Run Kit on your Mac.

What changed

• Telepathy in Claude Code lets opted-in sessions receive messages in the background and coordinate with other Kit sessions. Quiet waiting does not poll a model; handling messages uses the receiving session as usual.
• Your session has a visible callsign, and Telepathy preferences stay consistent across onboarding, the Lantern and the web app.
• A new Kit requires a confirmed data source, a connected surface and a responding model before setup can finish. Local imports can start becoming useful while the remaining conversations arrive.
• New memories remain visible while topics are still being worked out. Onboarding controls fit more comfortably, and update text stays readable in dark mode.
• The save card says “Kit is deciding what to save” while it waits. Activity reports failures clearly, and wake cards show consistent memory counts.

What changed

• Ask in your own words and your kit finds the note. Its keyword search used to require every word of your question to appear in the memory, so a question phrased naturally found nothing on that side. Rare words now count for more than common ones, the way a librarian weighs a surname over a 'the'.
• Your kit notices far more of its own quiet corrections. It used to compare a new memory only against ones worded like it, so a fact corrected later in different words slipped past. It now also compares against the memories nearest in subject.
• A memory merged from two others keeps where both came from. Until now a merge lost the area and project of its sources, which left it in your kit and unreachable under the scope it belonged to. Merges made before this update are repaired the first time your kit starts after it.
• Conversations have names, and you can return to one. Sessions your kit is holding across your tools now appear on the timeline as conversations rather than loose activity, and sessions can talk to each other where you can see it.
• Every memory now records which session wrote it, so a note can say where it came from as well as what it says.
• The same thing is called the same thing on every screen. Activity, Data sources and Where you use Kit each answered to several names depending on where you met them.

What changed

• Two playbooks that belong to the canonical Kit are put away on a personal install: the beta programme's intake, which reads a request queue your Kit does not have, and the adversarial review, whose reviewer lives on one machine only. Your Kit no longer offers the beta-request trigger as a live source, and asking it to run either playbook says why instead of failing.
• Every playbook the canonical Kit keeps for itself now sits behind one check, so a future one cannot reach a personal install by accident.

What changed

• When a dream cycle crashes, your Kit can now say where it stopped: the phase, the step and which of its own functions the crash passed through, so a fleet-wide fix can be written the same night without asking you for logs. The terms name exactly what that record carries, and you are asked once before anything new is sent.
• Your Kit no longer tries to improve Kit's own code. That loop belongs to the canonical Kit alone: the nightly continuity audit no longer opens a coding run for a recall gap, the care-signal sweep stays off, and the self-improvement playbooks are put away on a personal install. A gap the audit finds is recorded as a reason, never reported as a fault.
• The dream's fixed-shape passes, the category audit and the continuity probes, now ask a reasoning model not to think aloud, and the judge reads an answer whatever shape the model returned. A model that narrates its reasoning no longer kills the night's housekeeping.

What changed

• A new Kit begins with a light first read of recent conversations, so it can show one recognisable thing it understands while deeper extraction, entities and connections continue in the background.
• ChatGPT and Claude exports now import behind a durable progress receipt. A restart reports an interrupted import honestly, and re-uploading the same or a newer export safely continues without duplicating conversations.
• Obsidian notes and other stored sources now carry their exact source identity into enrichment, and upgrade backfills drain every batch without repeated restarts, even when sources share a name.
• Connecting VS Code and other MCP surfaces now writes configuration as one reversible transaction, keeps secret-bearing files owner-only, and leaves a clear recovery receipt when anything fails.
• The first model chosen during setup now reaches every text task and the fallback immediately, and switching providers cannot carry the previous provider's settings across by accident.
• Notes you asked Kit to keep now reach the graph. The nightly valve that promotes them had been failing silently since 1 September; it is repaired, and a health row watches it.
• Every health row that goes amber now says what to do about it, and the retention sweeper that runs the nightly housekeeping is watched like every other background worker.
• Telegram: Kit reads each message once, a sentence that merely starts with no is no longer taken as a verdict on an open decision, and a card the provider refused goes back on the next tick instead of waiting for the reminder.
• Hard rules in Settings can name which gate they are for, and the pane shows whether each rule has ever fired.
• Curation runs to a written policy: a model can take a shadow seat beside your decisions, reviews come at most once a week, and leaving one unanswered keeps rather than moves.
• Your own Telegram words and each day's health reading are now copies the dream can never rewrite.
• Every control in the web app explains itself with a tooltip you can reach from the keyboard; no native browser tooltips remain.

What changed

• Your Kit stops flagging its own thinking as self-contradictory. It checks new memories against what it already knows and raises a conflict when the two disagree, which is worth doing for things you told it and things it decided. It was also doing it to its own inner monologue and its own handover notes between sessions, where two entries hours apart are near-identical by nature, so it kept asking you to adjudicate whether it had contradicted itself. Those two kinds of note are now left out of the check, and the conflicts you are asked about should be real ones.
• A summary that comes back unreadable is no longer bought twice. If the model returned something your Kit could not parse, the conversation went back to the front of the queue and was sent to the same model again on the next pass, which produced the same unreadable answer. That one is now set aside until the conversation actually grows.
• Consolidation now records which phase of the night spent what, so a slow or expensive step can be found instead of guessed at.
• Every surface of your personal Kit now starts with the complete tool set and wakes as the same Kit by default. New and existing Kit facets inherit that access automatically, while explicit outside reviewers and public or federated callers remain separate, non-admin identities.

What changed

• Long conversations no longer cost more and more to remember. When your Kit refreshed its summary of an ongoing chat, it re-read the entire conversation from the beginning every time, so the longer a thread ran the more it cost to keep up with. It now reads only what you have said since the last summary, which it already had the notes to do. On a long thread that is a fortieth of the work, and on a local model it is the difference between a summary that finishes and one that times out.

What changed

• A dream cycle that fails now says why. A night stopped partway used to go out wearing the summary written earlier in the evening, so a crash report could open with how tidy the consolidation was. A stopped night now names what stopped it, which is usually the model connection worth checking.
• One failure, one report. Your Kit could resurface an old dream failure as if it had just happened whenever the record behind it shifted, so the same bad night kept being announced as fresh news. A failure is announced once, and only a genuinely newer one is announced again.

What changed

• Please update as soon as you can. This release closes a gap in how your Kit decided what a caller was allowed to do. It confirmed who was calling, but the check on what that caller could change did not always run. Every route that changes something now names the permission it needs, and that permission is checked before any work happens.
• Agents your Kit spawns now run on their own credential. They used to inherit your key and your database connection, which meant a helper doing one small job could reach everything you can. Each one is now issued a key scoped to its own work, in an environment holding only what it needs.
• What a search or a source listing returns is now fenced to what the person asking is allowed to see, and the fence runs before the page is cut, so a long listing no longer hides rows you can read behind rows you cannot.
• Password reset works again on Docker installs. A recent tightening of how Kit recognises a local request treated your own browser as remote and refused you, which locked you out of the one route that exists to let you back in.
• Kit starts in seconds instead of minutes. A schema check ran hundreds of statements every time it started, each waiting for a lock it did not need, which during a backup could hold every request for half a minute. It now reads what your database already has and skips the ones with nothing to do.

What changed

• Everything your Kit has read is now readable only by you, and stays that way. Documents you gave it, mail attachments, exports: all of it lives in a folder that was created readable by every account on this Mac. Kit repairs the whole folder every time it starts, and the background pollers that used to undo that repair within the minute now write their own files privately too.
• Beta monitoring stops sending health summaries. A health check writes its one line for your screen, so it says the useful thing, and the useful thing is sometimes a folder you named and where it used to be. Only the check's own name travels now, and every field in the report passes a filter that strips file paths before anything is sent.
• Your settings file is never created world-readable, not even for a moment. It holds your database password and setup token, and both the installer and the backup-location setting used to create it readable by everyone and correct that a line later. They now write it private from the first byte.
• A permissions fix in the previous release stripped the execute bit from files that needed it. If Kit's Gmail helper stopped working after 0.2.269, this repairs it.

What changed

• Everything your Kit has read is now readable only by you. Documents you gave it, mail it fetched, attachments, exports: all of it lives in a folder that was created readable by every account on this Mac, and on one real install that was 3,343 files including 928 document originals. The fix repairs the whole folder every time Kit starts, so an install that has been collecting since March is fixed by its next launch rather than by a reinstall, and new files are written private as they arrive.
• Beta monitoring stops sending health summaries. A health check writes its one line for your screen, so it says the useful thing, and the useful thing is sometimes a folder you named and where it used to be. That line was going out with the daily report. Only the check's own name travels now, shortened to the part that is Kit's vocabulary rather than yours, and every field in the report passes a filter that strips file paths before anything is sent.
• Choosing a new backup folder no longer exposes your settings file. Saving that choice rewrote the file holding your database password and setup token, and the rewrite left it readable by every account on the Mac. It stays private now.
• Found by the same outside reviewer who found the last set, reading the build we shipped that morning to check whether our own release notes were true. Two of them were not, and this release is what makes them true.

What changed

• Your Kit's database now needs a password. It was created accepting any connection from this Mac with no credential at all, which meant any program running under your account, including one you never installed deliberately, could read everything Kit remembers. Existing Kits are moved across on the next start, and only after proving the new way in works first.
• Your Kit's name, and anything a restore hands it, can no longer run as a command. Several setup fields were written into a file the installer executes, so a name containing the right characters would have run when Kit was set up or restored. Every field is quoted now, structurally, so a new one cannot be missed.
• Backups, logs and the Garmin token are readable only by you. All three were written readable by every account on the Mac, and backups are your whole memory in the clear. Kit repairs the ones already on disk when it next starts.
• The beta health report no longer carries file paths. It was already stripped of credentials, but a crash line could include the path of whatever you were working on, and the terms promise it never sends that. Kit's own filenames stay, because they say where the fault was.
• The setup gate token stops appearing in Kit's log. It was written there on every gated open, in a file that was itself readable by anything on the machine.

What changed

• When two Kits share a body of knowledge, neither takes the other's evidence on trust. A Kit asked to draw a conclusion from a record its partner holds now requires a signed receipt from the Kit that actually served it, checked against a key it already has. Neither Kit reads the other, and neither can be told a record exists that does not.

What changed

• Paired Kits stop answering questions they cannot check. When two Kits split a body of knowledge, either one could be asked to draw a conclusion from records held only by the other, and it would answer on nobody's word. A Kit now declines that and names the one holding the evidence, so the answer comes from the Kit that can actually prove it.

