Your Own AI v0.9.0

Your Own AI v0.9.0

Private AI on your own machine - offline by default, online only when you choose.

Highlights

• Your AIs on every computer you own. Sign in with the same Flowsta Vault on a second computer and your AIs, their settings and your conversations arrive there. A conversation started on one device appears on the other; a delete on one removes it from both. (Devices)
• Memory that knows what kind of thing it is keeping. Rules you set, facts, decisions and what happened in a session are kept as separate, dated entries, follow you to every device, and are tidied while the app is idle, with Undo for anything it changed. (Memory)
• Replies keep going when you leave the chat. Ask, go to another page, and the answer is there when you come back; the activity tray shows what is still running and lets you answer an agent's ask from there. (Working in the background)
• Web access per AI. Each AI has its own switch for reading web pages, and a reply that needs the web asks once instead of guessing. (Web)
• Signing in is one step at a time. See plans opens the dashboard already signed in, a closed Vault gets an Open button instead of "get it", and nothing restarts without asking. (Flowsta)
• Online models that keep answering. When a model's provider is busy or down, the question goes to another provider you allow; Settings lists the models routing may use, by provider, and shows where your answers came from. (Online models)

Everything in this release (10 sections)

Devices

• Settings has a Devices section: this computer and your others, a dot for each that is reachable right now (green: its Vault unlocked; grey: locked, asleep or off), when its last change arrived, and what travels between them. A lock on this computer when its own Vault is locked. On the first computer it says how to add another. Just after a start or an unlock it says it is catching up with your other devices instead.
• The check that keeps this device listed for your other devices no longer warns about the open fallback servers: the conductor registers only with the server it started on, so being listed on a fallback is extra reach, said once, not a stalled registration.
• Signing in on another device shows what is syncing in the activity tray: your AIs, their pictures and your memory arrive at once, then one Restart brings in their conversations.
• One Flowsta identity, several computers: a second device signed in with the Vault joins your AIs' private network and picks up their conversations, settings, pictures and hand-given knowledge.
• Each AI keeps its first address on every device, so its conversations list in one place wherever you open them.
• Settings shows which devices this one syncs with and when it last did, or that another device needs the update.
• A conversation from another device opens read-only until it has fully arrived.
• A device you used before signing in joins your identity: its conversations are carried over. It asks before the restart this needs (Restart now or Later).
• When Flowsta's meeting point is down, devices still find each other through an open community node.
• A device that drops off the meeting point, for example after a spell with no connection, puts itself back within a minute and tells its conductor about devices it had not heard of. No restart needed. The conductor's logs from the previous run are kept.
• A new or changed AI, or a new picture, reaches your other devices on its own. Before, only a conversation carried it across, so an AI you had not yet talked to stayed on one device.
• Changes from your other devices show here within a minute or two, not five: each minute the app checks what moved and reads only that.
• Your AIs, their pictures, their knowledge and your document cards now travel to your other devices as a small object of their own, written when they change. The full backup record (every conversation's index, your memory) is rewritten at most every ten minutes, and nothing is rewritten unchanged. Before, a new AI rode a 2 MB record that changed with every conversation, and each rewrite made the other device start its copy again. Settings' "Back up now" always writes the full record. A change to an AI or its picture is backed up after ten seconds, not sixty, so it reaches your other devices in about a minute and a half.
• Documents an AI was given on another of your devices are listed on this one too, marked with the device they are on. Add the same file here and it links itself; until then the document's words stay there. Shown as a plain note, not a warning: "On ", with "Add the file here".
• An AI arriving from another device shows its own picture at once; until this device had set the AI up (up to a minute), its card showed the character's stock face.
• When you switch identity in your Vault, Your Own AI restarts into that identity and opens again. It closed and did not reopen.
• Every request Your Own AI makes to your Vault names the identity it belongs to, so a Vault switched to another identity refuses the request instead of answering as that identity, even in the seconds before the app notices the switch.

