now grounds on meetings and indexed files too, with real semantic search over what was said, not just literal keyword matches.
• Import a transcript, get a meeting. Have a .vtt transcript from Zoom, Teams, or Otter but no Hindsight recording? Meetings → Import Transcript… turns it into a real meeting — searchable and recallable just like one you recorded, no audio required.

Requires macOS 14.2 or later, Apple Silicon.

1.10.0Release#

Meet the Concordance — your memory, one keystroke away, from anywhere on your Mac.

• Press ⌥Space anywhere. A floating slip of the book appears over whatever you're doing — another app, another Space, even a full-screen call — without stealing focus from your work. Type to search your days, meetings, beliefs, and settings instantly; arrow keys and Return jump straight to the page in Hindsight.
• Ask without asking. Pause typing for half a second and the Concordance starts recalling: an AI answer grounded in your captured days streams in right above the results. It's ephemeral — panel answers never clutter your Second Brain conversation. Press ⌘Return to continue the thought in the full app when it deserves a longer conversation.
• Scopes and files. ⌘1–4 narrows to everything, days, meetings, or files. File matches search your indexed folders and reveal the document in Finder.
• Your shortcut, your call. Rebind the summon in Settings → The Apparatus. (We default to ⌥Space rather than a global ⌘K on purpose — a system-wide ⌘K would silently eat the command palette in your editor, browser, and Slack. Inside Hindsight, ⌘K remains the Index.)

Requires macOS 14.2 or later, Apple Silicon.

1.9.1Release#

A quick fix on the heels of 1.9.0.

• Extract To-Dos works on any meeting. Extracting to-dos from a transcript silently did nothing for meetings older than three days — a Today-rail recency rule had leaked into the explicit action. It now works regardless of age, tells you the outcome ("3 to-dos added to Today." or an honest none-found note), and re-extracting resurfaces to-dos you'd previously exported or dismissed instead of silently skipping them.

Requires macOS 14.2 or later, Apple Silicon.

1.9.0Release#

A design release. Every page got a pass from an expert review of the whole interface — smarter suggestions, honest insights, readable type, and dozens of paper cuts gone.

• Suggestions you can trust. Tell Hindsight your email addresses (Settings → Second Brain) and it stops mistaking your own name for a person to track — no more "assess your own relevance" suggestions. Speculative observations can't become Apple Reminders anymore, websites and people no longer masquerade as tasks, and the same insight stops appearing twice in different words.
• A quieter, sharper Second Brain. The You tab rolls websites up to the site (not every page), caps the lists, and speaks to you — no more raw tags, enum names, or numbered leftovers. Chat now owns its tab with one scroll instead of a scroll-inside-a-scroll, and duplicated model answers collapse automatically.
• Meetings name themselves. A finished recording takes its title from the overlapping calendar event — no more walls of "Meeting 23 Jun at 18:37". Rapid-fire crosstalk renders compactly, phantom "speakers" with five words step out of the speaker list, transcript previews read like sentences, prompts show as runnable actions and replace their old results instead of stacking duplicates.
• Days that read like a day. Timeline entries describe what was actually on screen (never keyword soup), quick message-checking hops fold into one "Checking messages" row, and a new MARGINALIA column surfaces the day's open questions and loose ends — only when there are any, and never invented.
• Easier on the eyes. Long AI answers now read in the book's serif with proper line spacing instead of typewriter-style text, numbered lists align properly everywhere, and the Archive gained date groups, clickable favorite stars, a clearer filter bar, and word-by-word search.
• Settings you can navigate. The section index stays pinned while you scroll, ⌘F hint included; destructive actions look like the buttons they are; and the Lexicon finally dresses like the rest of the book.

Requires macOS 14.2 or later, Apple Silicon.

1.8.2Release#

A quick follow-up to the file-indexing rework: live status you can actually watch, and proper patience with local AI models that are still loading.

• Indexing shows its work again. The progress bar appears the moment indexing starts — no more long silent stretch on a fresh index — and a live status line tells you exactly what's happening: "Reading files…", "Splitting files by meaning…", "Embedding chunks (batch 2/7)…". Pause and Cancel are always visible while a run is active, and pausing now takes effect mid-run instead of waiting for the current batch of files to finish.
• Cold models are waited for, not skipped. If your local AI server is still loading a model when indexing starts, Hindsight now waits for it — retrying for a couple of minutes with a "Waiting for the embedding model to load…" status — instead of silently falling back to the weaker built-in embedder for the whole run. Meaning-aware chunking gets the same fix: a model that takes more than a minute to load no longer loses semantic splitting for every file.

