Local Character › Change log · Last reviewed 2026-10-09

Change log

Newest first. Each entry has a permanent link. If a claim on this site turns out to be wrong, the correction is recorded here with the date.

• 2026-10-09 - Group chats: Let them talk for now runs 1 to 8 rounds in a row (pick the number next to it; one round is one reply). With Voice on, each reply is read before the next character speaks. The button shows Stop while they talk; sending a message stops them too. #
• 2026-10-09 - Group chats: characters can now talk to each other. Let them talk, next to the member chips above the message box, makes one character answer the last line without a message from you - usually whoever was just spoken to. Tap a member's chip with the box empty and that character speaks. Each step is one reply, so the scene never runs on by itself. #
• 2026-10-09 - New: group chats. The + next to the character's name at the top of a chat brings in up to three more characters, and the chat keeps its history. Each message gets one reply: from the character you name, the one whose chip you tap above the message box, or otherwise the one the model judges should answer. Every reply shows who said it and is read in that character's voice; what you tell the group, every member remembers. Remove a member in Members; with one left, it is an ordinary chat again. On the device, a reply took 3 to 4 seconds in our tests with two and three characters. #
• 2026-10-07 - Smaller downloads. The chat model is now about 1.5 GB instead of 2.0, pictures 1.0 GB instead of 2.2 and the voice 0.11 GB instead of 0.37, so all four models take about 2.9 GB instead of 4.9. Replies in the app's languages are unchanged, and pictures look the same but take about half a second longer. Details and measurements are on the methodology page. If you already downloaded the models, nothing changes and nothing downloads again. #
• 2026-10-07 - Characters take the lead when you write little. A reply used to shrink to the size of your message, so "ok", "hmm" or "嗯" got a line or two, often asking you for more ("give me a name", "您想了解哪方面？"). Now the character reacts and then moves the scene on by itself - Bram brings out a rumour and a map, Mara starts on a case, a librarian takes a book down and reads you a line - in a couple of short paragraphs. In tests of 42 short messages across 7 chats, replies went from about 50 to about 100 words, and fewer ended by handing you a question. Short style still keeps replies brief. #
• 2026-10-07 - Fix: a short request such as "具体的时间线" (the exact timeline), "列出来" or "more details?" was treated as small talk and got a quick reply, so a character could say "I'll put that together for you" and stop, or give three points and ask whether you want more. Short questions and requests now get a full answer, and a request for a list, details, an example or a timeline is answered with the items themselves. "OK", "thanks" and "haha" still get quick replies. #
• 2026-10-07 - New: Think before replying, under Style (off by default). The model first thinks the reply through - what you said, the character, the facts it was given - and then the character answers with that in mind. The thinking appears folded above the reply ("Thought for 8 s"); tap it to read it. Replies take about 10 seconds longer. It runs on your device like the reply itself, and Stop works while it thinks. #
• 2026-10-07 - New: Knowledge, in a character's details panel. Add documents the character should know (text, Markdown, Word .docx, EPUB or web pages, or paste text; up to 30 documents and about 1.5 million characters of text per character). With every message the app looks up the parts that match, and before answering a question the model first pulls the exact answer out of them; the reply shows which documents were looked up. Small talk only draws on them when it is clearly about something in them. The documents stay in this browser: they are not in your account backup, not in a Community share and never sent with a cloud reply. Tested on 35 questions about documents the model could not know (invented lore, and Wikipedia articles from after its training): 33 answered correctly, against 1 with no documents, where it made answers up. PDFs are not read yet; copy their text and paste it. #
• 2026-10-04 - Fixed: a chat deleted while its reply was still being written, or just after (while the reply was read aloud), came back the next time the app opened. Deleting now stops that reply and the chat stays deleted. New: Delete all chats with a character, in the character panel and at the end of History (what the character remembers about you is kept; Memory clears that). Delete and other confirmations use the app's own dialog, so they also work in browsers that block pop-up dialogs. #
• 2026-10-04 - The starter characters have fuller descriptions: what they do, how they talk (with a sample line), who you are to them and, for group chats, a distinct voice for each person on their own line. In blind side-by-side tests on the on-device model, replies with the new descriptions were preferred about two to one. The Create page now suggests writing your own characters the same way. Also fixed: after a reply that reached the length limit, the next message failed with an error. #