Memory

• Your Memory and each AI's memory page open with a row of faces: You (what every AI knows about you), then each AI with how much it keeps for itself. Pick a face to see that memory. When an AI remembers something while working, the step shows whose memory it went into: every AI's, its own, or the project's.
• A note an AI saves for a project that corrects an earlier one (the build command changed, say) takes the old note's place; before, both stood and the next session read both.
• Project memory lists its sessions: what each one left behind, in its own words, with a date. Open one to read it; forget one and the next session starts without it.
• A correction replaces the note it corrects, whatever kind the AI filed each one under. Before, a "make build" saved as a fact left the earlier "make all" command standing beside it.
• Rules, facts, decisions and session notes are separate, dated entries in your own network, sealed to you, and they reach your other devices, even from a device that is switched off (through its backup copy).
• Your rules lead every reply's memory, ahead of the facts.
• Say "remember this" or correct the AI and it is kept at that moment; the remember card says what the new line would replace.
• Projects: a project is the same project on every computer, with "How to work here" in the project window and Save as AGENTS.md; a project's AGENTS.md leads its sessions.
• Dreaming: while the app is idle (and not on battery) a local model dates, splits and tidies your notes. Your Memory shows every standing entry and what dreaming changed, with Undo.
• Each installed model is graded for memory work; the best local one does it, and a small helper model is offered only when none is good enough. No memory work runs on an online model.
• The memory component is now EmbeddingGemma-300M; existing memory is re-indexed once, with progress shown on its card.

Working in the background

• The agent program is v0.4.2: a tool called by its short name (the way small models do) is matched to the one tool server that has it, instead of failing and being looked up again.
• A command that printed its output and then exited with an error code (the last command of a chain, git status outside a repository) shows as "exit 128" on the step, not as a failed step. Red is for a command that printed nothing, timed out or was stopped. The upgrade line counts only those.
• The first question in a project on a model this computer runs no longer ends with "The agent stopped before finishing": when the model was being reloaded with more room, an earlier start of the session could replace the one the question was running in.
• Opening a project the AI has no memory of yet answers at once; it used to read every AI's whole conversation list first (three minutes on a busy install) and forget that it had found nothing.
• A program that asks the AI at :11435 to call one named tool (the OpenAI tool_choice object) now gets the tool call from a model on this computer; the engine understood only auto / required / none and was dropping the request.
• When a model this computer runs gives up on project work - steps fail, or it answers a question about the folder without reading it and says it has no tools - one line above the input says so, with the way up: a bigger model on this computer, or your plan's online models. One door for this, in place of the pop-up and the line in the working rail. The line comes as soon as a second step fails, while the turn is still running; "Let this AI use online models" stops the turn and asks the same question again on the online models.
• A model this computer runs that sends nothing for four minutes in a project (no step, no text, no command of its own running) is stopped, with a note on the turn and the same line above the input. Before, a small model could write one file for ten minutes with nothing to say whether to wait.
• Replies and agent turns keep running when you leave the chat; coming back opens the running reply at once and lands on your question.
• The activity tray lists replies you are not looking at, never repeats the page you are on, and lets you allow or decline an agent's ask.
• Resending a question replaces the old answer instead of adding a second one.

Web

• An AI set to Offline Only no longer searches the web: a search is answered by an online model, so it follows the model setting, not the Web access switch. Reading a page you point the AI at is done on this computer and stays with the switch. When a question needed a search, the line above the input offers the online models.
• A switch per AI for reading web pages; when it is off, a reply that needs the web says so in one line above the message box, with "Just this once".
• When one page is not enough the AI reads on or searches once, and can offer the Build helper for bigger jobs.
• An agent asks before reading an address you did not give it.
• A question about a page you shared is answered in the chat; Dig deeper hands it to an agent with the web allowed.

Online models

• When an online model's provider is busy, erroring or not answering before its first word, the question goes once to a model of the same kind from another provider; the reply's model note says which and why. An AI set to one online model is never switched: its error offers the other model beside Retry.
• Settings › Routing lists the models automatic routing may use, grouped by provider with where each is based. DeepSeek, Moonshot AI, Qwen and Z.ai start off. A model you pick by name is used either way.
• Each card on the Online Models page has the same switch, a Use for menu, and labels for what routing uses the model for.
• A model that is failing for many people right now is left out of automatic picks until it recovers, and its card says so.
• Where your answers came from: Settings shows your recent questions by where they were answered, and this plan period's online cost by kind of model.
• A provider's failure reads as one plain sentence, and Retry asks the question once.