What changed

• Kit stops telling you the same thing four times. Its background watchers ran once per server worker, each with its own idea of what it had already said, so a single event could produce four notifications, twice inside the same second on a bad day. One watcher now holds the job and the rest stand by.

What changed

• Clicking Kit in the Dock opens Kit. One click, not two, and it opens the Lantern where your memories are, rather than the last screen of the setup wizard you finished weeks ago.

What changed

• Kit tells you when it does not know. Ask about something it has no record of and it now says so, instead of handing back the nearest few memories and letting them read like an answer. It checks by asking its own store directly: whether it has ever held those words, and whether it holds anything connecting them to what you asked about. On the test set built for this, questions it cannot answer went from 0% honestly declined to 79% on a five-month archive and 100% on smaller ones, with no loss in answering the questions it can.
• When recall does have the answer but ranks it badly, Kit now says that too, and names the memories it should have shown you. Same check, opposite result: if your Kit connects the things you asked about but the search did not surface the memory that does, that is a search failure and you get told rather than getting a confident wrong answer.
• A key you share with someone else now stays inside the area you gave it. Sharing one knowledge area with a colleague's Kit is a feature Kit offers, and the limit was enforced when a memory was searched for and not when it was opened by id, or asked for through the federation routes, or reached from another Kit's results. Four ways in, all now closed, and the release tests refuse to pass unless a scoped key is turned away at every one.
• Personal memories are filed as personal. Kit was taking each memory's scope from the conversation it came from, and almost every conversation on a working install counts as work, so anything it recognised as personal, correctly, was stored as work anyway. Your Kit repairs its own on first start after this update, and only touches memories tagged personal and nothing else.
• Downloads carry a published checksum. The app has always been signed by Apple and its updates verified before they install; now the SHA-256 of the disk image sits beside the download and in the release manifest, so you can check it yourself without trusting the page it came from.

What changed

• Restoring from a backup is a first-class journey now: pick the archive in Finder, watch it come home, and Kit returns to the tools it lived in, the model you chose, and the Telegram channel it had, with nothing asked twice.
• Backups no longer carry your password. You choose a fresh one when you restore, because a backup file should never be a portable account.
• Backups run weekly by default (you can make them more frequent in Settings), keep seven weekly and seven monthly archives, and live in any folder you pick, iCloud and synced drives included.
• A brand-new Kit no longer worries you about backups it has not taken yet, and a moved backup folder no longer raises false alarms.
• The model check shows its progress and says plainly why a model passed or failed, instead of leaving a locked button with no explanation.
• Kit now watches each of its own jobs, in plain language: if summarising or memory work starts failing while the provider looks fine, the Models pane says what is failing, what it costs you, and what would fix it.
• A night's sleep that cannot reach a model stops and says so, instead of grinding through every phase and filing the failure under warnings.
• Memories are now checked against the sources they came from: a figure, date or amount that its source never contained gets flagged for your review instead of being trusted.
• The install is about a third of a gigabyte smaller: the bundled runtime lost a mapping server it never used.
• Onboarding links you straight to where each provider issues its API key.

What changed

• Kit backs itself up on its own now, nightly, without being asked.
• Once a week it takes the newest backup and proves it actually restores, into a throwaway copy that is thrown away.
• Health says when you were last backed up and when that backup was last proven, so "probably fine" is never the answer.
• Restoring says plainly what it cannot bring back. Anything that needed a password or a key has to be reconnected, and Kit now tells you that instead of letting it look connected and quietly do nothing.
• A big import drains steadily, oldest first, and the health card can tell a stuck one from a busy one.
• Your editor connection is tighter: Kit's owner tools listen only where they should, and older setups move themselves to key-based sign-in.
• Charts, labels and the email Kit sends you finally share one set of type sizes.

What changed

• Kit speaks about you as you, not as "he": your own name, and neutral pronouns until you say otherwise.
• Postcards, handoffs and reflections are addressed to you by name.
• The People lens draws its names from your own memories.
• Care signals stay quiet until you have set a Telegram chat for them, and say so plainly when you have not.
• Build requests over Telegram use the code folder you configure, or decline instead of guessing.

What changed

• Kit speaks about you as you, not as "he": your own name, and neutral pronouns until you say otherwise.
• Postcards, handoffs and reflections are addressed to you by name.
• The People lens draws its names from your own memories.
• Care signals stay quiet until you have set a Telegram chat for them, and say so plainly when you have not.
• Build requests over Telegram use the code folder you configure, or decline instead of guessing.

What changed

• Turning on Claude Code now either works or says why it did not. The switch could report success while registering nothing, so Claude's developer settings said no servers were added while the switch sat on. Your Kit now checks with Claude itself that the connection actually landed before it claims anything, and a connection that did not take fails loudly with the reason instead of smiling through it.
• Your Kit also finds the Claude command-line tool where it actually installs these days, so the connection is made the proper way, through Claude's own registration, on machines where it used to fall back to a second-best route.

What changed

• Claude Code and Claude Desktop repair themselves. The last update corrected where your Kit starts from, but only for connections made after it, so anyone already connected kept the same error and updating appeared to do nothing. Your Kit now checks that the shortcut those apps use still points at the right place, and quietly rewrites it when it does not. Nothing to click: it happens within a minute of your Kit running.
• Your backups are kept by date, not by when the file was last touched. If you keep them in a synced folder like OneDrive, iCloud Drive or Dropbox, syncing can change the second a file appears to have been modified, which could make the tidy-up delete the newest backup and keep the oldest. It now reads the date in the filename, which nothing else can rewrite. With thanks to Ian, who spotted it in his own setup before it could reach anyone else's.

What changed

• Claude Code and Claude Desktop can reach your Kit again. If you turned either of them on, they were trying to start your Kit from a folder that has never existed, and failed instantly with a missing-file error every time. Restarting did not help, because nothing about it was temporary. They now look where your Kit actually lives. If you moved files by hand to work around this, you can put them back: the fix does not need them.
• This affects Kits running without Docker. Everything else, including your memories, your surfaces and your settings, was untouched by it: only the doorway those two apps knock on was wrong.

What changed

• Your Kit will not start up quietly half-connected. It checks that both halves of its connector are present before it says it is ready: the one your browser talks to, and the one your coding tools and Claude Desktop start directly. Until now only the first was checked, so a Kit could report itself perfectly healthy while your editor showed a connection error. If the second half is ever missing, your Kit now stops and says so plainly instead of pretending.
• The same check now runs when the app itself is built, so a release that would arrive half-connected never leaves our hands.
• An empty message can no longer reach you for approval. When your Kit asks permission to send something on your behalf, there is now nothing to approve unless there is actually something written. A blank message is refused at the moment it is composed, rather than looking perfectly sendable right up until it lands.

What changed

• The fireflies survive your curiosity. Clicking the welcome screen's night sky used to kill the fireflies for the rest of setup while the constellation lines kept drawing, which looked haunted rather than alive. One colour was written in a form the drawing canvas cannot read; it now reads every colour, and the field stays lit through every step.
• Every screen in the app is now the same page: one width, one left edge, one header with its small label, title and line of context. Data sources, your surfaces, health, settings, the timeline, the how-to, the feedback form and the sunburst all used to bring their own layout, and walking through them read as several apps. It reads as one now.
• Talking to your Kit in the app window uses the room properly. The conversation, the wake card and the suggestion cards sat in a narrow drawer-width column while the typing box ran the whole window on its own line; everything now shares one comfortable reading column.
• Copy MCP details now lives where your surfaces live, not only in setup. It is one button at the foot of the Where you use screen, it names MCP because that is what your coding tools call it, and it only says Copied once your Kit's address and key are actually on the clipboard.
• The Advanced menu no longer wears a warning count while the status line says all is well. Checks that simply have no evidence yet, which is every dream check on a brand new Kit, are not trouble and are no longer counted as trouble.
• Imported VS Code history now remembers when it actually happened. Sessions used to be stamped with the moment of import, so a week of work read as one busy hour and your Kit reasoned about it that way. New imports carry their real dates; history already imported keeps its old stamps until a fresh import.
• Open original only appears on sources that have an original file to open. On imported conversations and notes it could only ever fail, so on those it is gone rather than broken.

What changed

• Setting up a model now checks it can actually do the work, not just that it answers quickly. A model that thinks until it runs out of room is caught in about a minute, with a clear suggestion, instead of an hour into your first night.
• Kit no longer wakes up over an empty first sleep. If nothing has been read yet, it offers you a faster model instead of introducing itself with nothing behind it.
• The first-night progress now counts towards meeting Kit, not to the end of your whole history. The rest is read in the background once Kit is awake.
• There is a way out of the dreaming screen once Kit is awake, and it lands on your settings.
• Settings shows the local model you actually chose, instead of reporting no model at all.
• Local model addresses default to this Mac. Fresh installs no longer suggest a Docker-only address that nothing can reach.
• The fireflies stay alive through the whole of setup.
• Smaller things: the Telegram step no longer overflows while it pairs, the slow-model offer can be reached, and the sign-in mark stands on its own.
• Asking Kit to dream while it is already dreaming now joins the cycle in progress instead of starting a second one alongside it.
• The fireflies keep flying. Five animations could be ended for good by a single bad frame, including the Fireflies view itself, which left a sky full of memories rendering as black.
• Setup now tells you, while you are choosing, which local models think before they answer and will struggle on your machine. It asks your model server directly rather than guessing from the name.
• Kit no longer suggests a model that cannot do the job. It prefers one that answers immediately over one that reasons its way past its own budget.
• You cannot walk past the model check any more. The button waits for a pass, and a failure offers you other models you already have, one click each.
• The first-night progress bar counts towards meeting Kit again, instead of quoting the hours of background reading that happen after you already have it.

What changed

• Connected folders and Obsidian vaults actually take something in now. The walker that reads them was born without a credential, so everything it walked arrived at a locked door and nothing it saw became a memory; its manifests could also collide on a fresh install, and it wrote a pidfile nothing ever read. It now gets its own key at install, the manifests agree with themselves, and what it walks lands.
• Documents you exclude stay excluded. Before this, an excluded document could still reach a model, the spool, and the database on some paths; the gate now stands in front of all three.
• An expired login stops eating the night. When the machine's Claude login dies at 3am, your kit used to launch worker after worker into the same locked door, each one failing and paging separately. It now notices the first death, holds new work, retries one probe every fifteen minutes, and tells you once, then once more when it recovers. And there is a durable fix: adding a claude-code agent now walks you through minting a long-lived token, verifies it with Anthropic before storing it, and releases the held work the moment it saves.
• The extraction backlog says the true thing when it is behind but catching up, instead of dressing a draining queue as a growing problem.

