Changelog.

This release focuses on making model loading more flexible and the Settings experience cleaner and easier to navigate.

More flexible AI model loading

• Added a new “Allow low-memory model loading (unsafe)” switch in Settings → AI Models.
• If enabled, Canto can try loading built-in local models even when you don’t meet the full recommended free-memory target.
• This is intended for advanced users who want to trade safety/performance for flexibility.
• Clear warning text is shown so it’s obvious this mode may run slower or fail under pressure.

Read all the notes

Better AI editing reliability

• Improved how AI applies note edits so multi-step changes are more consistent.
• Added stronger protection for note titles so AI edits are less likely to accidentally break or overwrite them.
• Improved handling of repeated “insert after” operations so AI-generated structure lands in the expected order.

Markdown now renders properly after AI writes

• When AI writes markdown-style text (headings, lists, bold, links, code, etc.), Canto now converts it into proper rich formatting automatically instead of leaving raw markdown text behind in many cases.

Smarter model context sizing (stability/performance)

• Model loading now uses more accurate real memory measurements before/after load.
• Context size is chosen more intelligently based on available memory, reducing avoidable failures and improving load decisions.

New Notes setting

• Added a Notes settings section.
• You can now control how much text the AI reads from attachments (Attachment Read Limit) for better quality/performance trade-offs.

A New Editing Engine Under the Hood

This release ships a completely new way for the AI agent to read and edit your notes. Previously, the agent navigated your note like a text file — counting characters and guessing positions. It now uses a structured block engine that understands your note the same way the editor does: as a sequence of typed blocks (paragraphs, headings, lists, tables, code, math, diagrams), each with a stable identity.

Before touching anything, the agent reads the document's structure first. Only then does it apply targeted, block-level edits — never overwriting content it didn't intend to change.

What you'll notice:

• Edits land in the right place, even in long or complex notes
• Tables, code blocks, math expressions, and diagrams survive edits to surrounding content
• The agent no longer loses its place mid-edit on large notes
• Insertions go exactly where you asked, not one paragraph off
• Write operations are validated before they're applied — bad edits are caught and rejected before they touch your note

---

Significantly More Accurate AI Edits

The new engine enforces a strict read-before-write discipline. The agent must snapshot the document and confirm what it sees before any mutation is allowed. Revisions are tracked, and if the document changed between read and write, the operation is rejected rather than applied to stale content. This eliminates an entire class of garbled or misplaced edits that could occur in previous versions.

This beta introduces Evernote import for the first time, plus key polish to make imported notes feel right in Canto.

New in beta.5: Evernote Import

• You can now import Evernote .enex files into Canto.
• Imported notes keep more of their original structure and formatting.
• Evernote links between notes are restored more reliably, so your notes stay connected.
• Rich Evernote link blocks are handled better during import.

Better Export Experience

• PDF export is now much more reliable for large notes (especially notes with big embedded images).

Link Display & Compatibility Polish

• Wiki links now display more cleanly in the editor.
• Better compatibility with older/imported wiki-link formats.

🧠 AI & Agent Improvements

• Upgraded the AI “Find Notes” behavior to deliver more relevant suggestions.
• Improved vault/note listing behavior used by AI so results are more focused and less noisy.
• Added better run/composer controls in the AI panel so it’s easier to steer responses and workflows.
• General improvements to tool-driven AI reliability for longer or more complex requests.

🤝 LM Studio Improvements

• Selecting an LM Studio model now loads it immediately (instead of “one step too late” behavior).
• Canto no longer forces its own context size into LM Studio requests.
• Canto now follows LM Studio’s runtime context setting and syncs it back into the app.
• Reduced first-message failures caused by too-small default context windows.
• Context indicators are now more trustworthy during active sessions.

🧩 UI & Quality-of-Life

• Context badge on model cards now updates more accurately as runtime settings change.
• Context usage display now aligns better with real available context.
• Improved loaded-model handling so users can free memory more easily when needed.

✨ Highlights

Smarter, more reliable AI editing

• Improved the AI chat assistant so edits are applied more consistently and with fewer “stuck” or repeated actions.
• Improved how AI finds and updates text, especially in longer notes or content with mixed formatting.
• Reduced cases where AI would over-correct itself or retry the same step in an unhelpful way.
• Made multi-step AI edit flows feel more stable and predictable overall.

Better approval flow for AI changes

• Improved edit approval handling so AI tool actions are processed one at a time, making behavior more dependable.
• Removed overly strict selected-text retry behavior so responses feel less forced and more natural.

Tagging and editor usability improvements

• Added/improved manual tag suggestions to make tagging faster and easier.
• Improved tag input and tag suggestion interactions for a smoother note organization experience.
• Fixed selection toolbar positioning so it stays within view more reliably instead of drifting off-screen.

Sidebar and file tree polish

• Refined sidebar and file tree layout/styling for a cleaner, easier-to-scan vault experience.
• Improved overall navigation feel with visual and spacing tweaks in key sidebar areas.

🛠️ Under-the-hood improvements

• Multiple stability and guardrail improvements in AI runtime behavior.
• General UX refinements across editor and note-management surfaces.