Requires macOS 14.2 or later, Apple Silicon.

1.8.1Release#

A focused reliability release for reminders — the Today rail's suggested to-dos and calendar meeting alerts.

• "Add to Reminders" works now. Exporting your to-dos to Apple Reminders crashed the app the moment macOS asked for permission — a packaging bug, fixed and now guarded by a test. If access is denied, Hindsight explains why and offers an Open System Settings shortcut instead of failing silently.
• Suggested to-dos you can trust. AI-extracted tasks are now checked against the meeting or screen text they claim to come from — an invented task never reaches the rail or your Reminders. Rephrased titles that drift from the original task fall back to the plain version, owner-style items ("Jane: send the brief") are picked up, and filler lines ("None", "No action items identified") no longer become to-dos.
• Dismissed means dismissed. To-dos you dismiss, complete, or export stay gone across relaunches — and a suppression quietly expires after a month, so a recurring task with the same name isn't hidden forever.
• Meeting alerts actually arrive. Reminders now show as banners while Hindsight is the frontmost app (they were silently suppressed before), a Mac that slept through the reminder window still alerts you on wake while the meeting is joinable — with honest timing ("starts in 2 minutes" / "has started") — and declining notification permission no longer switches the whole calendar feature off.
• Fresher suggestions. Meeting action items older than a week stop resurfacing on Today as if they were new.

Requires macOS 14.2 or later, Apple Silicon.

1.8.0Release#

File indexing got a ground-up rework — dramatically faster and easier on your Mac — and your indexed files now back up the Second Brain's answers directly.

• File indexing is fast now. Folders index in parallel waves instead of one file at a time, re-indexing an unchanged folder finishes in seconds with zero wasted work, and a single huge file can no longer stall the run. On machines running a local AI, indexing also stops swapping models back and forth for every file — the biggest source of slowdown.
• Answers cite your files. Ask the Second Brain something that lives in an indexed folder and the answer now draws on those files, with [file:n] citations pointing at the exact file and lines it used.
• Snappier everywhere. Sending a question paints instantly instead of waiting on retrieval, the Second Brain tab opens without a hitch, and recall, day chats, and the Days list no longer block the interface while they read the journal.
• Retention that means it. A new toggle — "Also delete their text & recall data" — makes the retention setting remove expired captures' text, search index, and recall entities too, so old data truly expires and the database stays bounded. It's on for new installs; existing journals keep their history unless you opt in. Cleanup now also runs daily, not just when a session starts.
• Meetings, tightened. Stereo playback now lines up with the transcript timeline, and the live preview cuts at natural speech pauses instead of mid-word. Plus a batch of fixes: the AI vision-descriptions toggle actually works now, importing from the legacy app no longer freezes the window, and chat vs. embeddings API keys no longer overwrite each other.

Requires macOS 14.2 or later, Apple Silicon.

1.7.0Release#

Export your meeting transcripts, get real speakers out of single-microphone recordings, and watch — and pause — file indexing while it runs.

• Export any transcript. Save a meeting's transcript to a file — plain text, Markdown, subtitles (SRT or VTT), or JSON — from the new Export button on the meeting, or "Save Transcript…" in the Library's right-click menu. Everything is written locally to a file you choose; nothing is uploaded.
• Single-microphone recordings get real speakers. A meeting captured from a single audio source — just your microphone, or one where the system audio dropped partway through — used to collapse into one speaker labelled "Me". Now Hindsight splits it into separate speakers whenever it hears more than one voice, while a genuine solo recording still reads as "Me". Re-analyzing an existing meeting picks this up too.
• See — and pause — file indexing. Indexing a folder now shows a real progress bar (files done out of the total) instead of an open-ended spinner, and you can pause a long index and resume it later from where it left off. Pausing never leaves a half-built index behind.

Requires macOS 14.2 or later, Apple Silicon.

1.6.0Release#

A more AI-driven Second Brain — pick your own embedding model, smarter suggested to-dos, and meaning-aware file indexing — plus a wave of reliability fixes across reminders, your days, and the Second Brain.