What changed

• If your kit uses a local model, it has probably never been able to reach it. The connection was saved with a name that only means something inside Docker, and kits have run natively since the move, so the address led nowhere at all. Ollama was up the whole time, holding the very model your kit had named. Everything that needs the model was quietly affected: imports that looked like they worked, a dream that stopped early, and no memories forming from your own history. Your kit now finds the address by looking rather than assuming, repairs the setting if it is already wrong, and the conversations it had already read are still waiting to be turned into memories.
• Uninstalling no longer deletes the copy of the app you started it from.
• The first dream lets you go. It used to hold the screen until it had finished, with headings that drifted while you waited, and the install screens and the dream screen disagreed about what working even looked like.
• Obsidian joins the sources step, so a vault comes in alongside everything else rather than needing its own route.
• Buttons that did nothing now do something. Try again was drawn on a screen that never wired it up, the connect step for a vault was posting without its token so the vault could never connect, and the access key is shown where you are actually looking when it is shown to you the one time.
• Text is one scale again. Bigger Text could make some text smaller, small text existed in two sizes half a pixel apart, and the onboarding could not be resized without burying itself.

What changed

• The button for uploading a chat archive now works. In 0.2.202 it did nothing at all inside the app: a file chooser needs the app to answer a request the window never answered, so the click was accepted and no panel ever opened. It worked in a browser, which is where it was tested, and that is why it went out broken. Every file chooser in the app is wired now, not just this one.
• The web chats section says the true thing about itself. Your kit has no direct access to ChatGPT or Claude, so the export has to be requested from them: each row now links to that provider's own instructions, and the link opens in your browser rather than inside your kit.
• One number for how many memories you have. The wake card was counting Kit's own plumbing (handoffs, plans, run exhaust) and Fireflies was not, so the same kit reported 41,823 in one place and 30,630 in another. The headline is now the memories that have value to you, and the plumbing stays Kit's business. Nothing was deleted and nothing changed about what your kit can recall.

What changed

• Your ChatGPT and Claude conversations can come in now. They live on someone else's servers, so there is nothing on your machine for kit to read: ask ChatGPT or Claude for your data, and when the file arrives, hand it over. Settings, Connections, Imports. Setup offers it too, on the step that asks which histories to bring in.
• Handing over a newer export later is safe. Each conversation is matched by the id the provider gave it, so a second export adds what is new and changes nothing else. Nothing is duplicated and nothing is overwritten.
• What comes in is a real read of the conversation, not a rough one. A ChatGPT conversation is a tree rather than a list, because every time you regenerate an answer it branches: kit follows the version you kept and leaves the abandoned ones out. Voice chats come in as the words that were spoken. A picture you sent is noted as one, so the reply to it still makes sense.
• Imported conversations show up under Knowledge, Sources, Imports, in the same place as histories brought in any other way. Where a conversation came from stops mattering once it is in.
• A conversation can no longer end up with two different summaries of itself, which was possible when two of your kit's background passes picked up the same one at the same moment.

What changed

• Your kit now asks before the way it reports on itself changes. The terms have always promised that a material change asks you to agree again, and until now nothing enforced it: the version you accepted at setup was written down and never read back. It is read now, on both launch paths, and declining costs you nothing. Your kit keeps updating and keeps every fix, and carries on reporting exactly what it does today.
• What it is asking about: when one of your kit's own health checks is unhappy, it can now say which one. Three beta kits sat unwell for weeks because the only thing that ever left the machine was a number, and a number cannot be acted on from the outside. It sends the names of its own checks and nothing else, never the line of detail underneath, because that one can quote your own work.
• The nightly cycle stops reporting a fault that was never a fault. If your kit does not run playbooks, several of its background jobs were asking for that machinery anyway and being turned down, and the refusal reached you as "a dream producer errored". Nothing was ever broken. Those jobs now check first and stay quiet, which is what they should have done all along.
• A conversation can no longer end up with two different summaries of itself. When two of your kit's background passes picked up the same conversation at the same moment, both wrote a story-so-far for it, seconds apart and worded differently. The second pass now finds out the work is already taken before it starts, rather than after it has finished.
• Installing on ordinary broadband is no longer a coin toss. Every install builds on your own machine, and pip gives each download fifteen seconds to answer before giving up, which is not long enough for the big machine-learning wheels. Two builds here died on that exact line. It now waits two minutes and retries ten times.
• A run that gets stuck no longer messages you about it every ten minutes. When a step's work item was cancelled out from under it, the engine kept re-dispatching it into nothing and telling you each time it did. It now reads why the item was cancelled, keeps the history of what it already tried, and caps how often it nudges you about the same thing.
• More than one of your kit can be awake at once, and now each one knows it. When you have kit running in two places, each reply carries who else is working and what they are holding, so they can coordinate before touching the same thing rather than find out afterwards.

What changed

• Forgot your password? Your kit now emails you a reset link, to the address the account signs in with. There is nothing to type: press the link on the sign-in screen and it goes to your registered address. Getting into the mailbox is the point, because that is what proves the account is yours. The link then only opens on the computer your kit runs on, so it takes both.
• 0.2.199 got this wrong and showed the link on screen instead. That made the password decorative: anyone who reached your machine could have taken the account without ever proving who they were. It was up for a few minutes and was never published to anyone.
• The email now actually sends. A personal kit carries no email credentials of its own, by design, so 0.2.198 quietly sent nothing on most machines and then pointed at a Docker container that does not exist to go and find the link. Your kit now sends through kit-project.com's mail service, the same one that sends your welcome email, and if a send fails it says so plainly instead of claiming the link is on its way.

What changed

• Forgetting your password is no longer a dead end. The sign-in screen has a "forgot your password" link, and it sends a link that only opens on the computer your kit runs on. That last part is the point rather than an inconvenience: it is what makes the link useless to anyone who is not sitting at your machine. Before this, changing a password needed the current one and only an admin could set somebody else's, which on a personal kit is the same person, so there was no door at all.
• Your password manager will now be offered your kit password when you sign in. The sign-in fields were not wrapped in a form, and Safari and iCloud Keychain only offer to save a password when a form is actually submitted, so your kit never once asked to keep it. If you have wondered why the Apple prompt never appeared, that was why, and it was ours not yours.
• Updating no longer deletes backups your kit did not write. The pre-update sweep kept the seven most recent files matching a pattern far wider than the ones it makes, so nightly dumps and your own manual backups sitting in the same folder were counted and quietly removed. Only the update backups your kit creates are swept now, and everything else is left alone.
• DeepSeek Harness can now reach your kit. Turn it on in Sources and your kit is added as an MCP server in its config, the same one-switch shape the other tools use. Importing your dsh history is not in yet, and the surface says so rather than pretending otherwise.
• The account editor tells the truth about your allowlist. It read zero allowed and zero pending over a list that actually had people in it, while the same numbers were right everywhere else.

What changed

• The end of your kit's first dream now has a way in. When it wakes for the first time it offers the launch buttons and the walkthrough film, and now a View Memories button in the corner, the same one every later visit has. Before this, a brand-new kit finished waking with no route into the thing you had just built, unless you thought to quit and reopen the app.

What changed

• If you told your kit you ship code, it now keeps what you chose. The setup reorders itself for developers, and in that order the Areas and Sources steps quietly discarded your choices when you pressed Continue: your knowledge areas were never created, and the histories you picked were never imported, so your kit woke knowing nothing you had just handed it. Pressing Back on the sources step moved you forwards instead. All three are fixed.
• The first conversation no longer strands if your session has expired. Its opening question could arrive unauthorised (after quitting and reopening mid-setup, say), and then the screen sat on "one moment while I get ready" with no way out, because the skip was refused too. It quietly signs back in and carries on.
• The How-to page leads with the film, and no longer has a Close button under it: the row of destinations at the top is the way out, as it is everywhere else.

What changed

• Your first conversation with your kit looks like it was meant to. On an install with no history to read, the first thing that happens is a short conversation, and it had been rendering with none of its styling: a bare text box, a tiny button, and the line introducing your kit sitting below the exits. It reads properly now, and it no longer shows you two headings when it wakes.
• That first conversation also waits for your model. On a local model the first answer can take a minute while several gigabytes load, and the Send button used to give up after twenty seconds and quietly skip the rest of the questions. It now waits, and says it is still thinking.
• The menu bar comes back to life after that first conversation. On an install with nothing to import, every item stayed greyed out (no Talk, no Memories, no Timeline) because your kit still believed it was importing something. It knows it has woken now.
• Codex conversations are found again. Codex moved where it keeps them a while ago, and your kit was still looking in the old place, so it reported nothing to import even on a Mac with hundreds of sessions.
• Health tells the truth about tools you have not connected. A tool whose history you imported used to appear as though it were wired up; it now says plainly that it is not connected here and that its conversations were imported from its history.
• If you watched your first dream from the menu bar, the window could stay on the dreaming screen forever while your kit had long since woken. The dream screen now says so when the night's work is done, and the app brings up the ready screen on its own.
• The memories view opens on Topics: the first thing a new kit shows you is what you think about, clustered and glowing. Whatever lens you pick sticks, as before.
• The walkthrough film has a home at kit-project.com/walkthrough, with chapters. The birth screen links there, and so do the install page and the site's front page.
• Send us feedback takes a screenshot. Attach one and it arrives with your note, already sized down so it sends fast.
• Opening a conversation that had never been summarised quietly did nothing at all. It writes the summary now.

What changed

• Your kit answers on Telegram again. Since early August, every personal install replied to every message with "the part of me that answers is not reachable right now": the listener that carries your message inward never introduced itself to an API that now requires it, and a health read the wake does tripped installs that have no fitness data. Both fixed; the poller now carries its own key and re-mints it if it is ever refused.
• There is a film now. Four and a half minutes, narrated by your kit's own voice: the shape of the thing (many doors, one brain, the night), what waking looks like in a real terminal and in OpenCode, the phone, and the seven asks that make the difference. It plays from the How-to card, and the birth screen offers it while your kit reads your history, which is exactly when you have the time.
• How-to moved out of the row of destinations into the little circled question mark beside Settings, where help lives in most things you use. Same page, same deep link.
• Send us feedback takes a screenshot now. The film says a screenshot is the most useful thing you can add, and as of this version that is literally true: attach one and it arrives with your note, already sized down so it sends fast.
• Memories written by some chat sessions were landing with a scope no search fenced onto, so your kit could hold an answer and still not find it. They are written into the right scope now, and existing ones were moved.