What's New

Pin Notebooks

You can now pin notebooks to keep them at the top of your vault — just like you've always been able to pin notes. Right-click any notebook and choose Pin, or hover over it and click the pin icon. Pinned notebooks stay anchored above the rest, sorted by when they were pinned.

Smarter "Get Info" for Notebooks

The Get Info panel (right-click → Get Info) now shows meaningful stats for notebooks, not just notes. For a notebook you'll see the total number of cells, how many are code vs. markdown, what languages are used, and how many file attachments are included along with their combined size.

Cleaner Vault Item Cards

Notes and notebooks in the Vault sidebar got a visual polish pass. Selected rows have a crisper highlight style, tag chips now use the full available width before collapsing into an overflow count, and the inline rename field no longer causes the row to jump or resize when you start editing a title.

Nicer Code Blocks in Dark Mode

Code blocks inside notes look much better in dark mode. The background blends more naturally into the editor surface, the striped highlight artifact is gone, and syntax colors now follow a Monokai-inspired palette — magenta for keywords, cyan for types, green for function names, yellow for strings, and purple for numbers. Light mode is unchanged.

Smarter AI Agent — Structured Content Routing

The AI agent now reliably picks the right tool for structured content. When you ask it to insert a Mermaid diagram, a math equation, a table, or an inline element, it no longer falls back to a plain-text edit — it routes directly to the dedicated tool for that content type. If a model tries the wrong approach, the agent catches it, explains what went wrong, and steers the model to the correct tool automatically.

---

Read all the notes

Bug Fixes & Improvements

• Inline rename in the vault now fits neatly within the row width instead of overflowing or collapsing adjacent content
• Notebook sorting now respects pin order first, then recency — consistent with how note sorting works
• Various internal cleanup to improve startup speed and reduce memory overhead

This beta focuses on making Canto feel cleaner, faster, and easier to navigate day to day.

Editor experience improvements

• The writing space is now cleaner and less distracting.
• The old fixed top formatting bar is gone.
• A Notion-style floating formatting toolbar appears near selected text when you need it.
• The Chat button stays easy to reach in the pane header.
• Text selection and toolbar appearance feel smoother and less jumpy.

Read all the notes

Notebook page refresh

• Notebook headers have been redesigned for clearer visual hierarchy.
• Notebook titles are more prominent.
• Tags are back under the title where they’re easier to scan.
• File controls were moved into a cleaner popover flow with a visible file count.

Major sidebar and vault redesign

• Canto now uses a simpler single-sidebar layout.
• Main sidebar tabs are now Overview / Vault / Tags.
• Vault/File Tree rows are richer and easier to scan (better previews, metadata, and timestamps).
• Hover actions (like pin/delete) are cleaner and less visually noisy.
• Sidebar visual polish: spacing, icon treatment, and divider thickness were refined.

Folder structure simplification

• Folder behavior is now streamlined to a root-level model (no nested subfolders).
• Existing nested folders are automatically flattened on upgrade.
• Name conflicts are handled automatically during migration.
• Move/open flows are simpler and more predictable with the new structure.

Split-pane resizing (new)

• Borders between panes are now draggable, like the sidebar resize handle.
• You can resize panes directly in the editor layout.
• Pane widths are remembered across app restarts.
• Double-clicking a divider resets adjacent panes to an even split.

✨ AI Experience Improvements

• Safer memory handling across platforms so AI features behave more predictably on different machines.
• Cleaner AI status display in settings/model UI, making model readiness easier to understand at a glance.
• Added a persistent “Model Off” mode — if you turn AI off, Canto remembers it and avoids unexpected AI actions.
• Introduced broader AI availability checks so actions fail more gracefully with clearer guidance when AI isn’t ready.

Read all the notes

🌐 External Model Endpoint (Ollama + LMStudio + OpenAI Compatible)

• Expanded support for external AI endpoints with a more complete runtime flow for Agent and Vault experiences.
• Improved endpoint reliability with better model-switch behavior and fewer conflicts when changing providers/models.
• Added incremental model refresh for external endpoints to improve stability and responsiveness.
• Strengthened cross-endpoint unload behavior so old model state is cleaned up properly when switching.
• Included major under-the-hood runtime unification for external endpoints to reduce edge-case failures.

🤖 Agent & Chat Workflow Enhancements

• Agent chat state now persists better: history and docked-panel behavior are preserved more consistently per tab.
• Added a new “Open Agent Chat Panel” action from editors (including notebooks) for faster access while writing/coding.

⌨️ Shortcut Customization

• Added configurable AI and voice shortcuts.
• Shortcut labels now update dynamically across the app (including command surfaces/UI text), so help text matches your actual keybindings.

🏷️ Tags & Note Organization Polish

• Redesigned tag editing with a modern compact chips-first inline UI.
• Notebook tags now use inline compact editing (less popover-heavy flow).
• Note tags now appear in the title area flow (under the note title), with improved spacing and visual separation.
• Improved tag creation visuals with more varied tag color assignment in Vault/tag creation flows.

🛠️ General Stability & UX Polish

• Multiple consistency fixes and architecture cleanups across AI, endpoint handling, and UI interactions to make day-to-day usage smoother in this alpha.