• Choose your embedding model — and it's used everywhere. The embedding model you pick for Folder Context RAG now drives both indexing and search, so your files and your queries are always embedded with the same model (a mismatch used to silently break semantic search). Point it at a local model in Ollama, LM Studio, or oMLX and nothing leaves your Mac; a cloud embedding model only ever receives your file text if you explicitly choose one and grant consent.
• Smarter suggested to-dos. Today's task suggestions are rephrased into natural, imperative tasks; meeting action items are extracted by the model from the meeting text; and tasks you pull from a specific meeting now stay on the rail when the page refreshes. Each step falls back to the built-in, on-device behavior when no model is connected.
• Meaning-aware file chunking. With a local model configured, files are split into sections by meaning instead of by paragraph or page, so retrieval lands on coherent passages — and it falls back cleanly to structural chunking otherwise. PDFs and spreadsheets keep their page-accurate citations.
• File search actually uses your vectors. Semantic file search now runs whenever you've typed a query (multi-word queries used to silently drop to keyword-only), re-indexes only the files that changed instead of re-embedding the whole folder every time, can be cancelled mid-run, and tells you when a folder has nothing to index.
• Reminders that behave. "Add to Reminders" confirms when it worked, won't pile up duplicates, and writes the whole list in one pass — and an imminent meeting's reminder is no longer dropped when notification permission was still warming up.
• Faster, more honest day pages. Opening a day no longer hitches the interface, screenshots use far less memory, the timeline shows the whole day (not just the first stretch of it), per-app time and on-disk storage are reported accurately, and deleting a day takes a single confirmation.
• A steadier Second Brain. Beliefs no longer fade away prematurely, so your self-model and persona stay populated; chat citations resolve to the right frames after you reopen a conversation; and the self-model keeps updating even in Low Power Mode.

Requires macOS 14.2 or later, Apple Silicon.

1.5.2Release#

Debugging tools for the embeddings provider — so you can tell when your model isn't loaded and fix it without reading server logs.

• Test Connection for embeddings. The embeddings provider card now has a "Test Connection" button (just like the chat AI). It sends a tiny embed request to your configured endpoint and tells you whether it worked — including the served model and vector dimensions — or shows the exact error. No more guessing whether your model is loaded or the name is right.
• Diagnostic logging. The embeddings path now logs to Console.app (subsystem com.iannellomarco.hindsight.core, category Embeddings + FileIndexer). You can see exactly what model was requested, whether it succeeded, and when the app fell back to the on-device model — filter Console by the subsystem and send the logs when something's off.

Requires macOS 14.2 or later, Apple Silicon.

1.5.1Release#

A dedicated embeddings model picker for Folder Context RAG — and three fixes the adversarial code review caught before they reached you.

• Pick your embeddings model. Folder Context RAG now has its own provider config in Settings → Intelligence → File Index, parallel to your chat AI. Point it at Harrier OSS v1 0.6B (or any embedding model) in your Ollama, LM Studio, or oMLX instance — independent of your chat model. A refresh-from-provider list helps you pick. With none set, files use the on-device model.
• Saves take effect immediately. Configuring the embeddings model used to require an app relaunch (a stale-capture bug froze the old config). Now the index and retrieval read the live setting on every call — change the model, hit re-index, and it's used right away.
• File content never leaves your Mac. The fallback path that used your chat AI for embeddings is now gated to local providers only — a cloud chat provider will never receive your file text for embedding.
• Retrieval uses the right model. A matching fix ensures the query vectors and the indexed vectors come from the same model, so semantic search actually works (it was silently mismatched before).

Requires macOS 14.2 or later, Apple Silicon. We recommend Harrier OSS v1 0.6B (the MLX 8-bit build) in your Ollama or LM Studio instance for file embeddings.

1.5.0Release#

Your Second Brain now leads with genuine AI insight, reads your own files, and works reliably on local models — all on-device, all private.

