Changelog.

New: Vault Manager

A brand-new dedicated workspace for organizing your entire vault with AI assistance.

• Open from anywhere — Use the sidebar icon, command palette, the start menu, or press ⌘⇧M.
• Two-pane layout — See your vault overview (folder tree, note/notebook counts, tag cloud) on the left, and chat with the AI organizer on the right.
• AI-powered organization — Ask the Vault Manager to create folders, move notes, rename notes, or suggest tags. It analyzes your vault and proposes a concrete plan.
• Draft plan with approval — Every proposed change is shown in an interactive checklist. Check or uncheck individual actions, use "Check all / Uncheck all" to toggle everything, or clear the entire plan. Nothing happens until you click "Apply Approved."
• Preset prompts — Four quick-start buttons let you kick off common tasks instantly: organize into folders, fix unclear titles, clean up tags, or set up a weekly review structure.
• Smart folder & note resolution — The plan displays human-readable note titles and folder names instead of technical IDs, even for folders being created in the same plan.
• Safe by design — The Vault Manager never deletes notes or folders, never edits note content, and limits the number of changes per plan.
• Clear chat button — Start fresh anytime with the trash icon in the chat header.

Read all the notes

Improvements

• Sidebar buttons always visible — The quick action buttons (command palette, search, shortcuts, knowledge graph, vault manager) now stay visible even when no tabs are open.
• Welcome screen shortcuts — Added Knowledge Graph and Vault Manager buttons to the start menu for quick access.
• Rename persistence fixed — When the Vault Manager renames a note, the new title now sticks everywhere — in the editor, tab bar, sidebar, and database — without reverting.
• Tag sidebar sync — Tags added through the Vault Manager now immediately appear in the sidebar tag section.
• Vault Manager icon — Uses a distinctive archive icon across the sidebar, tab bar, command palette, and start menu.
• Keyboard shortcut registered — ⌘⇧M opens the Vault Manager, shown in the Settings shortcuts list.

✨ New features

• Added Web Search to the AI assistant, so it can pull in up-to-date information from the internet when needed.
• Added a dedicated Web Search settings section in Settings (separate from AI Models).
• Added support for the assistant to insert web images directly into your notes when relevant.

🧠 Smarter AI behavior

• Improved detection of “latest/current/news” requests so the assistant is much more likely to use web search when it should.
• Better handling of time-based requests (for example, keeping your requested year/timeframe instead of drifting to older info).
• Clearer guidance when Web Search isn’t configured yet (so setup issues are easier to understand).
• Improved image-insert retry flow: if an image result is rejected, the assistant can try again with a different image instead of getting stuck.

🛠️ Stability and polish

• Fixed the empty pane state so “No active tab” is centered correctly.
• Cleaned up a large amount of unused/legacy code to reduce potential bugs.
• Continued reliability improvements around agent tool execution and tool-call loop handling.

⚡ Performance & maintenance upgrades

• Removed legacy Python-web runtime leftovers, making setup/install leaner.
• Updated core dependencies (editor/runtime/build stack) for better compatibility and stability.
• General lint/type cleanup for a smoother development baseline.

This update focuses on making AI Chat much more reliable—especially with smaller local models like SmolLM3—while also cleaning up the chat footer UI.

✨ What’s better

• Smarter AI behavior for smaller models

  • Chat now adapts its internal strategy automatically based on the selected model and available context.
  • SmolLM3 should behave more consistently in multi-step note tasks.

• More reliable note editing with AI tools

  • The assistant is better at deciding when it should actually use tools (instead of only replying in chat).
  • If a request clearly needs note edits, the assistant is now more likely to execute the edit flow correctly.

• Major reduction in “tool call loops”

  • Added protections against repeated tool spam when nothing new is happening.
  • Added protections against “success loops” (repeating the same already-successful tool call over and over).
  • When loop-like behavior is detected, the assistant is pushed to move forward or finish the response.

• More stable chat generation

  • Chat now applies generation settings more consistently, improving response stability and reducing erratic behavior.

• Cleaner, condensed footer under the chat input

  • Context usage now shows a simple used / total (percent) display.
  • Runtime status details were condensed into compact, color-coded chips for faster scanning.
  • Fallback/retry states are easier to see at a glance.

• Improved instruction quality for the assistant

  • Strengthened behavior rules to avoid fake/pseudo tool narration.
  • Better guidance to stop repeating identical actions and finalize clearly when done.

