Changelog.

✨ New

• Safe Mode for memory-constrained Macs. If your Mac is short on free memory at launch — or if Canto crashed on the previous launch — Canto now boots into a stripped-down but usable mode instead of failing. Notes, search, and the editor all work normally; the heavy AI pieces (chat model, vision, voice transcription, link search, embeddings) stay paused until you opt back in.
• Safe Mode banner. A clear orange banner across the top of the window tells you Safe Mode is on, why it triggered, and how much free memory there was at launch. Three buttons keep the path forward obvious:
  • Try full launch — persists your preference and relaunches Canto in normal mode.
  • Why am I here? — expands an inline plain-language explainer of what's been paused and what your options are.
  • Load chat model — brings the AI back online on demand. When the model you had loaded is too big for current memory, the button automatically swaps to "Switch to <smaller model>" so you don't bounce off a memory error.
• "Switch to a smaller model" suggestion. When the model you're currently using is too big for the memory you have right now, the Memory panel surfaces a friendly suggestion (e.g. "Current Qwen 3.5 9B → try Qwen 3.5 4B") with a one-click Switch button. Picks from models you've already downloaded, picks the largest one that comfortably fits, and stays stable instead of flickering as RAM wobbles.
• Per-process memory breakdown. The Memory modal has a new collapsible Processes section that shows one row per Canto process — main app, each window, the vision worker — with its role, memory usage, and percent of the total, so you can immediately tell which part of Canto is heavy.
• Vision worker has its own live row. When the vision server is running, its memory usage shows up in the Memory dashboard in real time. No more wondering whether vision is "really" loaded.
• Flash Attention and GPU layer tunables. For users who want to fine-tune the memory-vs-speed trade-off under Low Memory Mode, the Memory modal now exposes inline pills: fa (auto / on / off) on the Context window row and gpu (auto / half / all) on the Weight pinning row.

Read all the notes

⚡ Improvements

• Low Memory Mode actually saves memory now. Flipping the master toggle reaches into nearly every memory-hungry corner of Canto at once, typically freeing more than a gigabyte of RAM — often closer to two on machines running a chat model and the vision server side by side:
  • Chat models use a more compact representation of their working memory on supported Qwen 3.5 models. Same answers, roughly half the cost.
  • The vision worker spawns with a tighter context window, a sensible per-model GPU offload, and a more compact cache.
  • Canto's local document database keeps a smaller per-connection cache, and grows back when you flip the switch off — no restart needed.
  • The semantic-search (MiniLM) cache halves at runtime, and entries older than 10 minutes evict on next access instead of pinning forever.
  • When you switch away from Canto, the editor revokes leaked image/video URLs, disposes preview-only editors, and asks the runtime to clean up — so the OS can reclaim memory while you're elsewhere.
  • Diagnostic events flush immediately instead of buffering for batch sends.
• Memory pressure triggers a real cleanup pass. When your Mac itself starts running short, Canto follows up its eviction wave with a forced cleanup so freed memory shows up in the dashboard immediately rather than minutes later.
• The memory total finally matches Activity Monitor. The "Canto X GB" readout in the dashboard header and status bar now sums every Canto process, including the vision worker — instead of only the main app process. The number Canto shows and the number Activity Monitor shows are now the same number.
• Clearer state badges in the Memory dashboard. Each row used to display INACTIVE for two completely different things (the master switch is off, vs. this feature is loaded but resting). Those are now distinct labels — IDLE for sleeping subsystems, LM OFF for "the master switch is off" — with matching dot colors, so you can tell at a glance whether you actually need to flip anything.
• The Low Memory Mode toggle takes effect instantly. Flipping it now fans out across services in real time — the vision worker restarts with new settings, the database resizes its cache, embeddings re-evaluate their cap — instead of waiting for the next operation to pick the change up.
• Chat model auto-fits when GPU memory is tight. If loading the chat model with full GPU offload fails because of memory, Canto retries once with half offload before giving up. A marginal launch that used to fail now succeeds.
• Memory toolbar pills are easier to scan. The Context window and Weight pinning rows wrap their controls cleanly so the new tunables don't push other elements around.

🛟 Why this release matters

If your Mac has felt borderline tight running Canto — beachballs at launch, the AI quietly never loading, or that nagging suspicion the memory readout was lying — 0.5.12 is the release for you. It doesn't ask you to upgrade your hardware; it asks Canto to be a better guest on the hardware you already have.

A small, focused release that makes long AI runs feel less frozen, easier to follow, and harder to lose to overzealous timeouts.