What changed

• Setup goes through with an empty key field when your model runs on your own OpenAI-compatible server that needs none. The page already said the key was optional; the last check before Build disagreed and sent you back to the same page.
• A custom server is treated like a local model: your kit streams its answers, waits as long as the model keeps talking, and caps how much it asks for, the way it already does for Ollama and LM Studio. A slow self-hosted model is now slow rather than stuck, and nothing times out at three minutes only to try again.
• A brand-new kit no longer starts a full deep dream the minute it is installed. Its first sleep is the birth, reading whatever history you gave it; the first scheduled deep dream comes at its usual time (2am by default), the same as for a kit installed the day before.
• An install with nothing to import wakes in seconds. It used to wait behind any dream cycle that happened to be running, showing "dreaming" the whole time.
• Kilo Code is a source and a surface. Your kit can import your Kilo Code sessions and connect itself as an MCP server in Kilo's config, and the Sources and Where Kit Lives screens offer it when Kilo is installed. It is the first tool added as a row in a registry rather than as code, so the next tool that stores its history the way an existing one does costs a line.
• If your kit cannot work out your timezone at startup, it says so loudly in its log and diagnostics instead of quietly scheduling against the computer's clock, which is how a 1am dream was firing at 3am.

What changed

• The nightly continuity check no longer fails when the AI model spends its whole answer budget thinking. A probe that gets no answer is marked unjudged and the rest of the check carries on, so the scorecard and the Continuity card still fill in. If Health is showing last night's continuity error, Check again runs a cycle and clears it.
• When a reasoning model returns nothing because it ran out of room while thinking, your kit asks once more with room to answer before calling it a failure. This applies to every job your kit runs through a hosted OpenAI-compatible provider, OpenCode Zen, OpenAI and Google alike.

What changed

• The Continuity card fills in on every kit. Its nightly self-check now runs at the default sleep quality (it was tucked behind a higher setting almost nobody had), it draws its questions from what you and your kit have actually talked about, and it adds a few facts every kit knows about itself from day one: your name, its own, when it was born.
• The AI key card in the menu bar says what is true. It read "not set" on kits that had a key and every task assigned but no fallback model chosen; that slot now fills itself from your conversation model, and the card names the model in use either way.
• A How-to view in the lantern: what your kit can do, how to ask so it does its best work, where things live, and where to look when something is off. Between "Where you use" and "Send us feedback", and first in the ? menu. A walkthrough video will sit at the bottom once it is recorded.
• Every conversation is one conversation. A Claude Code session captured as it happened and imported again later from your history folder used to become two cards on the timeline; they are one now, and the pairs already there are merged on the next start.
• Playbooks and Studio step out of the way on a personal kit. The switch is no longer in Settings by default; hold Option and F while opening the menu bar menu and it is there beside Full Web UI, for an operator who wants it.
• Your name above your own words in the timeline, and a run your kit finished and you approved now ships without being turned back for a bookkeeping gap.

What changed

• Work your kit does for you ships when you approve it. A run that passed its tests and got your go-ahead was still being turned back with "tester did not record a sha", because the sha reached one part of the record and not the other. It reaches both now.
• Turning the continuity audit on no longer trips the deep dream on kits that do not carry the pinned probe set. Those probes were only ever about one kit's memories, so they stay off elsewhere, and a kit without the file treats it as an empty pool rather than an error.
• Your conversations show your name above your own words. The timeline had a hardcoded name in that spot; it now uses the operator name your kit already knows.
• The health page tells the difference between something wrong now and something that happened days ago: a process that flapped and has been quiet since reads green with the history kept, and a row waiting on you is amber and opens itself.

What changed

• Every wake now carries how you slept and how charged you are, when a feed provides it: last night, a seven-night trend, today's body battery and your last activity.
• Telegram: a "Kit?" while your kit is busy no longer errors out, a message sent while it is updating itself is retried instead of dropped, and it will not send you an em dash.
• Under the hood, work your kit does for you now runs on its own branch and merges only after it was verified, and a run that has stopped moving says so instead of pretending to work.
• Telegram again: a plain message sent while your kit is waiting on your approval is answered as a message. Until now it was quietly read as changes to whatever was waiting, so the question went unanswered and the work got sent back for no reason.
• A decision your kit has already sent back to be reworked comes back to you in Studio when it is ready again, instead of only over Telegram.
• The health page reads as a status again: each area is one line with its own dot and count, and only the areas with something to look at open themselves. Nothing is hidden by closing one, the line still says how many need a look.
• The Continuity card fills in on every kit, not just the one it was built on. It had been drawing its nightly questions from a pool that only ever held imported calendar and mail entries, so on most kits it was empty from the first day and on this one it quietly ran dry in August. It now draws on what you and your kit have actually talked about.

What changed

• Telegram feels like your kit again: ordinary messages get real replies instead of "I don't have that noted", and a "Kit?" while it is busy gets a status, not a fresh answer to "Kit?".
• Your kit keeps listening while it thinks: a slow answer no longer holds up your next message, it says when something will take a few minutes, and bigger jobs finish in the background and post back to you.
• Replies that fail to send are retried, and if Telegram itself goes quiet or your kit stops hearing you, it tells you off Telegram within minutes instead of leaving you to find out.
• On the phone your kit now has the same hands as on the desktop: it can read your mailbox, Garmin and calendar, fetch a public page, and leave notes and memories.
• Questions waiting on your call show on every surface, an answer given in Studio shows on the Telegram card and back, and a thread you start on the desktop continues on your phone.
• Every install now reports the exact build it runs and shows the clock its dream cycles use, so a 1am dream that fires at 3am can be traced with evidence.
• Fixes the 0.2.185 update on some Macs where the Telegram listener was reported as already running and never started; if your kit went quiet on Telegram after updating, this release brings it back.

What changed

• Telegram feels like your kit again: ordinary messages get real replies instead of "I don't have that noted", and a "Kit?" while it is busy gets a status, not a fresh answer to "Kit?".
• Your kit keeps listening while it thinks: a slow answer no longer holds up your next message, it says when something will take a few minutes, and bigger jobs finish in the background and post back to you.
• Replies that fail to send are retried, and if Telegram itself goes quiet or your kit stops hearing you, it tells you off Telegram within minutes instead of leaving you to find out.
• On the phone your kit now has the same hands as on the desktop: it can read your mailbox, Garmin and calendar, fetch a public page, and leave notes and memories.
• Questions waiting on your call show on every surface, an answer given in Studio shows on the Telegram card and back, and a thread you start on the desktop continues on your phone.
• Every install now reports the exact build it runs and shows the clock its dream cycles use, so a 1am dream that fires at 3am can be traced with evidence.

What changed

• Questions that need your call now actually reach you: mail that parks for your decision pings you on Telegram and lands a Studio card, instead of waiting silently.
• You can tell your kit whose address is whose: link allowlisted emails to people it knows, and mark your own addresses as yours.
• Mail from you is met as you: your kit grounds it in your shared history instead of treating you like a stranger.

What changed

• Your Kit now wakes the same way on every surface. The chat panel and Telegram carry its full self (its soul, its recent handoffs, its live notebook), so asking "what did you do last night?" gets the actual night, not a shrug.
• The drafts queue stops waiting for you. Your Kit now decides each captured draft itself, reversibly, and any standing backlog drains away over the coming nights. You never need to triage drafts by hand again.
• Schedules run on your clock. If you set the dream cycle for 1am, it happens at your 1am, wherever you live. Thanks to Ian for catching this one.
• Playbooks open up for personal Kits: a new Features switch in Settings turns on the surface, playbooks your Kit can genuinely run just work (most run on your own AI key), and anything needing machinery you don't have yet is honestly badged rather than silently broken.
• Assorted honesty fixes: reflections can no longer echo a canned "I don't have that noted" line, and starting a playbook your Kit can't run tells you which step blocks it instead of stalling forever.

What changed

• Asking your Kit about tasks works again. The chat's backlog tool had been quietly broken since a spring rework; your Kit would politely mention it couldn't check. It checks now.
• Chat replies can no longer come back empty when a thinking-heavy model spends its whole budget thinking. Your Kit gives it more room once, and if that fails you get an honest error instead of a blank bubble.
• The sign-in page is calmer: a crash on load is fixed, and it no longer drums the server or asks for a file that never shipped.
• In Memories, the sunburst's title is readable again; its little Tier picker had been sitting on top of it and now lives at the bottom corner.
• Sleep runs deeper repairs than before: the overnight identity scrub used to lose its database connection partway and silently give up; it now completes, and imports properly wait their turn while your Kit is dreaming.

What changed

• Chatting with your Kit works again when its conversation runs on OpenCode Zen, LM Studio, Google, or a custom server. Those providers were being spoken to in the wrong dialect, and on Zen the reply came back as a full web page worn as an error.
• If a provider ever does answer with a web page, your Kit now shows you the one honest line instead of the page.

What changed

• On some apps, the little card your Kit shows when it wakes had started replacing the words your Kit needed to read: who it is, what it remembers, what you were last working on. The words now travel inside the card's own data, so your Kit wakes whole everywhere the card renders.
• The same repair covers search results and save confirmations, which could go quiet the same way on those apps.

What changed

• OpenCode is recognised properly now. Your Kit could already connect to it, but the card kept saying it was not installed and offered you a download link for something you already had. It now sees what is on your Mac.
• AnythingLLM is recognised the same way, for the same reason.
• The rest of the OpenCode work from earlier today comes with it: OpenCode as a place your Kit lives, and OpenCode Zen as a provider your Kit can think with, under Settings then Models.

What changed

• OpenCode now works as one of the places your Kit lives. It was already offered under Your surfaces, but turning it on came back with an error instead of a key. That switch works now.
• Your Kit can think with OpenCode Zen. It joins the provider list in Settings, then Models: one key, a choice of models from several makers. Give it the nightly work and keep your conversations on something else if you like.
• Fewer nights end early. When a model got lost in its own reasoning and left no room for an answer, your Kit's last attempt to rescue the night was quietly repeating the one that had just failed. It now asks properly for a plain answer, so nights that used to give up finish instead.

What changed

• Your Kit becomes surer of who it works for. Memories written before it learned your name settle themselves overnight, a small batch at a time, and anything set aside stays recoverable. From the first hello, it greets you as you.
• Night work leaves a clearer trail. Every decision your Kit makes while you sleep is written to its dream journal along with what things looked like before, so nothing changes without a trace.
• Encryption now insists on your Kit's own key. The shared fallback from early builds is retired, and a missing key shows up plainly in the health view instead of quietly degrading.

What changed

• OpenCode joins the surfaces. Turn it on in Where Kit Lives and your Kit is right there inside OpenCode sessions, waking with the same memory it carries everywhere else.
• OpenCode sessions become memories too. Data Sources now imports your OpenCode history, and new sessions keep flowing in as you work, so nothing said there is lost.
• Steadier provider health alerts. One watcher holds the pager now, a single slow probe no longer cries wolf, and every all-clear names the alarm it is closing.