AI writing and editing got much more reliable

• The AI can now edit rich note content much better, including headings, checklists, bold/italic, links, separators, tables, and inline note/link pills.
• Mermaid diagrams and math blocks are handled more safely, with better create-vs-update behavior so the AI is less likely to insert raw code when you wanted a rendered block.
• Wiki links and link pills are now more resilient (including auto-recovery when metadata is missing), so they behave more like manually created links.
• Attachment reading is more robust: the AI is better at resolving the correct attached file even when it only has a visible filename.
• The AI now avoids “fake completion” cases where it claims content was added without actually writing to the note.

Read all the notes

Better AI run experience

• Clearer AI run status/progress visibility while a task is in flight.
• Improved cancellation handling so stopping generation is more predictable and immediate.
• Better persistence guidance and behavior around generated file content.

Voice features improved

• Added local speech-to-text and text-to-speech workflow with better focus-aware behavior.
• Read-aloud now has a toggle-style stop action, so stopping playback is faster and more intuitive.

Notebook experience improved

• Markdown cells now render LaTeX math expressions better.
• Math formatting in notebook markdown is auto-normalized for cleaner output.
• Auto-scroll now keeps AI-edited notebook cells in view after updates.

UI and app behavior improvements

• Update notifications now include direct changelog links.
• File tree/recent activity refresh behavior is more consistent and timely.
• Fixed an issue where switching tabs could incorrectly trigger edit events.

Overall quality

• Large round of reliability hardening for agent tool calls and edge cases.
• Improved guardrails and fallback behavior across advanced edit flows to reduce confusing outcomes.

Better editing for diagrams and formulas

• You can now edit existing Mermaid diagrams directly, and your updates are saved reliably.
• You can now edit inline math more easily through a dedicated editing flow.
• Mermaid and math blocks now show a clear hover edit badge, so it’s easier to discover where to click.

Smarter, more consistent AI chat behavior

• AI chat history is now separated per document, so conversations don’t get mixed between notes.
• AI generation progress is preserved per document when switching tabs, making multitasking smoother.
• Mermaid and math-related AI tools were improved, with more stable behavior and friendlier Mermaid error handling.

UI polish

• File tree and secondary sidebar layout were refined for a cleaner, more consistent look.

New Model: SmolLM3 3B

• Built for 8 GB Macs — A new 1.9 GB model that runs comfortably on base-model MacBooks. It rivals Qwen 3 4B in quality at a smaller footprint, with full support for thinking and tool use.

Better Model Recommendations

• "Recommended" badge is now realistic — The model picker used to recommend the biggest model your total RAM could theoretically hold, ignoring that macOS and other apps are already using several gigabytes. It now accounts for OS overhead and recommends a model that will actually run with comfortable headroom for the KV cache.

AI Agent Improvements

• Agent now writes into your notes instead of just chatting — When you ask the agent to "write a summary" or "draft a project update," it now uses write tools to put the content directly into your note. Previously it would sometimes just dump the text into the chat panel, which wasn't useful.
• Better cross-note research — When the agent searches your notes and needs to synthesize content from them, it now reads the full text of each note before writing. Previously it would sometimes work from 150-character snippets, producing shallow or inaccurate results.

AI Model Loading

• Fixed crashes on machines with limited memory — The AI engine now accounts for how much memory the model itself uses before deciding on a context window size. Previously it only looked at total system RAM, which caused "too large for available VRAM" errors for many users — especially when loading larger models on 16 GB or 24 GB Macs.
• Smarter memory usage — Removed an unnecessary 4x memory overhead from the AI context allocation. Models now load with a much smaller memory footprint, leaving more room for the rest of your system.
• Graceful fallback when memory is tight — If the AI engine can't fit the ideal context window into memory, it will automatically try smaller sizes (up to 3 times) instead of crashing outright. You'll still get a working AI — just with a shorter conversation memory for that session.
• Clearer error messages — If a model truly can't fit on your hardware, the error now tells you plainly that the model is too large, instead of showing a cryptic internal message.

Keyboard Navigation

• Quick tab switching — Press Ctrl+Tab to jump to the next open tab and Ctrl+Shift+Tab to go back. This works across all split panes, so you can cycle through every open note in one smooth loop.

AI Agent

• Smarter note editing — The agent now uses a more reliable find-and-edit workflow when making changes to your notes, reducing the chance of edits landing in the wrong spot.
• Cleaner conversations — The agent no longer leaks internal tool syntax into its responses. Conversations read more naturally, without cryptic bracket codes showing up mid-reply.
• No more 30-second timeout — When the agent proposes a change and waits for your approval, it will now wait as long as you need. Previously it would time out after 30 seconds.
• Better change previews — The diff review panel now shows richer descriptions for insertions and full-content replacements, making it easier to understand what the agent wants to change before you approve.
• Improved search accuracy — When the agent searches your note for a specific section, markup tags are now stripped so searches like "FAQ" match correctly even if the heading has formatting.