• 2026-10-05 - Fix: in Chinese and Japanese replies the narration was read in the character's voice (a whole *action* was taken for one emphasised word); it now goes to the narrator. #
• 2026-10-07 - Fix: a character's own words written without quotes (like a greeting that opens with "Hello, welcome to my study.") were read by the narrator. Words in the first person are now read in the character's voice, and shown as speech. #
• 2026-10-07 - Replies no longer come in fixed sizes. On the Normal style the character decides how long to write, in step with you: a line or two for small talk or a thanks, a few paragraphs when something happens, more for a big scene or a story you asked for. Short and Long in Style still set it yourself. #
• 2026-10-06 - Characters follow through: asked for a book, a librarian now takes one off the shelf in the same reply instead of stopping at "let me see what we have". A short answer to the character's own question is no longer treated as small talk (Chinese and Japanese messages are measured by how much they say, not by character count), and when you switch language mid-chat the reply switches with you. #
• 2026-10-06 - Asked about something from early in a long chat, characters now answer it instead of dodging. Before replying in character, the model first reads the earlier lines the app found and pulls out the exact answer, then the character replies with it in mind. On 33 recall questions this went from 50 to 53 of 69 details; asked at the end of 11 long chats, from 12 to 20 of 27. It adds well under a second, and only when you ask a question about the past. #
• 2026-10-05 - Characters remember the early part of a long chat much better. When you ask about something from before, the app finds the earlier lines that answer it - with the sentences around them, and matching word forms ("texted" finds "text") - and hands the model a longer list of them, right where it starts its reply, with a reminder to answer with the exact names and numbers. On 33 questions about the start of 11 long chats, correct details went from 26 of 69 to 50 of 69. #
• 2026-10-05 - Long chats keep their story. When a chat grows past what the model can read at once, its older part is folded into Story so far notes (see and edit them in the chat's details panel), and before each reply the app looks up the earlier lines that match what you just wrote - "how many crystals do I carry?" finds "I always carry three" - and shows them to the model next to your message. Tested on 11 long chats with a small window: the characters recalled more of the early facts you asked about. #
• 2026-10-05 - Replies now mark everything: actions and description in *asterisks*, speech in quotes. The model is asked to, and the app fixes the few it leaves unmarked, so the grey narration and the narrator voice always line up. #
• 2026-10-05 - Persona moved next to AI Models, and logging in now also backs up and syncs your personas and narrator voices with your characters (chats and memories still never leave). Cloud replies no longer send your persona or memories, as the privacy page always said. #
• 2026-10-05 - Community has category filters (Adventure, Romance, Companions and the rest, each with its count), and the list fills the screen first and loads more as you scroll down. #
• 2026-10-05 - The narrator is always the other gender from the character's voice (now for voice mixes too), and you can pick both narrator voices, or turn the narration off, under Persona. #
• 2026-10-05 - Reading aloud now includes the scene: the grey *action* and description text is read by a narrator voice, and the character's lines in their own voice, one after the other with no gap. The narrator is a man's voice, or a woman's for characters with a man's voice. Cloud voices still read only the spoken lines. #
• 2026-10-05 - Recommended on the home page now shows six characters picked for you: ones like the characters you made or chat with, mixed with a few at random, so each visit brings different ones. Search still finds every character. #
• 2026-10-05 - Animated portraits: an uploaded GIF or animated WebP now stays animated (it used to keep only its first frame), and sharing an animated character to Community keeps it moving, as a small animated WebP. Browsers that cannot re-encode animations (Safari, Firefox) still save a still picture. #
• 2026-10-05 - The search box on the home page also finds characters shared in Community. The shared list is downloaded once when you search, and what you type stays on your device. #
• 2026-10-05 - Long chats hold up better. Replies no longer retell your message, stop asking "what do you do next?", stay a readable length (shorter for small talk, fuller for role-play) and finish their last sentence instead of stopping mid-word. A character no longer repeats the sample line from its description over and over. Memory keeps lasting facts about you (your name stays) instead of passing moments, and the model keeps more of the earlier chat after it learns something new. A character's first chat opens with its own greeting. #