• The page now leads with what the AI actually did. The Second Brain used to open with templated counting ("X sits at the center of a lot of what you do", cycling a few phrases). Now the AI Reflection — grounded observations written from your real captured moments — leads the page, and the templated belief cards are rephrased into varied, specific observations by the model. The vanity stat row is gone. With no AI connected, you get an honest "connect a model" prompt instead of fake templated confidence.
• Your Second Brain can now read your files. Point Hindsight at a workspace folder — a repo, a docs folder, a papers folder — and it indexes everything inside on your Mac: code, markdown, PDFs, images (via OCR), spreadsheets, rich text, email. The Reflection can then connect your screen activity to your own work artifacts, cited by file. "You switched to Slack right after the compile errors your plan flagged." Nothing leaves your Mac; embeddings run on your local model.
• Local AI is now reliable enough to be the primary brain. Hindsight now asks your local model (Ollama, LM Studio, oMLX) for structured JSON output instead of prompting and hoping — the change that makes a 3–8B local model dependable as your main Second Brain, not a fallback. Cloud stays optional and opt-in.
• The Reflection checks itself, and sees the right frames. Every reflection is now verified against your actual captured data before it sticks — a fabricated name paired with a real citation gets stripped. And it retrieves semantically-relevant frames (not just the most recent ones), so a reflection about "the migration" pulls migration moments from across the week.

Today's Focus is sharper and lands in Reminders:

• Cleaner, more useful todos. Citation markers and raw URLs/IPs no longer leak into your todo titles, and machine-internal detail text is replaced with human guidance. A tighter, higher-quality list.
• Send to Reminders, not Notes. The export now creates one checkable reminder per todo in a "Hindsight" list — native checklist, individually completable, syncs to your iPhone. Action items belong in Reminders.

Requires macOS 14.2 or later, Apple Silicon. Folder Context RAG and the local-AI improvements work best with a local embedding model (we recommend Harrier OSS v1 0.6B, the MLX 8-bit build, in your Ollama or LM Studio instance) — but everything degrades gracefully to the on-device model when none is running.

1.4.1Release#

Today's Focus is now smarter, always visible, and reachable from every meeting.

• Extract Todos from any meeting. A new checklist button in the meeting detail reads the transcript and pulls out the action items — things you said you'd do, were asked to do, or agreed to. They land in Today's Focus immediately, each tagged with the meeting it came from.
• Todos from real content, not just patterns. When an AI provider is connected, Today's Focus now reads the actual text of your meeting transcripts and on-screen captures — not just the statistics — and extracts specific, actionable tasks ("Send the brief to Jane by Friday") instead of pattern inferences ("You haven't touched plan.md in 8 days").
• Always-visible, always-refreshable. The Today's Focus rail is now always on screen — even when empty, it shows what to do to surface tasks. A refresh button lets you regenerate on demand.
• Auto-extract after meetings. When a meeting finishes, its action items are extracted automatically — you don't have to do anything.

Requires macOS 14.2 or later, Apple Silicon.

1.4.0Release#

Your Second Brain now genuinely thinks about your work — and the whole app got sharper, more honest, and more consistent.

• The Reflection. After you finish a day (or when you open Hindsight in the morning), your Second Brain reads the actual moments it captured — not just the counts, but what was on your screen, in what order, at what time — and writes a short Reflection: two or three genuine observations a thoughtful colleague might make. Every claim points back to the specific moments it noticed it in. It runs on your Mac if you use Apple Foundation Models or Ollama — free, private, no cloud needed.
• A self-model that develops over time. The Reflection's observations now become lasting beliefs. When the same pattern recurs across days, it firms up; when it doesn't, it fades. Recurring patterns promote to stable traits, and those traits shape what the Second Brain notices next — so its understanding deepens, not just accumulates.
• The story of becoming. The "Who you are" card on the You tab now tells a narrative — what the Second Brain has come to understand about you, how that's changed recently, and where it sees tension.
• Today's Focus. A new panel on Today surfaces suggested to-dos from your patterns — files you opened then dropped, topics that faded, action items from meetings. Check one off, dismiss it, or send the whole list to Apple Notes with one click.
• Hide your voice from transcripts. A one-click "Hide my voice" toggle on any meeting transcript. Your turns disappear from the view, but nothing is destroyed — bring them back anytime. It survives re-analysis too.

Trust and privacy hardening:

• Every answer is now verified. The Second Brain's honesty guard — which strips any claim it can't ground in your actual captured data — now runs on every chat answer, not just the relational ones.
• Settings you can trust. Three toggles that promised behavior but never actually did anything have been fixed or removed. The cloud-consent controls are now one clear decision instead of three cross-referencing toggles.