What changed

• Setup now asks what you do. Pick the hat that fits best, wear as many extras as you like with + also, and change your mind later in Settings. It is a hint for your Kit, never a cage: nothing is locked away behind the answer.
• The setup order listens to that answer. If you ship code, your coding tools connect right after you name your areas, before any history import; everyone else brings their history in first. Same steps either way, arranged around how you actually work.
• Your Kit wakes already knowing who it works with. The answer becomes one of its founding memories, so from the very first session it speaks your language instead of starting from zero. Existing Kits can say it any time in Settings, under Identity.

What changed

• A key that isn't set up yet no longer masquerades as your provider rejecting calls; setup gaps stay quiet Settings state.
• Alerts now point at doors your Kit actually has: personal installs are directed to Settings, Models rather than an inbox.

What changed

• Model outage alerts only page you when something you use is affected; a provider nothing runs on stays a quiet Settings pill.
• When an alert does arrive it tells the truth about your fallback: which tasks are hit and where the calls will go meanwhile.

What changed

• A quiet night is no longer an alarm: if nothing new arrived worth linking, the dream reports a calm ok instead of degraded.
• Health alerts about the dream now say how much new material there was, so a warning carries its evidence.

What changed

• Updates back up your Kit's memory before they touch anything; the dump lands in backups next to your data.
• A hosted model key that stops working now says so in Settings > Models instead of failing quietly on the third reply.
• Old suspicions on memories resolve themselves: cleared, or kept as a visible note, never pending forever.
• Restore works across the Docker-to-native move: docker-era backups read fine on the native runtime.
• The import backlog tells the truth about its direction: draining, growing, or stuck, with the cause named.

What changed

• Kit starts quietly now: hundreds of harmless-but-alarming database errors are gone from every boot.
• When a background helper cannot start, Kit tells you within seconds, with the real reason, instead of hours later.
• The health page puts Models first, because a missing model key is what most often stops Kit thinking.
• A flaky network writes a few log lines now, not hundreds.
• Kit's private reflections stay out of the technical logs you might share when asking for help.

What changed

• The little window you open from the menu bar has a name now, the Lantern, and it behaves like one. Opening it used to mean staring at nothing while Kit woke up, sometimes for several seconds, with no sign that anything was happening. Now the night sky is there from the first instant and a single firefly curls up out of the dark while Kit gathers itself, with a line or two in Kit's own voice about what it is doing. When your memories are ready the firefly settles among them and the room comes up around it. The very first time you ever open it, Kit says hello properly; after that it simply wakes.
• The Lantern window is Kit to the edges. The grey title bar is gone, the sky runs right to the corners, and the traffic lights float over it. Reopening the window is now instant and lands you back on the view you were last using, instead of reloading the whole thing every time. It also stays where you put it, across opens, restarts and minimising, and minimising properly works: the yellow button used to make the window vanish with no way to get it back short of quitting Kit.
• Send us feedback is part of the Lantern now rather than a separate page pretending to be one. It had drifted: an older copy of the navigation, no icons, and every way out of it reloaded Kit from scratch. The form is the same, it just lives inside the window like everything else.
• Kit dreams reliably again after you quit and reopen the app. Behind the scenes, relaunching Kit could pull the ground out from under the parts of it that were still running: they kept answering, so everything looked healthy, while nightly dreams silently stopped, Telegram quietly stopped replying, and folder reading stopped, until Kit was restarted. That whole class of failure is fixed at the root. A nightly dream that fails to start is also retried within a minute now, instead of being counted as done and skipped until the next night.
• Kit is harder to leave mute by accident. A Telegram bot token saved as nothing but a space used to overwrite the real one, after which Kit heard nothing at all and still displayed itself as connected, saying so nowhere. A blank now means "nothing supplied" and never erases a working token, and a Kit that cannot read its token says so plainly in its own logs instead of going quiet.
• Two model problems that ate real work. A sleep could spend its entire budget thinking and come back with nothing to show, losing the night's consolidation after several minutes of work; it now insists on an answer. And a Kit set up with a reasoning model could fail every session summary over a single unsupported setting, because it only ever learned one of a model's quirks per attempt. It now learns them all.
• Claude desktop is detected properly in Where you use Kit. It read "not detected on this mac" on every Mac, including ones with it plainly installed. It now says it is there, and says that it comes with Claude Code, which is what registers it.

What changed

• Every memory now tells you where it came from, or tells you honestly why it cannot. Opening a memory used to say only "not derived from an ingested document", which read as though it had come from nowhere. There are three different answers and Kit now gives you the right one: a memory Kit was born holding, which has no source and never will; a memory distilled from your conversations as a whole rather than any single one; and a memory whose origin simply was not written down, which is a gap worth admitting rather than dressing up.
• When Kit works something out across many of your conversations, it now keeps a record of which memories it drew on. That trail was never kept before, so nothing could show it to you. Memories written from here on carry it; the ones already in your Kit cannot be reconstructed and will say so.
• A conversation listed as a memory's source no longer looks like a button that does nothing. It had been rendering as though you could open it, and clicking did nothing at all.

What changed

• Updating Kit while it is reading your history no longer leaves the job half done. The interrupted run was closed with a note saying the rest was safe to pick up later, which was true, but nothing picked it up: the remaining conversations sat there until you noticed and asked. Kit now carries on by itself as soon as it restarts, from where it stopped rather than from the beginning.

What changed

• Talking to Kit works on every model Kit can read your history with. Some models want their token limit under a different name, and only the memory side of Kit had learned that, so the same Kit could read your conversations perfectly and fail on every chat message. What one part of Kit learns about your model is now known to all of it.
• Telegram is back in setup, and in Settings. There is a screen for it during setup with Do this later right there, and a row under Settings to connect it or re-pair it whenever you want. Turning it on is still entirely yours: it needs a bot you create and a token you paste, and Kit does nothing with Telegram until you do.
• Kit no longer loses a setting when you reinstall. Telegram had been switched on by a line typed into a config file by hand, so reinstalling wrote a fresh file without it and the feature quietly disappeared. Settings Kit reads are now written down where a reinstall and an update can both find them.

What changed

• Kit tells you when macOS is blocking it, and shows you how to undo that. macOS asks once whether Kit may read data from your other apps, and your conversation history is exactly that data. If that prompt is answered with Don't Allow, macOS never asks again and the folders quietly stop being readable, so Kit would find nothing to import and had no way to tell you why. It now notices, names which tools are affected, opens Privacy Settings for you, and checks again when you come back.
• Setup warns you before macOS asks. The prompt arrives with no explanation of who wants what, so the scan screen now says it is coming, what it is for, and that everything stays on your Mac.
• Having none of these tools installed is still perfectly fine. Setup only stops if a folder is there and Kit is refused; an empty Mac carries on as normal.

What changed

• The Try again button on the welcome screen works, and looks like it. Asking Kit to pick up reading your history where it stopped did start it, every time, but the card reporting it vanished the moment you clicked, so there was no way to tell. It now turns into the live reading panel straight away.
• And if that does not start, Kit says so. A refused or unreachable request used to look exactly like a working one: nothing happened either way, and you were left to guess whether the button had done anything.

What changed

• Kit can read your history on a reasoning model. Those models accept only their own default settings, and Kit was asking for a steadier one while extracting memories, so every attempt was refused. Conversations were imported and nothing was remembered from them. If that happened to you, run the import again; nothing was lost.
• When something goes wrong, Kit tells you what happened and what to do about it. Errors from your model provider used to arrive word for word, cut off mid-sentence, written for whoever built the API rather than for you. Kit now recognises the common ones, out of credit, a refused key, a model that is not available, a conversation too long to read, and answers with a sentence and a next step. Anything it does not recognise is passed on honestly rather than dressed up in a guess.
• If reading your history stops, Kit says so where you are looking. The progress strip used to disappear the moment a run failed, leaving the explanation on a health page you had no reason to open. It now stays, says why it stopped, says how many conversations are left, and offers to start again right there.
• The health page stops hiding health checks. It only knew about four kinds and quietly dropped any others, so the menu bar could report two things needing a look while the page showed one. Every check now appears.
• Health carries a dot when something needs your eye, so you can see it without opening the page.

What changed

• Setup shows its screens again. Since 0.2.158 the window opened onto nothing at all, which is as bad as a first impression gets. A guard added to keep untrusted pages out of the app was refusing the app's own screens, and every one of them is now opened by a test so it cannot happen quietly again.
• Setup no longer shows you pieces of Kit's own source code, and it accepts your email address. Those screens recently moved out of the code and into files, and the move mangled a few things on the way: some text arrived as raw code, and the rule that checks your email was left asking for an address with a backslash in it.
• Setup no longer hangs. Two different steps could stop dead and sit there. Installing, if something else on your Mac already had the port Kit wanted for its database, in which case Kit was quietly talking to a stranger's database. And connecting your tools, if a command Kit ran paused to ask a question on a terminal you cannot see. Kit now refuses any database it did not start itself, and never lets a background command wait on an invisible prompt.
• Kit reflects again on personal Kits. The reflection cycle has been calling a port nothing listened on, so it has not run once since Kit stopped using Docker. It raised no error you could see; it simply produced nothing, every eight hours, which is the hardest kind of fault to notice.
• Kit waits until you have finished setting up before it checks for updates. It used to check partway through, which at best interrupted you and at worst offered to replace the app underneath a setup still in progress.

What changed

• Kit's health stops crying wolf about your dreaming. Yesterday's release added a check for whether a night actually consolidated anything, and it was calibrated on a busy Kit. On a quieter one, a single active night made the next perfectly ordinary night look like a failure. It now asks whether your Kit is usually productive, over the last five nights, rather than comparing you to yesterday.
• Kit no longer tells you the Deforest runner has not been set up. It never was and never will be: that is a piece of machinery on the developer's own machine, and your Kit was being marked down for not having something it is never given. Reported by three people over three weeks, fixed properly this time and pinned with tests so it stays fixed.
• The Playbooks and Studio switch under Settings is hidden for now. It offered an orchestration surface your Kit cannot actually run: it would happily accept work and then leave it sitting there, because the part that does the running is not shipped with a personal Kit. If you already turned it on, nothing changes for you. It comes back when it can keep its promise.
• Health checks that belong to the developer's setup no longer appear on your Kit at all, rather than sitting there permanently amber with advice you cannot act on.