Routing

• A question about now on an AI that can't look things up (Offline Only, or no plan) gets one line above the message box saying the answer may be out of date, with the way to look it up online on a plan.
• The check for questions that need current information was rebuilt for the new memory component and catches far more of them, on every routing setting.
• Settings › Routing has two parts, Offline models and Online models. Choosing a model, health questions and project work on your server show without a plan. The dial's first option is now Online-first.
• In projects, the helper that plans uses the Planning model, and the helpers that change files and run tests use the agent's model.
• Offers of online models link to the Online Models page, say the plan is optional, and put the offline fix first.

Flowsta

• See plans opens yourownai.net signed in as you, with this computer linked to your plan, and the dashboard says who is signed in.
• With the Vault locked, See plans waits for you to unlock it instead of falling back to a separate sign-in in the browser.
• The first time, See plans opens Flowsta's "Share with Your Own AI?" page already signed in; Allow takes you to the dashboard.
• When Flowsta Vault is installed but closed, Settings and Online Models offer Open Flowsta Vault.
• When you disconnect Your Own AI in the Vault, Settings says so instead of quietly signing out.
• The welcome wizard offers an optional sign-in on its last step.
• Clearer sign-in messages that name the fix; the plan card names OpenAI and Meta among the online models.

Engine and models

• llama.cpp v0.7.0 (b11669): the new-model support, leaner embedding memory, and the speculative-decoding fixes.
• New in the catalog: NVIDIA's Nemotron 3.5 Bolt (74B mixture-of-experts, 9B active, 22 GB, 256K context, images; for a 24 GB graphics card or a 32 GB machine) with NVIDIA's own speculative-decoding draft, and Nemotron 3.5 Super VL (120B, 12B active, 1M context, images; for a 96 GB-plus workstation) with its multi-token-prediction head.
• Background work never loads a model, and loads wait for a reply that is streaming.
• Agent sessions load the model once with the context window they need, and retry with a smaller one when it does not fit.
• Model downloads use about 30% less processor per megabyte.
• Each model's reasoning setting follows its own template.
• Project work is no longer given to Qwen 3.8 Distilled 2B, which answers in words where a tool call is needed; the 4B and 9B still take it.

Fixed

• A model whose template always starts a reply by thinking, with no way to turn that off, could spend the whole reply thinking and end there; the page then said the response was all internal reasoning. The app asked it not to think, but that request never reached such a model. A chat turn on one now hands it an empty, already closed thought, so the reply starts with the answer. Reports still think.
• Auto could rank a small model by its family's flagship score and pick it over a better, faster one on the same computer. A model's size now bounds its score: under 3B at most 6 of 10, under 6B at most 7, under 12B at most 8.
• Gemma 4 E2B graded "runs slower" on a 4 GB card it fills to 1.7 GB: its per-layer embedding table, kept in system memory by the engine, was counted on the card. It now grades, and ranks, as the model that runs fully on the card that it is.
• The app log now records each offline model pick for a chat turn, with the score and fit that decided it, and whether the loaded model was kept.
• Going back to the chat page right after watching a reply finish, then pressing New, could bring that conversation back with its question doubled: the page reopened a reply it should only reopen when it finished while you were away, and a read of its records that landed after New still wrote the page. Neither happens now.
• A reply you left with New in its first second came back onto the new page when it finished, and the activity tray did not list it. The reply now stays where it started; the tray says "answered" with Open, and the reply keeps its Model, Tokens and Sources when you open it.
• A question in a project or a tools session stopped at "working" with nothing sent and nothing to stop it. The turn's bubble went up and then the page failed on its next line, before the question reached the agent, so no ceiling and no Stop applied. The page's own errors now reach the app log as well, so a stop with nothing on screen leaves a line that can be read.
• A question could sit at "working" with nothing to stop it: before a prompt is sent, the question is embedded to find the AI's documents, and when the embedding model stopped answering (a chat model had just filled the graphics card and the helpers were being moved) that wait had no end. The wait is now 15 seconds at most; past it the question goes without document passages, and the embedding server's own checks and requests time out instead of hanging.
• AI pictures from another device now show on this one.
• An AI that has no picture on this device takes the one your other devices have, and a built-in face never covers an AI's own picture: its own picture replaces the face when it arrives, after a restore too.
• After a device joins, your AIs show their own pictures straight away (they showed the built-in faces until the page was reloaded).
• The activity tray no longer says "0 conversations" joined your identity when none did.
• Project memory saves and reads in well under a second, even for an AI with a thousand conversations (it used to take minutes on a busy computer), and each AI keeps one memory conversation per project.
• A note the agent saves during a session is no longer written over by the session's summary when the project closes.
• A conversation from another device that can't open yet says why.
• A computer that opened Your Own AI before signing in now offers to join your other devices, and their conversations arrive after the restart.
• macOS: the window comes to the front once the app has started, instead of opening behind other apps.
• Personalities describe how they greet you instead of quoting a line the model then repeated.
• Each bundled personality is listed once.
• Welcome wizard: the AI cards keep their order and names while they are set up, and step 1 continues a download instead of starting it again.
• A reply lost to a duplicated request ends the turn instead of hanging.
• A reply left mid-way with New is recorded in its own conversation, and the next question starts a new one.
• The activity tray lists only replies you did not see finish, and its download rows follow every update to the end (an engine download showed 11% after it had installed). The NVIDIA engine and the Projects helper say when they are installed.
• The NVIDIA engine offer and Settings show the real download size.
• The conversations drawer opens quickly with a thousand conversations.
• Online models are named by their catalog name in the model menu.
• Pausing an online model keeps automatic routing from using it, as its button says.