Plus: the Apple Notes Automation entitlement for one-click todo export, LLM-assisted phrasing of suggested to-dos when a provider is connected, meeting action items flowing into the Today's Focus rail, and dozens of accessibility improvements.

Requires macOS 14.2 or later, Apple Silicon.

1.3.1Release#

A smarter privacy boundary for on-device AI — and a more consistent feel throughout.

• Local AI now grounds on the real thing. When you use an on-device provider (Ollama, Apple Foundation Models, or LM Studio), Hindsight no longer strips contact details from what the model sees — names, emails, and phone numbers are preserved so your Second Brain and chat answers reference real people and projects instead of [REDACTED:name]. Nothing leaves your Mac, so there's no privacy cost. Secrets (cards, keys, tokens) are still scrubbed for every provider.
• Meeting chat speaks up when it fails. A question to a meeting that can't reach the AI provider now shows the reason inline, instead of going silent or leaving a placeholder message behind.
• One Send, everywhere. The ask boxes in the meeting panel, meeting detail, and journal chats now share the same look and the same Send control — a consistent composer across the whole app.
• Keyboard tabs in meetings. Switch a meeting's Transcript / Results / Chat / Speakers with ⌘1–⌘4, matching the rest of the app.

Plus the usual polish: standardized section headings and spacing, a centered empty state on Meetings, and sharper accessibility labels.

1.3.0Release#

Find any day in your history at a glance, and never lose one to an accidental delete.

• The Day Index. Press ⌘L anywhere to open a chronological overview of every captured day — title, date, and a thumbnail for each, with a density strip showing how much history you've built. Use ↑↓ to move and Return to turn straight to a day.
• Recoverable deletions. Deleting a day now removes it from the book but offers an Undo right after, from Today — so a mistake never costs you a captured day.
• Always-on sidebar. The book's Parts are visible by default on launch; unpin from the masthead toggle if you'd rather have the room.
• Richer timeline. Every activity row now shows a thumbnail you can click to open full-size — not just the starred moments.

A real privacy hardening for AI vision:

• Secret-bearing frames never leave your Mac. If a screenshot's captured text revealed a card, key, or token, its pixels are no longer sent to a vision provider — the text redactor can't see into the image, so the whole frame is withheld.
• Honest about what vision sees. The vision setting and the on-day redaction badge now say plainly: captured text is scrubbed, but visible pixels sent for descriptions are not.

And a sweep of reliability and polish: the recording control no longer lets Discard look like the main action; a failed Second Brain question now tells you instead of hanging; the "relaunch to apply Screen Recording" state stays visible until you act on it; a new multi-step progress bar in setup; and consistent ask fields, status dots, and delete confirmations throughout.

1.2.0Release#

The Second Brain now answers from what your screens actually showed — not just who and what kept recurring.

• Answers from real content. Ask "why do I keep working with this client?" and the chat reads the actual text of the relevant moments — emails, docs, messages — and grounds its answer in them, with citations you can click to jump straight to the frame.
• It knows you're not your own coworker. The Second Brain no longer lists you among the people you work with.
• Tighter privacy at the edge. When a cloud model is in use, email addresses and phone numbers are scrubbed from what's sent — on top of the existing secret redaction.

The Days page is also redesigned:

• A sidebar for the book. An always-available table of contents. It auto-hides to give the page room — hover the left edge to peek it open — or pin it from the toggle in the top-left.
• Timeline previews. Each moment in a day's timeline now shows a thumbnail you can click to open full-size.
• A bigger delete control, and clearer labels distinguishing asking the whole book from chatting with a single day.

1.1.1Release#

Two fixes to the Second Brain.

• Dismiss sticks. When you dismiss something the Second Brain inferred about you, it no longer quietly comes back the next day.
• Sharper recall. Asking about a specific person or file now surfaces the right ones to jump to, instead of coming up empty on a normally-worded question.

1.1.0Release#

Your Second Brain now understands you — not just the people, files, and projects in your work, but who you are and how your days have been going.