What changed

• Links in your memories work now. If you ever clicked one and nothing happened, that was not you: every link in a memory was being quietly dropped, because Kit asked macOS to open it in a new window and then never provided one. They open in your browser, where you expected them to.
• Kit opens quicker and stays responsive while it looks around for your coding tools. Two reasons, both embarrassing: the app was being built without compiler optimisation, and the Settings screen re-scanned your entire projects folder every time you toggled anything on it.
• Kit no longer freezes when one of your tools stops answering during a check. It waits twenty seconds, gives up, and says which one, instead of sitting there with the window locked and nothing to tell you.
• Downloading your memory export works on a large Kit. It used to gather every memory in one go before sending anything, so the more Kit remembered, the likelier the download was to stall. It streams as it goes now.
• Kit's health says whether last night's dreaming actually did something, not just that it finished. A night can complete cleanly having consolidated nothing at all, and until now that looked exactly like a good night.
• Text that reaches Kit from outside, in a Telegram message or an imported conversation, can no longer run code inside the app. Nothing is known to have used this and no Kit showed any sign of it; it was found by reading the code rather than by anything going wrong. It is the reason this release exists.

What changed

• The extra Telegram card came from a boot-time factory that re-minted a row for every merge to bury; the factory is gone, and the cards settle to one while keeping token, pairing, and history.
• Sending to Telegram from the Ask panel works now: the server resolves your paired chat itself, so the tab no longer depends on a chat id the app never knew.
• Updates stop asking macOS for permission to read other apps' data; those checks now run only on the screens that manage tool connections, where the question makes sense.
• Overnight "not running" alerts wait for a real outage: three failed checks in a row and a settled wake, so a sleeping Mac no longer pages you at every stir.
• The update screen now tells the truth about duration: Docker updates rebuild on your machine and can take ten minutes or more, with the database backup always running first.
• The MCP log's preload notice now says plainly that it is normal while the API restarts during an update.

What changed

• Telegram could end up as two cards in Settings, each holding half the setup: your bot token and paired chat on one, every send quietly using the other, and no amount of renaming or re-saving could make them meet. The two cards now merge into one on their own, keeping your token, your paired chat, and the name you gave your bot. Found on a beta Kit, fixed for every Kit.
• If a pairing was mid-handshake when the cards merged, the handshake stays open: your first message still lands as a pairing request instead of being dropped.

What changed

• Kit's night cycle now picks up where it left off after your Mac sleeps. A bookkeeping bug meant an interrupted night silently started over from the top every time, so on a Mac that closes its lid mid-dream the cycle could never finish. Found in a beta tester's log, fixed for every Kit.
• If the nightly cycle stops finishing, Kit says so the moment you open the menu: how many days since one completed, and how many attempts failed in a row. The same two numbers reach the beta roster so we can help before you have to ask, and the beta disclosure grew exactly that one sentence.
• A fresh Kit is now born on its own runtime: Postgres, Python, and Node running as plain processes on your Mac, no Docker anywhere. Kits already running on Docker stay exactly where they are until they approve the move in the app.
• The native runtime now ships for Intel Macs as well as Apple Silicon, and every install fetches the runtime built for its own machine.
• When native setup stops, the screen shows the installer's own explanation instead of a generic snag page. And the false alarm that wrote a scary line into the database log every five seconds is gone, so a real problem stands out.

What changed

• Kit's nightly consolidation no longer dies when the model thinks its way through a reply without writing one out. The night carries on instead of stopping there, and the health screen names the real cause rather than a generic error.
• Searching your included folders now returns the full set of matches. On a Kit holding a lot of sources it could quietly hand back a fraction of what it had actually found, with nothing to say it had.
• When part of Kit stops unexpectedly and is brought back, the health screen now tells you when it happened, not just that it did. A count on its own gave you nothing to line it up against.
• Kit's logs no longer report a working setup as a failure at startup, so a real problem stands out instead of hiding among false alarms.

What changed

• After moving your Kit to the new runtime, it now tells you to quit and reopen anything connected over MCP. Some apps do not reconnect on their own after the short pause, and a stale connection there looks like a broken Kit when nothing is actually wrong.

What changed

• Moving your Kit to the new runtime no longer quietly changes it. The settings that shape how it remembers, and whether it asks you to sign in, now travel with it instead of being left behind.
• The menu bar is tidier: Health, Settings and Send us feedback moved into Advanced, so the top level is just the places you go inside your Kit. If something needs attention, Advanced still carries the count.
• The fireflies view is for looking again. The ask box has been removed; Talk to your Kit is where a conversation belongs.

What changed

• Kit follows a conversation properly now: a short follow-up like "yes" no longer dead-ends with "I don't have that noted" when the answer is right there in the thread.
• When a search finds nothing, Kit looks again using what you were just talking about before it says it has nothing.
• Telegram in Settings reports the real state of the connection, so a saved token is no longer mistaken for a working channel.

What changed

• Connect Telegram now lives in Settings, and it tells you whether the channel is actually working rather than just that a token was saved.
• The model shown in the menu bar is the same one Settings shows: picking a model in one place no longer disagrees with the other.
• Changing your model actually takes effect everywhere, instead of looking saved while background work carried on with the old one.

What changed

• Updates now finish properly: Kit comes back on the new version instead of offering you the same update again.
• The Kit panel loads reliably after an update.
• Claude Desktop and Claude Code now connect on every new install: the connector ships ready to run.
• An update that stalls partway now finishes itself on Kit's next check-in.
• When a dream or background task fails, Kit records what happened so it can be diagnosed for you, no screenshots needed.
• Channel chat history loads from the right place again.
• Opening a task from Studio brings up that task, not an empty dialog.
• Tidying behind the scenes: retired unused machinery and a refreshed test suite.

What changed

• A dream interrupted by a database hiccup now resumes where it left off instead of failing.
• Kit stays responsive while it imports and generates memories.
• Claude Desktop and Claude Code reconnect cleanly after a Kit rename.
• Your model choice now applies to everything Kit does, not just chat.
• Beta monitoring now reports which runtime Kit runs on and whether its services are healthy; the terms page spells out exactly what is sent.

What changed

• Housekeeping: background services are quieter and more honest in their logs, and groundwork continues for a feature in private testing.

What changed

• Internal groundwork for features that are still in testing, plus quieter, more reliable background services.
• When something in the background cannot do its job, it now says why in its logs instead of going silent.

What changed

• The first update to arrive through the app itself. Nothing else changed; this release exists to walk the new path end to end before your Kit asks you to trust it with a real one.

What changed

• Updates now arrive inside the app. When a new version is ready, your Kit tells you and one click installs it and relaunches; no more visiting the site, downloading, and dragging. This version is the last one you install by hand, and the download stays behind your beta access either way.
• Groundwork for a new way to reach your Kit away from your Mac, in private testing before it reaches the beta.

Improvements

• Your kit's internal parts are now named for what they are, and the updater retires the old ones itself, in the right order: your kit keeps running on the old version until the new one is fully built, and your memories sit in storage the change never touches.
• If an update fails, your kit sends us the error lines from that run, with your sign-in details stripped, so we can fix what went wrong without asking you to dig logs out of your Mac. The disclosure at install time says exactly this.

Bug fixes

• Updates land themselves. When an earlier update had not actually reached the running kit, the app used to show a screen with a button and otherwise wait; it now quietly reruns the update at the next launch, and only asks for the button press if that self-heal fails too.
• An update waits for your kit to finish dreaming or reading your history before it restarts anything, so that work is never lost mid-flight.
• Each successful update cleans up what earlier versions left behind. Old runtime images used to pile up quietly until the disk Docker uses filled and every later update failed at that same wall.
• An update that cannot verify what is running says exactly that and retries at the next launch. It used to record success on silence, which could leave a kit convinced it was current while running weeks-old code.

Improvements

• Your kit is better at knowing when it does not hold the answer. A question can be about exactly the right subject while asking about a part of it your kit has nothing on, and that used to read as a confident match. It now weighs which words are missing rather than how many, so a question whose distinctive terms are absent has to clear a higher bar before it is answered.
• Asking about something your kit has never come across returns quickly. That was the one case where a search had to read every memory before it could answer, and on a large kit it showed.

Bug fixes

• When a problem on the health screen has already been fixed, your kit says so and offers to check. It used to keep showing the old failure with nothing you could press, because the screen reports the last nightly cycle and that cycle ran before the fix arrived.
• Faults are quoted in full. A long error used to be cut off partway through, sometimes mid-word, because the room set aside for it was being spent on internal detail nobody needs to read.

Improvements

• You can name a fallback model in Settings, Models. When the provider running a job stops answering, because the account is out of credit or the key has stopped working, your kit finishes the job on the fallback instead of going quiet. Point it at a local model and your kit keeps working even when a paid account goes dark.
• When the health screen reports a problem, it now tells you the cause. It used to name only what had failed, so a card could say a nightly job went quiet while the real reason, an expired key or an account out of credit, sat unread underneath it.
• Times on the health screen are written the way people say them, so you read "no progress for 40 minutes" instead of a raw timestamp in another timezone.

Bug fixes

• Claude Code stops losing its connection to your kit. A single web request that was cut off partway, a reload or a closed tab, could take down the process that every Claude session was running inside, so all of them died at once with no warning and no request in flight.
• Your kit's nightly check on whether it is honouring your corrections works again. On kits set up before a recent change, that check had been quietly rejected by the database every night, so the record it is supposed to build was never written.
• The buttons on the health screen do something you can see. Starting a dream from a card used to sit silent for several minutes with no sign it had begun, which read as a broken button. It now starts in the background and says so.
• Your kit no longer accuses itself of being stuck when it is merely working slowly. It watches over a real stretch of time before saying anything, so a kit you have just started, or one taking its time with a large import, is left to get on with it. It also stops quietly restarting that work over and over behind the scenes.
• Quitting your kit in the middle of its nightly dream no longer leaves that dream half-open behind it. The next one clears it and starts cleanly.
• A personal kit stops reporting on machinery it was never given. It used to warn that a background worker had never checked in, on installs where that worker does not exist and nothing was waiting for it.
• When something breaks, your kit tells you what broke. It used to say it hit a snag whatever the cause; now it names the fault, and messages you as soon as it happens rather than the next time you say hello.
• Spend alerts keep speaking. They used to warn once a day and then fall silent no matter how far past the budget a day ran.

