Changes

Since v0.3

Memory

• A loaded model and the image server can stop by themselves when nobody uses them, so they don't sit on your memory for days: Settings tab, Unload after (Models) and Pictures idle (Programs), in minutes. Unused means no answer, no Writing and no agent run in that time. Both load again when needed. Off ("never") until you set them.

Passwords

• Entries have a type, the top folder: web, mail, social, shopping, finance, gov, games, media, app, wifi, pc, server, db, vpn, crypto, other. The list is grouped by type with headings that fold away, the menu above it shows one type, and the form has a type menu.
• More → Import CSV brings in a browser or password-manager export: the logins are listed without their passwords, each with a type guessed from its site, to untick, retype or rename; then they are encrypted and committed at once, and the CSV can be shredded.

Pictures

• The persona writes the image model's whole prompt itself: from your one-line request and what the chat is about, 150-250 words with the camera, depth in layers, named light with a warm/cold contrast, and the important things in the middle of the frame (pictures often get cropped at the top and bottom). What you asked for stays as you said it, and there is no text in the picture unless you ask for it. Same image model, much better pictures.
• A persona's own face. Give it a LoRA trained on the Z-Image model (Personas tab, LoRA, with its trigger word; the files go in the LoRA folder, ~/AIcove/loras/). When the persona is in the picture itself, it describes itself as its instructions say and the picture is made with its LoRA, so it looks like itself every time. The image tool needs --lora-model-dir "$AICOVE_LORAS" on its start line.

Sync

• Fetch and push may take up to 15 minutes, for big LoRA files in the AIcove folder.

Chats

• Delete removes a chat for good, with its audio and pictures (no trash any more).
• No more blinking while an answer comes in: only the answer is redrawn, not the whole chat, and you can scroll up and read while it writes; it follows the answer only while you are at the end.
• If an answer finishes while you are reading further up, nothing jumps: the persona's picture shows a ↓, and a click on it goes to the start of the answer.

Writing

• The persona's picture is next to its name in the talk panel, with the ring that pulses while the model thinks, writes or loads.
• After Accept, Decline or Edit the cursor goes on to the next change. Accept all / Decline all ask only until you have decided a change by hand; after that they take the rest.
• New chapter goes at the end, not after the selected one.

v0.3 (2026-10-25)

A lot was reworked since v0.22 (2026-10-19): agents stand on their own, personas keep settings per model, the Writing tab learned to work chapter by chapter, and the Tools tab got groups.

Agents

• Agents stand on their own: a picture, an optional voice, files they read, no persona behind them. In the Chat tab they are below the personas (as @name) and know their reports.
• Leads: a run brings many headlines, each with why and one link, for you to pick from. A lead can follow up an earlier report.
• Research…: one topic in depth, older sources welcome. It shows source cards as it reads, then writes a dossier: what is solid, what is disputed, open questions, proposed articles and all sources. Write articles… has a persona write them from the dossier, facts only from its sources.
• Write it with… starts a Writing project from a report. Both it and Work on it… offer the persona you picked last time.
• Stories need articles from two sites, no front pages, at most three days old; the agent knows today's date.

Chat

• 👍/👎 under answers, counted per persona and model. Again with… answers the last question anew on another model.
• Personas keep their settings per model, one file per machine: Keep for , and Settings for: on the Personas tab. Clean up… forgets models unused for 30 days.
• The model list offers every launcher, the persona's own on top.
• Summarise next to the context meter; read web pages and pictures are counted, so a chat is summarised before it overflows.
• A ring around the persona's picture pulses while the model thinks, writes or loads.
• Models with strict chat templates (Mistral Nemo, Gutenberg) no longer fail with "roles must alternate".
• The box you write in can be made taller, and stays so.

Writing

• One talk per chapter: go to another chapter and you get its talk. New talk starts afresh.
• A first prompt waits in the box on a chapter with no talk yet (first-prompt.txt, yours to change).
• Cast…: tick who is in the chapter, and only their cast entries go to the model, plus anyone you name in your message. Without a pick, the people named in the chapter go.
• When the story memory has to shrink, the chapter notes keep their facts.
• The talk sends as many recent messages as fit, so a long talk never overflows the model.
• A Folder button opens the project's folder.
• Fixed: a rewrite longer than the chapter crashed; many waiting changes froze the window.

Tools

• Groups: a group: line puts a tool under a heading; click a heading to fold it (it shows how many are running). The groups are yours to name.

Everything else

• Sync: the AIcove folder and its stories follow their git remote on their own.
• A SearXNG among your tools is started by the first web search.
• Fixed: AIcove could close when an agent's model came up; Save on a persona could land on the wrong one.

---

Made from CHANGES.md in the repository.