Highlights

• Long tool calls no longer look frozen. When the AI is preparing a big action (a long edit, a structured plan, a meaty lookup), the activity card now shows a live "Preparing… N characters" readout with a ticking elapsed-time counter, so you can tell it's actually working.
• Every step shows how long it took. Each item in the agent timeline now carries a wall-clock duration next to its timestamp, so slow steps stand out at a glance and fast ones get out of the way.
• The AI's reasoning shows up where it happened. Pre-tool thoughts are now interleaved into the timeline at the moment they occurred, right before the tool call they led to, instead of being collapsed into one bundle at the end.
• Heavy edits aren't cut off mid-flight anymore. The safety timers that watch for stuck runs were too aggressive on long, healthy work. They've been widened so a slow-but-legitimate step (a large structured rewrite, a multi-block proxy edit) is no longer mistaken for a crash.

Read all the notes

Improvements

• Tidier activity cards — the state chip (working / done / failed) now sits in the corner instead of pushing the title around, and the timestamp and duration share one clean line.
• Collapsible step contents — when a step touched a long list of blocks or files, you can expand it on demand instead of having it dominate the panel by default.
• Standalone thought messages get their own card, so a pure reflection from the AI doesn't visually merge into the next answer.

Behind-the-scenes

• General housekeeping and dependency cleanup so the app builds cleaner and ships smaller.

A small release that finishes a couple of things 0.5.9 left half-done, plus one big new way of talking to the AI in chat.

The headline: type @ in chat to point at anything

Until now, getting the AI to look at a specific cell, note, or file meant copy-pasting it into the chat or hoping a vague description ("the third code cell") would land. Not anymore.

Type @ in the agent chat box and a dropdown opens listing everything you can pin to your next message. Pick one, and it shows up as a colored pill right inside your sentence — so you can write things like "summarize @note Project Plan against @file design.pdf and what we said in @session yesterday's brainstorm" and the AI knows exactly what you mean.

You can pin as many as you want. A small token counter shows up so long context never sneaks up on you.

Five things you can @ today:

• @selection — whatever text you have highlighted in the editor. The fastest way to ask "what about this?"
• @cell — any cell in the notebook you're in (with code-cell output included)
• @note — fuzzy-search any note in your vault
• @file — any attachment Canto knows about (PDFs, images, audio, video, text files), grouped by type
• @session — a past chat session, so you can pick up a previous conversation without digging through history

You can also send a message with only mentions and no typed text — the AI will figure out what you want from the pinned context alone.

Read all the notes

Slash commands are now on for everyone

In 0.5.9 we shipped the new / command system but accidentally only flipped it on for development builds. The "Try it now →" button on the old Ask AI banner went nowhere for most users. That's fixed: type / in chat and the menu actually opens now. Free users get three commands; Pro unlocks the full catalog and multi-step skills.

Link reading is fast again

If you opened the Memory dashboard in 0.5.9, you may have noticed a yellow caveat next to Link RAG saying "sqlite-vec unavailable — JS fallback." That meant the AI was using a slow fallback every time it read links from your notes. The fast native version now loads correctly in the packaged app. Expect snappier link reading, especially on notes with lots of links.

🪶 Low Memory Mode

• New Low Memory Mode master switch makes Canto comfortable on 16 GB Macs by tightening a dozen things at once — smaller context windows, faster idle unloads, lighter caches, and gentler disk usage.
• New Memory indicator in the status bar opens a live dashboard showing what's loaded, what's idle, how much memory Canto is using right now, and what got freed up recently.
• Friendly tunables (idle timeouts, context size, cache size, link-search results) are shown as named presets instead of raw numbers.
• When your Mac starts running low on memory, Canto now frees things up in a sensible order on its own instead of fighting the system.

⚡ Slash Commands & Skills

• The standalone Ask AI modal is gone. Type / in the chat box to pick a built-in command and run it instantly.
• Build your own Skills — multi-step recipes (for example: read this note, draft a TL;DR, insert it after the title) that always run the same way.
• Each skill step can be marked read-only or allowed to edit, and trusted skills can auto-approve their changes so you don't have to confirm each one.
• Chat sessions now get smart, AI-generated titles so you can find them again in history.
• Delete a single past chat inline with a quick confirm, instead of clearing everything at once.

🔗 The AI Reads Your Links

• The AI can now actually open and read wikilinks and web links in your note — your other notes, web articles, PDFs, and YouTube captions — and use that content when it answers you.
• Ask the AI to "summarize each of these links" and it will, instead of guessing from the link name.
• When the AI uses a link, it shows you which one it pulled from so you can click through and verify.
• A new Link Indexing setting lets you turn the whole feature off, or disable individual source types (web pages, PDFs, YouTube).