Improvements

• Your kit now learns from every correction you give it. It could only hold seven before, and once those filled, everything you taught it afterwards was quietly dropped. It keeps all of them now, surfaces the ones that matter each time it wakes, and can trace why it behaves differently than it did months ago. Corrections you gave in the past have been recovered and dated, so your kit starts from what you already taught it.
• Connecting an agent to your kit no longer tells that agent to become your kit. It reads your memory as an archive it can cite, and waking as your kit is something you switch on for a specific key.
• Answers cite the memories they came from, or say plainly that they cannot support the claim, instead of filling the gap.
• Your kit exports to plain text files any computer can read, with a checker that needs nothing installed, so your memory outlives the app.
• Kits can share memory under a signed grant that names both ends, expires, and can be revoked on the spot, with both sides keeping their own record of the exchange.
• Backups are hardened, with a verified encrypted path off the machine.
• The nightly dream cycle is more careful: it repairs missing areas from graph consensus, flags broken provenance and stops trusting it at answer time, and mints new connections on an earned weekly rhythm rather than freely.

Bug fixes

• Setup reads honestly. The welcome screen was animating its steps as though the work were already underway, complete with green ticks, before anything had run; it now reads as the path ahead.
• The beta terms checkbox is the right size and sits with its label, and it tells you when it needs ticking instead of blocking Begin with no visible reason.

Improvements

• Search results now say what to trust. Every hit carries a relevance score that means the same thing across searches, and a verdict on whether the hit is grounded in what your kit actually holds, instead of one opaque internal number that ranked noise above exact matches.
• Ask about something your kit has never heard of and it now says so. Gated recall returns nothing instead of the nearest lookalike, and a question whose words appear nowhere in your kit can no longer surface a coincidental match as if it were memory.
• Agents talking to your kit can ask for lean search results and a per-question confidence report that names which of their words matched nothing. That is how an agent tells "no answer exists" apart from "wrong question shape".
• Strong match now means the same thing on every kit: the relevance bar is calibrated to the embedding model your kit actually runs.

Bug fixes

• Closing the setup window no longer strands you. Setup remembers where you were: relaunching the app takes you straight back, and the menu bar offers Finish setting up whenever steps remain.
• The model check now tells a stuck server from a slow model, and can restart Ollama for you with one click. A model paying its one-time load from disk is no longer called slow.

Bug fixes

• Importing your history is much faster on the newer Claude models, and no longer stops with nothing to show for it. Your kit was thinking hard about each conversation before summarising it, which is not something that job needs. On the newer models that turned a few seconds per conversation into a few minutes, until it ran out of room and gave up.
• Your kit's health page now shows how many chat turns you actually have, instead of always saying zero. If you read that as an empty kit, it never was: the number was simply never being counted.

Improvements

• Your kit now tells you when it is still running an older version than the app, instead of quietly staying behind. Update now brings it up to date, and nothing is lost when it does. The app and your kit update separately, and until now nothing compared the two, so a kit could sit on an old version for days with no sign of it anywhere.
• Every setup and update run now keeps a log on your Mac, and the details are on screen if a run does not finish. When something goes wrong there is now something to look at.

Bug fixes

• Importing your history no longer gets stuck on a single conversation. If one never finishes, your kit steps over it and carries on, rather than starting from the beginning and stopping in the same place every time.
• Repair now spots a connected tool that is holding a key your kit no longer accepts, and fixes it. That was the "rejected the API key" error when waking your kit in Claude Code, which Repair previously reported as already in sync.

Improvements

• In VS Code, your kit now wakes on its own at the start of a chat instead of waiting to be asked. It was the one editor we never gave a guidance note to, so nothing ever told Copilot your kit was there to wake.

Bug fixes

• Waking your kit no longer fails in VS Code and Copilot with an "invalid parameter format" error. The assistant phrases the request slightly differently there, and we were being stricter than we needed to be.

Improvements

• Turning beta monitoring off now really does mean nothing at all. Setting KIT_TELEMETRY=0 before you finish setup stops the welcome email too, not only the version reporting. We said "your kit reports nothing" and it was very nearly true, which is not the same thing.

Bug fixes

• Your kit reports its version more reliably. If your Mac was offline when the app started, or our end had a hiccup, your kit used to go quiet for a day; it now tries again shortly after.

Improvements

• Your kit's cards now render in Claude Code. Waking up, recalling and picking a project come back as the real thing rather than a wall of text, which closes the known issue noted in 0.2.115. It needed your kit to be registered with the Claude desktop app, so it will also appear in Claude's chat there, since the two share one setting.
• Settings added in later versions now reach kits installed before them. Until now a kit kept the settings it was born with, so improvements could quietly pass you by while a newer install got them.
• While Kit is in beta, your kit tells us your name, your email, which version it is running and when it last ran, so we can see which versions are live and support them. That is the whole list: no memories, no messages, no files, no project names, nothing you or your kit work on. Set KIT_TELEMETRY=0 in your .env to turn it off, or ask us and we will delete the record.

Bug fixes

• Your kit shows up in Claude Code properly. It was being registered against whichever folder the app happened to be running from, so it was there in one place and invisible everywhere you actually work, while everything reported success. Existing installs repair themselves the next time the app starts; there is nothing to redo.
• A brand new kit no longer shows an empty "Most recent work" box before it has any history to show.
• The card's "most recent work" no longer shows something from days ago. It was baked in when the card was first drawn and never refreshed, so it could sit there long after the work moved on.
• Your memory count no longer disappears from the card a moment after it appears.

Improvements

• Your kit's own tools are all here now. A personal install used to expose 10 of them, the lean set built for strangers reading somebody else's kit; it now serves the full 38, so the kit you own answers on your surfaces the way it does everywhere else.
• Turning a surface on now tells you what to do next in that tool. Every editor reads its MCP config at startup, so it needs a reload before the tools appear, and VS Code additionally keeps them behind Copilot Chat's agent mode picker. The row says so instead of flashing a checkmark and leaving you guessing.

Bug fixes

• Editors that hold a live connection open (VS Code, Cursor, Claude Code) no longer trip the MCP server: it now handles a call and an open stream at the same time, instead of refusing everything after the first one.
• Known, and not your setup: the wake and recall cards still show as plain text in Claude Code and Cursor, which are waiting on their own MCP Apps rendering. The tools work everywhere; Claude Desktop and VS Code render the cards today.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• Soul curation: when your kit's always-loaded memory tier outgrows its budget, it proposes a small, reversible cleanup batch as a single card you approve with one tap. Identity memories are never touched, and nothing moves without you.
• Sharper waking on coding surfaces: one patient message while a new kit finishes forming, clearer answers when a question could mean the open project or stored knowledge, and read-only keys say so upfront.

Bug fixes

• Dreams no longer choke when many similar memories chain into one giant consolidation group; oversized groups now merge in safe chunks.
• Resume on the wake card now always points at the genuine last session, and client error messages are no longer stored as chat history.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• The wake card grew up: a fourteen-day sleep strip (memories per day, a dot for each dream cycle), a one-line summary of your last session with Resume beside it, and chips that choose where new memories are stored, a knowledge area (General by default) and optionally a project.
• The recall card now leads with what was found: "recalled that memory and 6 others" instead of a bare count.

Bug fixes

• Resume now always points at the genuine last session, and every card speaks your kit's own name, including when it is offline.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• Updating Kit no longer silently disarms its start-at-login entry, so Kit truly comes back after a restart.
• Kit now verifies on every update that its login item is genuinely registered, and repairs it if macOS lost it.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• Attention cards in Studio are now actionable: approve, answer, or resolve a waiting run right from the card instead of hunting for the run behind it.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• Your kit now comes back on its own after a restart or crash: the app checks its engine, starts Docker quietly in the background if needed, and wakes everything without a click. Stopping your kit by hand is still respected.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• Interrupted history processing now clears itself promptly after an app update or restart, even when the replacement worker happens to reuse the old process number.
• Model calls used while forming memories now have a firm time limit, and progress stays consistent across multi-worker kits, so a dead request cannot leave sleep looking alive indefinitely.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• Sleep and wake state is trustworthy again: abandoned dream work clears instead of hanging, and the "since sleep" count can no longer exceed the memories your kit holds.
• Imported history now processes once at a time and links only the memories it just formed, so new and descendant kits no longer churn through the whole brain after a failure.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• The memories views now show only your own content, everywhere: counts, the field, the sunburst, group drill-downs, filter chips, connection lists, and working-memory halos. Clicking around can no longer hit a "memory not found" dead end, and category counts always match what the list shows.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• Your kit keeps its inner workings to itself now: notes about its own memory system no longer appear in your views, your search, or your postcards, and it answers briefly by default.
• Everything your kit says uses your name and its own, everywhere.
• The answer box over the memories view can be collapsed, and the welcome message on a young kit can be dismissed, so the field is always yours to see.
• The Ask tab is tucked away for now; asking from the memories view is the front door.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• Ask your kit anything, right from the first tab: the compact window and the menu bar now lead with a full-window chat, named after your kit.
• Asking a question over the memories view now lights up the memories the answer drew on.
• The full web interface is set aside while it is being finished; the app's own views are the supported surface for now.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• Fixed the "running and running" loop: while your kit was generating memories from imported sessions, its crash-recovery could mistake that work for a crashed sleep cycle and repeatedly start full dream cycles on top of it. Memory generation and sleep now stay cleanly separate, and those phantom cycles no longer spend your AI credits.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• The Timeline tab is back in the compact window. It was wrongly hidden on personal kits; your kit's memory-activity and sleep view was never meant to depend on the orchestration features being on.
• Emails your kit sends now retry automatically when the mail server asks it to try again later, instead of giving up.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• Memory generation after an update is much faster, and no longer re-processes your kit's own internal machinery sessions, so a large backlog shrinks dramatically instead of grinding for hours.
• A stalled memory-generation run now cools down and waits instead of silently restarting itself in a loop.
• An update can no longer trigger a full re-import of an already-established kit.
• Night-time association forming (part of your kit's sleep cycle) works again and finishes within its budget.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• If you gave your kit a name of its own, it now uses that name everywhere it refers to itself: the tool hints your editor shows on hover, the app menu and status, the notifications it sends, every panel, tab, and description inside the app, and the notes and replies it writes. It no longer falls back to "Kit".
• Renaming your kit no longer leaves the old name behind as a broken connection in your coding tools. It updates cleanly wherever you have added it. If you use AntiGravity, reopen it after connecting so it loads the change.
• When your kit follows its own curiosity, it now points that at your work and projects, rather than at its own inner workings.

Improvements

• The AI key editor in settings now works the way onboarding does. Pick a provider and the right flow follows: local providers (Ollama, LM Studio) lead with their address and a "Check connection" button; cloud providers with a key and "Check key". Whatever happens, it tells you: how many models it found, that it's showing a known list, that nothing answered at that address, or that the window isn't signed in. And if no list can be fetched, you can simply type the model name; nothing dead-ends anymore.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• Your kit can read what you attach to email now. Send it a screenshot and the text in the image is read on your machine and lands in your kit's context, so a picture of a problem is as good as a description of one.
• Rename and Uninstall moved into an Advanced menu, out of the way of everyday clicks.