Diagnostics

• The truth matrix's chat, answer-quality, think-switch and sampling legs now ask a model exactly the way a chat turn in the app asks it, with the same no-thinking shape, so a model the app answers with is not reported as thinking first or running out of room.
• The build's closure-order check also fails on a closure that reads a signal, store or computed value declared below it, the fault behind the stopped project questions.
• The truth matrix report has a Page errors section: the last ten errors the app's own screen code threw, each as time, message and code location. Never your words: anything that looks like a path or an address in a message is blanked.
• Running the truth matrix can no longer run the computer out of memory: each test load starts only with room for it (a model graded too big is not loaded to confirm it when memory is short), and the app's own model waits until the matrix finishes. The Devices section says which identity it describes and numbers your other devices; the report names no device or identity, so it stays safe to share. The helper check uses the routing classifier's real instruction and fails when it reads a message as the wrong kind of task. The tool-call check loads each model graded for agent work itself, and the computer is kept awake while the matrix runs (macOS and Linux).
• The routing classifier and memory extraction answer the same way every time: their fixed-form answers are no longer sampled at random, so a message is not routed one way on one run and another way on the next.
• The routing classifier tells everyday uses of words like debug, code, script or bug (a sourdough starter, a bike chain) from programming, so those questions are no longer sent to a coding model; questions about developer tools like git count as code.
• The truth matrix's Devices section says when the app has not yet checked your other devices since it started, instead of showing the time from before the restart.
• The truth matrix gives each model three minutes in the answer-quality step (a slower one is reported as too slow here, not failed) and shows a line when a question runs long; the keep-awake helper ends with the app however the app ends.
• When the system asks Your Own AI to quit (a logout, a closed terminal), it says so in its log and shuts down cleanly, stopping its model servers. Freeing the chat model's port only ever stops the server listening there, never the app itself.
• Re-indexing your document library (after the memory model changes) no longer holds up questions: it sends four passages at a time, so a question's own lookup waits a moment, not half a minute.
• On a Mac, the helper model starts when there is room by the same memory figure the model loader uses; it was refused with plenty free.
• The truth matrix has a Devices section: the Vault, when this device last backed up its AIs and everything else, how old this Vault's copies of your other devices are, the last check for their changes, and where this device is listed for them. It also asks the loaded model for one named tool the way a program does, and says when the agent program installed is older than the one this app ships.

Downloads

See assets below for platform-specific installers. Verify any download against SHA256SUMS.txt.