🎙️ Voice & Transcription, Reimagined

• Transcription has been rebuilt to run natively on Apple Silicon, so the larger, more accurate Whisper models are now practical to use in real time.
• New Whisper Large v3 Turbo (Korean fine-tuned) model — noticeably better accuracy on Korean speech for native Korean speakers.
• The Voice settings page is redesigned: see all available models with file sizes, switch the active one in place, and watch downloads with a progress bar.
• The Save button is gone — toggles, language, and other settings persist as you change them.
• If your chosen language and model don't match well, Canto now suggests a better one instead of silently doing the wrong thing.
• A clear "Download a transcription model to start" message replaces the silent failure that used to happen when no model was downloaded.

🤖 AI Models

• Qwen 3.6 35B A3B replaces the older Qwen 3.5 35B A3B as the 35B local option. Your selection is migrated automatically.
• Old, retired model files are now cleaned up automatically on startup so they don't keep eating disk space.
• Ollama Cloud models now show an inline indicator so they're easy to tell apart from local ones.
• The intent classifier no longer needs its own dedicated model — that work is now routed through your active chat model, freeing up memory.

Read all the notes

🎨 Improvements

Settings, unified

• Every Settings panel now shares the same buttons, status badges, and stacked model lists, so the whole window feels like one product instead of many.
• The AI Models, Vision, and Voice model lists all use the same clean, scannable stacked-card design.
• Web Search settings save themselves — the Save / Reset Defaults footer is gone.
• The trash icon is now hidden on the Voice model that's currently in use, so you can't accidentally delete the one you're listening to.

Editor

• Paste and Match Style (Shift + Cmd/Ctrl + Alt + V) now works in notes, so pasted text takes on the surrounding style instead of dragging in foreign formatting.
• The little badges and pills that float above blocks no longer jitter or get stuck in a position-update loop in long notes.

AI Chat

• Wikilinks survive AI rewrites — when the AI rewrites a block that contains a wikilink, the link is preserved instead of being flattened into plain text.
• Tooltips on the agent toolbar — the small buttons in the AI panel now describe themselves on hover, so it's easier to learn what each one does.
• A new background watchdog gracefully stops AI runs that get stuck repeating themselves, instead of letting them spiral.

Reliability

• Mermaid diagrams render more reliably — small markup quirks that used to break diagram images have been smoothed out.
• Internal logging is calmer in development builds, and credentials never appear in plaintext in the logs.

🐛 Fixes

• Fixed a position-update loop that could make link pills and badges flicker in long notes.
• Fixed wikilinks being lost when the AI rewrote the block they were in.
• Fixed Mermaid diagrams that wouldn't render due to malformed markup in the generated image.
• Fixed an issue where the clipboard could fail in production builds, breaking the new Paste and Match Style command.
• Fixed transcription falling back silently when no model was downloaded — you now get a clear message and a one-click way to fix it.

What's improved

AI chat history now stays with each note or notebook

You can now come back to earlier AI conversations for the specific note or notebook you were working on. This makes it much easier to continue where you left off instead of starting over every time.

Start a fresh chat without losing older conversations

You can create a new chat for the same note while keeping older chats in history. This is helpful when you want to try a different direction, rewrite something again, or keep separate threads for different tasks.

Chat history is easier to browse

Saved chats now show up in a clearer history view with recognizable titles, previews, and timestamps, so it is easier to find the conversation you want.

You can remove chats you no longer need

If a conversation is no longer useful, you can delete it from the history list to keep things tidy.

Closing Canto feels clearer

When you quit the app, Canto now gives better feedback while it finishes background work. Instead of feeling abrupt or uncertain, closing the app feels more deliberate and easier to trust.

AI-written note content comes back cleaner

When AI rewrites or inserts content into your notes, links, formatting, and code-style sections are handled more cleanly than before, so there is less cleanup afterward.

Canto 0.5.7 is a voice-and-flow update focused on making Live Transcription feel more natural in everyday note-taking, while also making Vault Manager easier to review during longer cleanup sessions.

What's New

Live Transcription feels much more seamless

• You can now stream speech directly into the note while you're still talking, instead of waiting until the end to place the transcript.
• This makes meetings, brainstorms, demos, and quick voice capture feel more like normal writing and less like a separate recording step.
• A compact minimized recording bar lets you keep working without losing track of the active session.

Easier ways to start recording

• Live Transcription is easier to open from the Welcome screen.
• Starting from the Welcome screen now creates a note first, so there is a clear place for your words to go right away.
• A new Cmd/Ctrl + T shortcut makes it much faster to jump into Live Transcription from anywhere.

Better language setup

• You can choose the transcription language before you start recording.
• Canto remembers your language choice for later sessions.
• The language stays fixed during recording, which makes the experience feel more stable and predictable.

Cleaner Live Transcription controls

• The recording controls are easier to read and scan during use.
• Buttons stay more compact and consistent, which helps the panel feel cleaner during demos and real use.