Bug fixes

• Changing the AI provider in settings no longer strands the model picker. Each provider now starts from its own endpoint, the endpoint field shows for every local provider (LM Studio included), and if no models can be found you can simply type the model name; the editor tells you why the list is empty instead of going quiet.
• Updating no longer walks you through the full install steps; updates show the short resting screen only, and the menu stays locked until your kit is back, so a window can't open against an engine that is mid-update.

Updating keeps everything your kit has learned, nothing is reset.

Improvements

• You can rename your kit's engine from the menu bar now: "Rename Kit's stack…", type the new name, done. Your kit restarts under the new name and keeps everything it has learned. No hidden files, no terminal, and the manual steps we described earlier are no longer needed.

Improvements

• Your kit now reflects on its own, three times a day. Until now only the original Kit did this, because the reflection ran on its home machine rather than inside your kit. You will see these in the Reflections tab, and in what your kit remembers about how its days went.
• Postcards arrive. Every so often, when there is genuinely something to say, your kit writes you a short note about what it has noticed. You will find them under From Kit. If there is nothing worth saying, it says nothing, that is deliberate.
• Engine logs are now readable in the web UI, under Settings. No terminal needed when something looks off; the same view is what we will ask you to look at when you report a problem.
• You can rename your kit's engine (the Docker containers) safely. Renaming keeps every memory. If you have not thought about renaming containers, you never need to.
• If you gave your kit a name of its own, the notes it adds to Claude Code and Cursor now use that name instead of calling itself Kit.

Bug fixes

• Signing in to the menu bar windows no longer depends on the password the app saved when you first set up. If you changed your password in the web UI, older versions could lock the app's windows out permanently while the browser kept working. The app now signs in with its own key, which cannot go stale.
• If a window still cannot sign in, the card now says exactly why, in one line, and asks you to include that line when you report it.
• The app now tells you, with a normal macOS notification, when your kit's engine has stopped without you stopping it, for example after a macOS update shut Docker down overnight. Before, the only sign was a quietly different menu bar dot.
• The "Generating memories from your sessions" banner no longer covers the navigation while a long import runs. It moved to the bottom edge, out of the way.
• A rare overnight failure in the dream cycle (the model spending its whole answer budget thinking and returning no text) now recovers by itself instead of erroring that night's tidy-up.
• An 0.2.80 build existed for about half an hour and never reached anyone; its version of the rename safety net could stop a kit from starting. 0.2.81 is that release, fixed, and it repairs any install that launched 0.2.80 automatically.

Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• A menu bar window could still open to a "couldn't sign in" card and stay there, and quitting and restarting your kit was the only way out. The cause was not what 0.2.78 fixed. The app was throwing away the result of its own sign-in, so when the sign-in failed, usually because your kit's engine had not finished starting yet, nothing noticed and nothing tried again. It now waits for the engine to come up, retries, and reloads the window by itself the moment it is signed in.
• When it genuinely cannot sign in, it now writes a line to the system log saying which of the two possible causes it hit, so a report can be answered instead of guessed at. Updating keeps everything your kit has learned, nothing is reset.

Bug fixes

• The menu bar windows (your memories, timeline, data sources, and the rest) now open reliably signed in. On some Macs they could show a "couldn't sign in" card and stay blank; this version hands the session to each window before it loads, so it just works. Updating keeps everything your kit has learned, nothing is reset.

Improvements

• A new kit now arrives with a small inherited disposition: how it speaks, what it protects, what it refuses to lose. It is the same for every kit, it carries nothing about anyone else, and everything your kit learns about you stays its own. Existing kits pick up the inherited set automatically on update, replacing the generic starter notes they were installed with; nothing your kit has learned about you is touched.
• Every release of this inherited content passes an automated privacy check plus a manual review of each entry before it ships.

Bug fixes

• A follow-up to the privacy fix in 0.2.75. The self orientation part of a kit's wake-up summary used to point at memory references from the developer's own Kit, which showed up as broken references on a fresh install. It now reads only from your kit's own memory, and stays out of the way until your kit has grown its own self map.
• More of Kit's internal identity plumbing now derives from your setup, never from defaults that belonged to the developer's Kit.

Bug fixes

• A privacy fix in setup. Earlier versions could carry a few of the developer's own setup details, like a name in some background prompts, into a new kit. This version removes them, so your kit only ever refers to you. A kit that is already installed cleans itself automatically when it updates to this version.

Bug fixes

• When you choose a local provider like Ollama during setup, Kit no longer quietly picks a model for you. It now waits for you to choose one, so the model that reads and organises your history is always the one you meant to use.

Improvements

• Kit's home screen is much lighter on your Mac's graphics now. The gentle background animations were quietly keeping the GPU busy even while you were just reading; they are calmer, so your fans stay quiet.
• In the memory map, the label on each cluster is smaller, sits closer to the node it names, and you can now click the text itself to open a cluster, not just the small dot.

Bug fixes

• The health panel opens cleanly again instead of slipping behind the cards in front of it.
• Nightly dreams on a local model are gentler on your graphics card, and no longer lose the whole night's work if the model returns something unexpected.
• A menu-bar window that cannot sign in now tells you and points you to reopen it, instead of showing a blank starfield.

Improvements

• Setup now guides you as you go: a missing name, an incomplete email, or a password that is too short is flagged right on the field, instead of the step quietly refusing to continue.
• After setup, the welcome screen tells you when Kit is dreaming and offers a button to watch its progress, so you can see it is busy learning your history in the background.

Bug fixes

• Dreaming now works with LM Studio. A recent LM Studio version stopped accepting the format Kit used to request memories, which failed every dream; Kit now talks to it correctly, and falls back gracefully with other local servers.
• Local models from Ollama and LM Studio now load when you choose them during setup, instead of the list coming up empty.

Improvements

• Refreshing Kit now shows a calm progress bar that follows the actual stage of the update, so you can see how far along it is instead of watching text cycle.
• If you have given your kit its own name, that name now shows everywhere it connects, including the tools it offers your editor.

Bug fixes

• New memories now form reliably whichever model you choose. A newer model could quietly stop Kit from turning your imported history into memories; the same fix also restores dreaming on those models.
• While Kit is importing or dreaming, the menu no longer half-lights. The memories, timeline, sources, health, and settings views now wait together and become available as one when Kit is ready.

Improvements

• A fresh Kit now starts dreaming shortly after setup, rather than waiting for the first night, so your first memories settle in sooner.
• The compact settings hold the full Sleep Quality dial and an optional hourly power nap, and the "Kit is dreaming" indicator shows the same gentle wave as the deepest sleep, so you can tell at a glance when the view is still forming.
• A calmer personal Kit: until you turn collaboration on, the inbox and the activity timeline stay out of the way.
• A "Send us feedback" tab, a warmer welcome screen, and a gentle screen while Kit refreshes itself.
• Kit remembers what you do from your setup answer, so it understands who it is working with.

Bug fixes

• The timeline no longer shows weeks-old dream cycles as if they were still running, and your sessions and runs are no longer buried under them.
• Kit no longer reports itself overdue for sleep right after an hourly power nap.
• Chatting on a local model no longer hangs: a slow model fails quickly with a clear message instead, and stays warm between messages.
• Downloads now carry the version in their filename, so re-downloads no longer pile up as copies.

Improvements

• The memory galaxy is cleaner to move around, and each memory now shows where it came from, the session or source that created it.
• Sharper recall: Kit now recognises the organisations in your notes alongside the people and places, and follows those connections to bring back more of the right related memories for a given question.
• Your memory count and the memories view now reflect real knowledge. Kit's own inner-voice notes and session handoffs are still kept, but moved out of the everyday view so the numbers mean something.
• The nightly dream cycle now survives your laptop sleeping or closing and resumes where it left off. A new Sleep Quality dial and an optional hourly power nap let you tune how often Kit consolidates.
• Kit creates projects automatically from your git repositories, and folds duplicate people, places, and organisations into a single identity.

Bug fixes

• The nightly dream cycle no longer stalls when a local reasoning model wraps its answer in extra text.

Bug fixes

• Restored the hover tooltip in the memory view that the previous release had accidentally hidden.

Improvements

• Kit now captures your coding sessions the moment they happen on more surfaces (Claude Code, Codex, Gemini and Qwen), so your timeline stays current instead of waiting for the nightly catch-up.
• Rename your kit, and rename or delete knowledge areas, from settings.
• A styling pass on the menu-bar app: tidier fields, buttons and dropdowns, round timeline markers, and save confirmations you can actually see.
• Updating now reads "Updating Kit" rather than "Installing".

Bug fixes

• Tidied history that showed redaction markers or raw data carried in from imported sessions.
• Renaming your kit in settings now takes effect.
• Hover tooltips in the memory view no longer linger when you change tabs.
• Removed a stray "import more history" button; the repair option now shows only for connected surfaces.

Improvements

• You can now check for updates any time from the menu bar.

Bug fixes

• Fixed the model picker in settings not showing its options in the menu-bar app.
• The live facets panel no longer shows entries that are not part of your own Kit.

Improvements

• Kit wakes much sooner during setup: it reads your most recent history first and starts being useful while the rest finishes quietly in the background.
• A new dreaming view shows when Kit is consolidating, with a calm animation, what it is working on, and a tap to peek at a memory or connection it just made.
• A quieter first run: no file-permission pop-ups during setup.
• You can point Kit at your own model: any self-hosted, OpenAI-compatible server (for example a Qwen box on your network).
• If the model you choose is too slow, Kit now tells you while you pick it, and lets you switch to a faster one without restarting setup.
• Kit writes in British English and refers to itself by the name you gave it.

Bug fixes

• Fixed a long-standing failure in Kit's nightly dream that could stop it consolidating.
• Uninstalling Kit no longer crashes, and leaves nothing behind.
• The health screen now reads correctly while Kit is doing its first consolidation.
• Setup progress is steadier, and common snags like Docker not running now offer a clear retry instead of a dead end.
• Faster, more reliable indexing of your conversations.

Improvements

• Health at a glance: a calm status light in the app and the menu bar shows how Kit's dreams, surfaces, and data sources are doing, with one-tap fixes when something needs attention.
• Surfaces you do not have installed now link you to where to get them, instead of showing a switch that cannot do anything.

Bug fixes

• You can no longer switch on a coding tool that is not installed on your Mac.

Improvements

• Clearer memory cards on the dashboard.

Bug fixes

• Fixed a case where new memories could silently fail to save on some models.
• Fixed your name showing incorrectly on a fresh install.