A·R·I Server v0.8.3

Changes since v0.8.2:

• On Discord, you can now send ARI an image and ask her to redraw or restyle it. She uses your image as the reference, and when she presents the result it arrives as an image attached to a Discord message.
• Release tags are now plain versions like v0.8.3.

A·R·I Server v0.8.2

Changes since v0.8.1:

• The A·R·I icon no longer has a grey border around it.
• The installer is now its own app in its own repo, and installs both the server and the desktop app. Get it from github.com/ari-labz/Installer. Releases here are only the server itself.

A·R·I Server v0.8.1

Changes since v0.8.0:

• ARI can now start subagents: small background helpers that work on a side task with a chosen set of her tools while she carries on, or while she holds her reply open until they report back. Their results show as chips you can open.
• ARI can DM anyone on Discord, wait for the answer, and close the conversation when it's done. The DM conversation gets a brief so it can answer follow-up questions, reports back to the chat that started it, and wakes that chat if it has gone quiet. If nobody replies for 30 minutes it closes itself and tells her.
• Tighter privacy. Guests on the web and people in server channels no longer get your memories, the memory, calendar, persona, project, git or GitHub tools, or push notifications on your phone. Her refusals are one short line in her own voice, and when someone says you've OK'd something she DMs you to check instead of saying she can't verify it.
• The status line follows what she is doing right now: Reading, Thinking, Researching, Typing, Generating, Waiting or Working.
• Images you attach are shown to her on the turn you send them (up to four, on a vision-enabled server), and attachments sent while she is replying arrive with the message. A message sent while she is replying now appears once, at the point you sent it.
• Git calls show as chips in the chat, and a failed one shows as an error. A rejected GitHub token is now reported with what to do, instead of "Invalid username or token", and public repos still work.
• GitHub connects from the control panel, commits ARI makes are signed as Author or Co-author per your setting, and destructive git commands ask for approval first.
• File safety: write_file only creates new files, edit_file no longer refuses a file she has just read, and the server's file tools stay inside the project. Reading anywhere on your machine is for the desktop app only and needs desktop 0.4.1.
• Project fixes: server-side projects are no longer treated as remote, desktop chats keep their project binding, and a project can be bound straight to an absolute server path.
• She no longer announces a step and stops, only says she did something when she can see she did it, and the Coder stops drafting code inside its thinking.
• Engram writes conversation log entries and one commit per note, with tighter prompts.
• Truncated replies are flagged in the UI, reasoning is kept when she is redirected, and pasted images with the same name no longer collide.
• Thinking budgets are switched off for now, while we test whether reasoning effort alone stops overthinking.

A·R·I Server v0.8.0

Changes since v0.7.0:

• ARI now has its own calendar in place of the generic scheduler, and modules can be started/stopped live from the control panel instead of requiring a restart.
• Voice training (StyleTTS2): checkpoints now save into an organised Checkpoints/ folder, the learning-rate schedule and the diffusion/joint-training phase boundaries now correctly persist across pause/resume instead of resetting, and a rare per-epoch validation line no longer gets silently dropped from the training graph. Retrain now genuinely starts from a blank slate (checkpoints, log, and TensorBoard history all cleared) while keeping the dataset and epoch target.
• ARI no longer crashes entirely if a voice model fails to load — it now falls back to running with no voice loaded, same as the existing "no voices found" case.
• Voices tab: the active voice is now highlighted directly instead of a plain badge, and a voice with a saved training run but no trained model yet stays visible instead of disappearing.
• Restored a file explorer for a project's server-side folder — browse, open, upload (including drag-and-drop), and delete files directly from the Projects screen.
• Ari's background dreaming no longer hammers a stopped LLM server with connection-refused retries every few seconds — it backs off after repeated failures and skips entirely while its server is deliberately turned off.
• WebSocket endpoints for the client and listener now require authentication.
• Fixed reminder timing bugs and reminder creation now rejects invalid dates instead of silently accepting them.
• Ari now sends a push notification when she replies to a thread nobody is currently watching.
• Capped Engram's delete-gate so a stuck sweep can no longer leak scratchpads indefinitely.
• Image generation is now gated behind a vision review pass before the result is shown to you.
• Sensitive content is now marked in place with a callout on the ordinary note instead of being split into a separate note.

A·R·I Server v0.7.0

Changes since v0.6.0:

• Engram (the background process that files conversations into Ari's memory) redesigned around three short, focused steps instead of one long wandering pass — a sweep that used to take 15+ minutes now finishes in a fraction of that.
• Background memory work (Engram, Refactor, Curiosity) now runs on a real priority queue and always yields to a live conversation — it only fills genuine idle gaps, never delays a reply.
• Private or sensitive memories are now kept in their own separate, linked notes rather than mixed into an entity's normal note — Ari won't bring one up in conversation unless you go there first, and it's never surfaced when talking to anyone other than the owner.
• Fixed a bug where a fully-written reply could occasionally be discarded and silently regenerated from scratch after it had already finished streaming.
• Fixed Engram getting permanently locked out of reading any brain note over 100 lines partway through a sweep.
• Curiosity and ProactiveMessage are gone, replaced by Dreaming — every agent-initiated thread now sends a push notification.
• Logs restructured: ChatLogs (human-readable transcripts), SystemLogs (per-run copies), DTILogs (full reasoning/tool trace, renamed from Sessions), DreamLogs.
• Image generation: added img2img support (reference images + denoise control), removed all video generation, images attached to messages are now available to image-gen tools automatically.
• DTI panel: completed sessions (like Engram sweeps) now show under a collapsible section; live auto-refresh when threads are created or deleted.
• Fixed a couple of chat rendering bugs (images appearing twice, a spurious "connection closed" message on minimize).