• 2026-10-04 - "New reply" is gone: the on-device model writes the same reply every time for the same chat, so asking again only repeated it. Instead, the pencil under the latest reply lets you rewrite what the character said. The original is kept as an earlier version you can flip back to, and the character's next replies follow the version you chose. Editing an earlier reply opens a new chat that ends at it, and the original chat is kept, so you can compare where each version leads. #
• 2026-10-03 - Auto-fill has a Model menu: on this device (Gemma 4 E2B, the default), or on the server - Gemma 4 26B or GPT-Sol - for richer cards. Server models are free (the app pays) and the note under the box says where your request goes. Phones offer the two server models. #
• 2026-10-03 - Delete everywhere now sticks in your account: the backup file is always read fresh (a cached copy from before the delete could put the character back), and this browser keeps re-applying its deletions, so an old tab or an old copy of the file cannot restore a deleted character. #
• 2026-10-03 - Phones and browsers that cannot run the on-device models now work straight away: no download offer, replies and voices come from the cloud models (free), and a note says that private on-device chat needs a computer with WebGPU. Pictures, speech input and voice-clip matching are hidden there. #
• 2026-10-03 - Interface fixes from a full pass on phone and desktop: character panels cover the whole screen on phones, their arrows line up, "1 chat" instead of "1 chats", Auto-fill and Find with AI show their whole hint on phones, AI Models rows no longer squeeze model names, the download window drops a progress bar that never moved, the Mine page puts Create before Log in, and a search no longer sticks after you leave the page. Chats that started answering in a JSON "character card" format (a leak in an earlier version) recover by themselves. #
• 2026-10-03 - While the voices download, replies written on your device can now be read by the cloud voice too: the first time a reply needs reading, the app asks, and nothing is sent unless you say yes. #
• 2026-10-03 - The cloud voice now also speaks Chinese, Japanese and French (MeloTTS on Cloudflare), so cloud replies in those languages are read aloud while the on-device voices download. The Log in button left the sidebar; log in or manage your account from the Mine page. #
• 2026-10-02 - The browser's built-in speech is no longer used anywhere. Replies are read by Kokoro on your device; while it downloads, cloud replies in English and Spanish use the cloud voice and anything else is read as soon as Kokoro is ready. #
• 2026-10-02 - While the on-device voices download, cloud replies in English and Spanish are now read aloud by a cloud voice (Deepgram Aura on Cloudflare, through SkillSafe; free), matched to the character's sex. Replies written on your device are still never sent out to be read; they, and other languages, are read by Kokoro as soon as it is ready. #
• 2026-10-02 - Reading aloud no longer waits for the voice download: while the Kokoro voices download, replies are read by your browser's own on-device voice, and Kokoro takes over when it is ready. #
• 2026-10-02 - Downloads no longer hold up the chat. Voices, speech recognition and the picture model download in the background once you press Download, and the chat model shows no progress card while the cloud is answering; follow any download in AI Models, where Stop using the cloud also lives. #
• 2026-10-02 - "Models & data" is now called AI Models, and it shows the progress of a chat-model download started from the chat. #
• 2026-10-02 - No more waiting for the 2 GB download: once the download is running and is measured to take more than about 10 seconds, the chat offers to use the cloud while waiting. Say yes and replies come from Gemma 4 26B on Cloudflare, through SkillSafe, until the on-device model is ready; then the same chat carries on on your device. It is free (the app pays from its sponsored budget), needs no sign-in, is off unless you turn it on, and cloud replies are marked "cloud". What each cloud turn sends is listed on the privacy page. #
• 2026-10-02 - The Mine page now lists all your characters in one place and shows where each one is saved: this browser, your account, Community. Each card's menu holds only the actions that apply - Edit, Save to account, Remove from account, Remove from Community, Delete from this browser, Delete everywhere. Delete everywhere now also removes the Community copy (a shared character used to reappear from there), and a character you deleted no longer comes back from another browser or an old tab. #
• 2026-10-02 - Manage the characters saved in your account: the account window lists them with Edit and Delete. Delete asks whether to remove a character everywhere or keep it only in this browser, and characters kept only here can be backed up again. A delete now sticks: another browser no longer brings the character back, and a character deleted elsewhere stays in this browser as "only in this browser" instead of disappearing. The newest edit of a character wins across browsers. #