• It knows you. Tell it your name and role (during setup, or anytime in Settings) and it speaks to you personally — greeting you by name and writing your day as someone it actually knows.
• It learns you. From your own days it quietly works out things like your role, the way you work, and how you've been feeling — and asks you to confirm the important guesses before trusting them. Nothing is assumed, and anything that feels off is one tap to dismiss.
• A new "You" tab. A home for your profile: a short portrait of who you are, the traits it has noticed, and a timeline of how your days have felt.
• Private by default. All of this stays on your Mac and works with on-device AI. Sharing your profile with a cloud provider takes a separate, explicit opt-in.

1.0.12Release#

A broad polish pass: more reliable meeting status, far wider accessibility, and a faster, more consistent interface.

• Meeting status updates instantly. The menu-bar recording glyph, the permission prompt when a recording is blocked, and the "transcribing…" indicator now reflect what's happening the moment it changes.
• Much better VoiceOver support. Buttons and controls across the app — Meetings, the Day Journal, the Second Brain, onboarding, and the Index — now announce meaningful names, and secondary text meets WCAG AA contrast.
• Snappier reading. Opening a finished day, searching the Index, and reading streamed answers are all faster and smoother.
• Clearer meetings. The section is consistently called "Meetings", larger calls get more distinct speaker colors, the connect-calendar prompt refreshes after you grant access, and a failed prompt now shows an error instead of doing nothing.
• Rebindable meeting hotkey. The global start/stop shortcut (default ⌘⇧M) can now be changed in Settings → The Minutes.
• Smaller fixes. Onboarding scrolls on small displays, the meeting audio player resets at the end of a clip, deleting a custom word asks first, and a failed app summary can be retried.

1.0.11Release#

A fix so a finished day always shows up where you expect it.

• New days appear right away. After you stop and save a day, it now shows up on the Days page immediately — previously a freshly finished day could stay hidden there until you quit and reopened the app.

1.0.10Release#

Much better speaker detection, plus new ways to fix up a meeting after it's recorded.

• Far better speaker separation. A one-on-one call is no longer split into a handful of phantom speakers, and genuine group meetings keep everyone — the speaker count is no longer capped, so a large call isn't truncated down to a few voices.
• Re-analyze a meeting. A new button on a saved meeting re-runs transcription and speaker detection on the original audio — so you can pick up these speaker improvements on meetings recorded earlier, without recording again.
• Rename a meeting. A pencil next to the title renames a meeting at any time.
• Name a meeting from your calendar. One click names a meeting after the calendar event it overlaps, so your recordings match your schedule.
• Refreshed meeting tabs. The Transcript / Results / Chat / Speakers tabs now use the same thumb-tab style as the rest of the app.

1.0.9Release#

A formatting hotfix for AI answers.

• AI answers are now formatted. Responses from Ask, the Second Brain, and meeting Q&A render their Markdown — bold, italic, headings, and bullet lists — instead of showing raw "**" and "*" characters.

1.0.8Release#

Fixes meeting transcription.

• Transcription no longer freezes. A scheduling deadlock could leave the live transcript blank during a meeting and the post-recording "transcribing…" step stuck indefinitely. Recording and transcription now run on separate lanes — the live transcript fills in during the meeting and the final transcript completes promptly.
• More accurate speakers. Tightened speaker separation so a single remote participant is no longer occasionally split into two, and a slow speaker-detection step no longer holds up the final transcript.

1.0.7Release#

Clearer day navigation.

• Visible page controls. The Days page now has previous / next buttons and a dotted day indicator beneath it — one dot per day, with the current day highlighted. Click any dot to jump straight to that day. (The page-turn controls were previously invisible edge zones.)

1.0.6Release#

Turning the pages — fixed, and nicer.

• Day navigation works again. The left/right arrow keys (and ⌘[ / ⌘]) turn between days without having to click the page first, and the optional two-finger swipe (Colophon → "Turn pages with a two-finger swipe") is reliable.
• A page-peel turn. Days now turn with a subtle page-peel animation instead of a flat slide. Honors Reduce Motion.
• New app icon. A Daybook cover with the rubric-red ribbon bookmark.

1.0.5Release#

Meeting recording now starts instantly.

• No more waiting to record. Previously, starting a meeting could pause on "Preparing speech model…" while the on-device model loaded. Now recording begins immediately and captures audio from the first second; the speech model loads in the background.
• Visible transcription progress. While the model loads, the transcript area shows a progress bar instead of a frozen screen, and the transcript fills in as soon as the model is ready. The full meeting is always transcribed when you stop.