A·R·I Server v0.6.0

Changes since v0.5.0:

• Reasoning effort control on the composer. For models that support it (e.g. Qwen 3.8), a thinking dial appears next to the message box with three levels — Low concludes thoughts quickly, Medium allows longer reasoning, High allows a nearly unbounded chain of thought for the hardest tasks. The chosen level both steers the model and scales its thinking-token budget.
• The dial only shows when the model behind the current pipeline supports it — the dialogue model in Default mode, the coding model in Code mode.
• Support is detected automatically from the model's chat template; no configuration needed.

A·R·I Server v0.5.1

Changes since v0.5.0:

• Reasoning effort control on the composer. For models that support it (e.g. Qwen 3.8), a thinking dial appears next to the message box with three levels — Low concludes thoughts quickly, Medium allows longer reasoning, High allows a nearly unbounded chain of thought for the hardest tasks. The chosen level both steers the model and scales its thinking-token budget.
• The dial only shows when the model behind the current pipeline supports it — the dialogue model in Default mode, the coding model in Code mode.
• Support is detected automatically from the model's chat template; no configuration needed.

A·R·I Server v0.5.0

Changes since v0.4.2 (released as v0.5.0):

• Web search via a self-hosted SearXNG instance — ARI provisions and starts the Docker container automatically on first launch, no configuration needed. Results are filtered for relevance before being returned to the agent.
• Web page reading replaced with a local Mozilla Readability extractor (SmartReader) — no external service required, no rate limits.
• Research agent now tracks a search/read budget to prevent runaway tool loops.
• SearXNG secret key is generated randomly on first run; a placeholder key is replaced automatically on upgrade.

A·R·I Server v0.3.2

Changes since v0.1.10:

• Breaking: config no longer lives in a PersistentData subfolder — servers, models, agents, persona, scheduler and coding-conventions state now sit directly under the server's app-data root. Existing 0.1.x installs will not find their old config here and start clean.
• Agent prompts moved out of code and into Agents.json, editable from a rebuilt Agents tab — every prompt each agent sends, grouped by agent, with the tokens it uses explained above each field. Bindings (which server and slot an agent runs on) are now separate from the prompt itself and default to the first server if unset.
• Discord's token, whitelist, watched channels and allowed guilds are now configured from the control panel instead of environment variables and a config file.
• The scheduler runs jobs at their actual scheduled time instead of firing everything that was missed while ARI was off; each job can be stopped from the panel mid-run.
• A fresh install now seeds a demo server and a small model (Qwen3.5-9B) instead of starting with empty lists — nothing is downloaded or started automatically. The chat composer greys out with an explanation while no model server is running, instead of accepting a message nothing can answer.
• Removed an unused second code-execution path (the old Coder agent), folding its still-used pieces (coding-conventions injection, thread budgets) into the agent that now does all the work.
• Voice: StyleTTS2 no longer reinstalls its Python dependencies on every launch, and its diagnostic output no longer floods the log at warning level.
• Removed personal/example data that had leaked into shipped default prompts.

A·R·I Server v0.3.1

Changes since v0.1.10:

• Breaking: config no longer lives in a PersistentData subfolder — servers, models, agents, persona, scheduler and coding-conventions state now sit directly under the server's app-data root. Existing 0.1.x installs will not find their old config here and start clean.
• Agent prompts moved out of code and into Agents.json, editable from a rebuilt Agents tab — every prompt each agent sends, grouped by agent, with the tokens it uses explained above each field. Bindings (which server and slot an agent runs on) are now separate from the prompt itself and default to the first server if unset.
• Discord's token, whitelist, watched channels and allowed guilds are now configured from the control panel instead of environment variables and a config file.
• The scheduler runs jobs at their actual scheduled time instead of firing everything that was missed while ARI was off; each job can be stopped from the panel mid-run.
• A fresh install now seeds a demo server and a small model (Qwen3.5-9B) instead of starting with empty lists — nothing is downloaded or started automatically. The chat composer greys out with an explanation while no model server is running, instead of accepting a message nothing can answer.
• Removed an unused second code-execution path (the old Coder agent), folding its still-used pieces (coding-conventions injection, thread budgets) into the agent that now does all the work.
• Voice: StyleTTS2 no longer reinstalls its Python dependencies on every launch, and its diagnostic output no longer floods the log at warning level.
• Removed personal/example data that had leaked into shipped default prompts.