• 2026-10-02 - Editing one of your earlier messages now continues the story from there: a new chat opens that ends at your edited message, the character replies to it right away, and the original chat is kept unchanged. Editing your latest message gets a fresh reply in the same chat. Editing a character's message still just corrects it. #
• 2026-10-02 - Characters can roll dice (a d20, 2d6+3...), flip a coin and pick at random from options you give. The roll happens on your device with fair random numbers, plays out above the reply so you can see it happen - dice spin like a slot reel, and for a coin or a random pick every choice is shown while a light runs across them, slows down and stops on the winner - and the character answers with that exact result. With reduced motion turned on, the dice show the result without spinning. #
• 2026-10-02 - Characters can now check the clock. When you ask the time, the date, the time in another city or how many days until a holiday or birthday, the chat model calls a small tool that reads your device's clock, and the character answers with the real figures. It all runs in your browser and nothing is sent anywhere. #
• 2026-10-02 - Optional login, which only backs up your characters (settings and portraits) to a private file in your own SkillSafe account, so they come back on another browser. Chats, memories and personas are never sent. The Mine tab now has three sections: In this browser, In my account and Recommended. #
• 2026-10-02 - New chats open with a greeting the character writes fresh each time, in your language, in the spirit of its saved greeting - so two new chats with the same character no longer start word for word the same, and you are not greeted in English when you use another language. #
• 2026-10-02 - Community: share a character you made, and browse and add characters others shared, ranked by how many times each is loaded. Only the character's settings file is shared (name, tagline, description, greeting, appearance, category, voice, portrait); chats, memories and personas never leave the device. Sharing needs a SkillSafe sign-in; browsing and adding do not. #
• 2026-10-02 - Traditional Chinese (繁體中文): the interface, character replies, Auto-fill and the Chinese voices. Browsers set to Taiwan, Hong Kong or Macau Chinese start in it, and speech input keeps Traditional characters for it. The language is now picked from a small globe at the bottom of the sidebar. #
• 2026-10-02 - The app now uses seven languages, the ones its chat, speech and voice models all handle well: English, Chinese, Japanese, Spanish, French, Italian and Portuguese. Other languages are answered in English. Speech input now favours the language you are already using (it used to hear Ukrainian as Russian and Slovak as Czech), and Chinese speech is written in Simplified characters unless your browser is set to Taiwan, Hong Kong or Macau Chinese. #
• 2026-10-02 - Kokoro now also speaks Japanese, Spanish, French, Italian, Brazilian Portuguese and Hindi, with that language's own voices, matched to each character's sex. Each language's pronunciation data downloads only when a reply in it is first spoken. #
• 2026-10-02 - Chinese replies are now spoken by Kokoro, with seven Chinese voices matched to each character, instead of the browser's built-in voice. The Chinese pronunciation is worked out on your device. #
• 2026-10-02 - Characters now reply in the language you use, and switch when you do. Auto-fill writes the card in your language, speech input detects the language you speak instead of assuming English, and replies in other languages are read by the browser's on-device voice for that language (Kokoro speaks English only). #
• 2026-10-02 - Auto-fill on the Create page and Find with AI for voices now run on the chat model in your browser, the same one that writes the replies. They no longer send anything to a server, need a sign-in or use credits. Auto-fill no longer searches the web, so it writes from what the model already knows. #
• 2026-09-29 - The microphone is now a switch. On, it listens hands-free: you talk, the character answers aloud, and it listens again once the reply ends. Off, the microphone is closed and the browser's recording light goes out. Leaving a chat turns it off. #
• 2026-09-29 - Renamed from Ours Alone to Local Character, at a new address, local-character.skillsafe.ai. Chats saved at the old address stay in that browser storage; backups exported from Ours Alone still import. Added the about, comparison, methodology, privacy and terms pages. Discover now says up front when a browser cannot run the chat model. #
• 2026-09-29 - Chat replies moved to Google's LiteRT-LM runtime: mean reply time fell from 17.8 s to 8.1 s on a 35-turn test. Browsers without WebAssembly JSPI (Safari, Firefox, Chrome before 137) no longer get chat. #
• 2026-09-24 - Talk instead of type: speech input with Whisper base, on the device. One tap records one turn, and the character answers aloud. #
• 2026-09-24 - First release, as Ours Alone: on-device character chat with 55 characters, pictures and voices. #

Corrections policy

Speed and size claims are re-measured when the models or runtimes change, and the methodology page shows the date of each measurement. Report an error to local-character@skillsafe.ai.

Open Local Character →