Vault Manager is easier to review

• The Draft Plan area now shows more items at once, so longer plans are easier to scan.
• As new suggested actions appear, the panel automatically follows the latest one.
• Reviewing, approving, and applying bigger cleanup plans now feels smoother and less cramped.

Highlights

More control over AI reply length

• You can now set an Agent output cap in Settings to control how long agent replies are allowed to be.
• You can also let Canto match the active model automatically, so longer AI runs feel more predictable without extra tweaking.
• This makes planning, rewriting, and multi-step editing feel easier to manage than in 0.5.5.

Clearer endings for long AI responses

• Canto now tells you more clearly when a response stopped because it hit its length limit.
• That means it is easier to tell the difference between:
  • the AI finishing naturally
  • the AI simply running out of room
• Live AI status messages also feel more informative during longer runs.

Smarter handling of connected AI models

• Canto now does a better job figuring out how much room a connected model has for replies.
• Model lists show clearer read and reply limits, so choosing a model feels less like guesswork.
• Longer prompts and longer replies behave more consistently across different connected models.

Read all the notes

Better note editing reliability

Structured AI edits land more cleanly

• AI-generated lists, tables, sections, and other structured content are more likely to appear in the right place.
• Multi-part insertions behave more naturally, so related content is more likely to stay grouped together.
• Follow-up insertions after freshly added content are more reliable than before.

Better recovery during larger edits

• If a bigger AI edit runs into trouble halfway through, Canto is better at keeping the progress that already worked instead of making the whole change feel lost.
• This makes complex note updates feel steadier and less all-or-nothing than in earlier 0.5 releases.

Titles and note body changes are handled more safely

• Renaming a note and editing the note itself are handled more cleanly, which reduces awkward mix-ups.
• This makes AI-driven note changes feel more dependable.

Table of contents updates are steadier

• Notes with a live table of contents now behave more reliably when nearby content is inserted or expanded.
• TOC-related updates are less likely to drift or land in the wrong place.

Clearer AI feedback

Easier-to-understand tool and edit feedback

• When an AI action fails, the message is clearer and more useful.
• Complex note-edit actions are summarized in a friendlier way, so it is easier to understand what the AI is trying to change.
• AI progress cards feel more readable and less opaque during structured edits.

Better saved chat details

• Exported or saved AI chat transcripts now preserve more useful context about how a response ended.
• This makes it easier to revisit a run later and understand what happened.

Smaller improvements

• Connected model names and size labels are displayed more cleanly.
• Long AI runs are better aligned with what each model can realistically handle.
• Structured insertions and rewrites feel more stable overall.
• The Agent settings page now feels more complete and useful than it did in 0.5.5.

Canto 0.5.5 is a quality-of-life update focused on making everyday writing, browsing, and AI-assisted work feel smoother and more organized.

What's improved since 0.5.4

Daily Notes feel much more organized

• Daily Notes now have a proper home in a dedicated Calendar view.
• It is easier to jump to today, browse past days, and open notes by date.
• Daily Notes stay more separate from your regular folders, so they feel more like a journal and less like ordinary files.
• Moving around your vault feels cleaner because Daily Notes no longer get mixed into places they do not belong.

The sidebar is cleaner and easier to use

• The sidebar layout has been refreshed to make navigation feel simpler and less crowded.
• Recent notes and notebooks are easier to scan at a glance.
• Switching between your vault, tags, and calendar is more straightforward.
• Common actions are easier to reach, and navigation feels better when moving quickly.

Writing feels more personal

• You can now choose from multiple editor font styles, including clean sans, system, serif, and mono options.
• This gives your notes a writing feel that better matches how you like to think and work.
• There is now a simple spellcheck setting, so you can keep underlines on or turn them off when you want a calmer writing experience.

Exports look more like what you wrote

• HTML and PDF exports now do a better job preserving the look and feel of your notes.
• Your chosen writing style carries through more naturally when sharing or saving notes outside Canto.
• Exported notes feel closer to what you saw in the editor.

AI chat is clearer while it works

• The AI chat panel has been refreshed to feel easier to follow during longer tasks.
• Progress updates are clearer, so it is easier to understand what the AI is doing.
• Step-by-step actions are presented in a more readable, more human-friendly way.
• AI-generated content such as tables, formulas, and diagrams lands in notes more cleanly.

Small polish and reliability improvements

• Notes are less likely to look freshly edited when nothing meaningful actually changed.
• Everyday note and vault behavior feels a little steadier and more predictable overall.
• Background cleanup related to attachments and note updates is smoother and less disruptive.

Read all the notes

In short

0.5.5 makes Canto feel more polished in the moments you notice most: opening Daily Notes, browsing your vault, writing for long stretches, exporting notes, and following along with AI-generated work.