AI Chat
Mira's AI chat doesn't just answer questions — it reads, edits, and creates notes in your vault.
How It Works
When you open the chat panel (Ctrl + Shift + C), Mioru sends the AI:
- The current note you have open (full content)
- Sister Notes — the top 30 most recently modified notes in the same project (summaries)
- Project context — the
.project.mdfile if it exists
The AI can then reference your notes, find connections, and take action.
Chat Modes
| Mode | Context | Best for |
|---|---|---|
| Normal | Current project (top 30 notes) | Everyday chat and edits |
| Organization | Current project, broader view | Restructuring, cross-note planning |
| Vault-Wide | All projects (top 60 notes) | Global search, vault-wide changes |
Trigger vault-wide mode by mentioning "all projects", "entire vault", or "alle Projekte".
Agentic Features
The AI can perform these actions on your notes:
Note Operations
| Action | What it does |
|---|---|
| Create Note | Create a new .md file with AI-generated content |
| Update Note | Edit the current note — add sections, reformat, expand |
| Delete Note | Move a note to .trash/ (recoverable) |
| Move/Rename Note | Move a note between folders or rename it |
| Append to Note | Add content to the end of an existing note |
| Multi-Read | Read up to 15 notes in one turn for cross-note analysis |
| Bulk Operations | Create, update, or delete multiple notes in a single turn |
Feedback Loop
Mioru closes the feedback loop between AI intentions and actual results. After each turn:
- XML action tags from previous turns are removed from the AI's conversation history
- Instead, the AI receives a compact confirmation of what actually happened
- This prevents the AI from assuming actions succeeded when they didn't
Example confirmation: "[ACTIONS EXECUTED: note_created: WLAN_Standorte.md, note_updated]"
Sister Notes & Token Efficiency
Mioru uses a three-tier context strategy to keep AI costs low:
| Tier | What | When |
|---|---|---|
| Default | Top 30 sister notes (summaries) | Every turn |
| On-Demand AI | Full note content via list_notes + read_note | AI needs more context |
| On-Demand User | @mention specific notes | User explicitly references |
This is the same principle as Claude Code's tool-based architecture: send only what's needed, fetch the rest on demand.
Web Search
When you have a Tavily API key configured, the AI can search the web live during your conversation. Mioru automatically decides when web search would help — no manual toggle needed.
AI Debug Details
Enable in Settings → System to see expandable error details when AI file actions fail. Debug info shows the action type, attempted path, parsed payload, and the OS error. Debug details stay out of the AI's conversation history so they don't pollute future turns.