1.0.4Release#

A reliability release for meeting recording — especially your first one after installing.

• Meetings wait for the speech model. If you start recording before the on-device speech model has finished its one-time setup, Hindsight now shows "Preparing speech model…" and begins once it's ready, instead of recording a meeting that comes back with an empty transcript.
• A more responsive Mac while recording. Day Journal capture pauses for the duration of a meeting, so screen capture and live transcription no longer compete for the Neural Engine — and it resumes automatically when the meeting ends.

1.0.3Release#

A small reliability release for first-run setup.

• Clearer Screen Recording setup. When Hindsight is opened from outside your Applications folder — where macOS runs it from a temporary, locked-down copy — setup now tells you to move Hindsight to Applications and reopen it, which is the only way the Screen Recording permission will stick. Previously it pointed you at the "+" button, which can't take effect in that case.
• Sturdier packaging. The disk image is now signed unambiguously, and the diagnostics behind the Screen Recording grant were tightened so first-run setup issues are easier to trace.

No feature changes — everything in 1.0.0 below is unchanged.

1.0.2Release#

• Fixed a launch crash on some Macs. The bundled Daybook fonts failed to load on Macs other than the one the app was built on, which could crash Hindsight at startup. Fonts now load reliably on every Mac.
• Smoother Screen Recording grant. After you allow Screen Recording, setup now shows a clear Relaunch Hindsight step — macOS only applies this permission to a freshly launched app, so the relaunch is what makes the grant take effect.

1.0.1Release#

A packaging release focused on a smoother, more reliable first launch.

• Notarized and stapled. The app is signed, notarized by Apple, and the notarization ticket is stapled into the build. macOS can verify Hindsight offline, so the first launch is clean even without a network connection — no spurious "damaged" or "cannot be opened" warnings.
• Steadier installs. Hardened the DMG so dragging Hindsight to Applications and opening it the first time just works.

No feature changes — everything in 1.0.0 below is unchanged.

1.0.0Release#

The first release of Hindsight: a private, on-device daybook for macOS. One continuous book you turn through page by page, with three pillars and an Index (⌘K) that jumps anywhere.

Day Journal#

A running record of your day, woven from periodic screen capture read with on-device OCR (Apple Vision). Turn between days with the left and right arrows, let the red Ribbon bring you back to today, and use Ask the book on any page to pose a question about your day or your history. Screenshots and the journal database never leave your Mac.

Meetings#

Record and transcribe meetings on-device. Start one from the Record tile, the menu bar, the Index (Begin minutes), or the ⇧⌘M hotkey. Hindsight can watch your calendar and offer to begin a recording when a meeting starts, recognizing Google Meet, Zoom, and Microsoft Teams links.

• Records your microphone and your Mac's system audio as separate, time-aligned tracks (the default; System Audio Only is also available).
• A floating pill and a live panel with Notes, a speaker-labeled Transcript, and an Ask tab.
• Saved meetings keep their transcript, notes, and any Ask conversation, and can be renamed and re-run through prompt templates.

Second Brain#

A page that reads across every day you've captured, links the things that keep recurring, and forms plain-language beliefs about your focus. It moves between Chat, Insights, Network, and Export. The beliefs, Insights, and Network are computed locally and work with no AI provider connected; recall grounds every chat answer in your real history with citations back to the moment each claim came from.

On-device by default#

• Capture, transcription, and storage stay local. Meetings are transcribed on your Mac with the on-device speech model (Parakeet, which runs on the Neural Engine), and your journal database never leaves it.
• No AI traffic without a provider you choose. Until you connect an AI provider, nothing is sent anywhere — Ask the book and the Second Brain simply wait.
• A local provider keeps everything on your machine. Connect Ollama or LM Studio and your text is processed locally.
• Cloud is opt-in. Sending anything to a cloud provider is off until you turn it on, and the same gate applies uniformly to journal text, meeting transcripts, chat, and drafts.

Requirements#

• macOS 14.2 or later.
• Apple Silicon (M-series) — the on-device speech model runs on the Neural Engine.
• A few gigabytes of free disk space for the speech model, which downloads once during first-time setup.