Key Concepts
NoteTalker combines a few distinct ideas that work together. This page explains them so you have a mental model before diving into the details.
A vault is a folder on your computer where NoteTalker stores everything. It’s full of plain markdown (.md) files.
NoteTalker reads your vault to answer questions, find relevant notes when you’re chatting, and write new content when you ask it to. But your files never leave your machine unless you choose to sync or share them yourself.
You can open your vault in any text editor, back it up with any tool, and keep it in a Git repository. Your data is always yours.
Notes are .md files in your vault. NoteTalker works with standard markdown, plus a few conventions:
- Wiki-links:
[[Note Title]]creates a link from one note to another - Tags:
#tagadds a tag to any note - Frontmatter: YAML metadata at the top of a file (between
---lines) for structured fields like date, status, or type - Tasks:
- [ ] task itemcreates a checkbox task that NoteTalker can track and toggle
You can write notes manually or ask the agent to create and edit them for you.
The agent is the AI assistant built into NoteTalker. It’s not a chatbot that just answers questions — it can take actions in your vault.
The agent reads and writes files, searches your notes, manages your contacts, sets reminders, creates calendar events, and more. It does this using tools (see below).
When you have a conversation with NoteTalker, the agent is always working behind the scenes: searching your vault for context, deciding which tools to use, and making sure its responses are grounded in your actual information.
Tools are actions the agent can perform. When you ask NoteTalker to do something, it chooses the right tool (or sequence of tools) to get it done.
NoteTalker has 42 built-in tools across several categories (the 14 most-used stay in the model’s context by default; the rest load on demand):
| Category | Examples |
|---|---|
| Notes | Create, read, edit, move, search notes |
| Memory | Remember facts, search memory, update your profile |
| Tasks | Add tasks, track completion, generate a daily report |
| Daily notes | Open today’s note, add to it |
| CRM | Look up contacts, find organizations |
| Calendar | List upcoming events, create events |
| Weather | Get current conditions and forecast |
| Voice | Create voice notes from recordings |
| Scheduling | Set reminders, create recurring tasks |
Tools run locally on your machine — there’s no round-trip to a server to execute them. The agent decides which tools to use, but the execution happens right on your computer.
Memory
Section titled “Memory”NoteTalker remembers things across conversations. This isn’t just chat history — it’s a persistent knowledge base about you, stored as markdown files in your vault, on your device.
Memory is organized into a few layers:
- Facts: Things you’ve told NoteTalker, like your job, preferences, or important context
- Goals: What you’re working toward
- Relationships: Notes on people in your life
- Projects: Active work you’ve mentioned
Over time, NoteTalker also generates a snapshot — a short summary of who you are and what you’ve been focused on lately. This gets included in every conversation so NoteTalker doesn’t start from zero each time.
Memory is stored in atlas/memory/long-term/facts.md. You can read and edit it directly.
Sessions
Section titled “Sessions”A session is a single conversation thread. Sessions have a few important behaviors:
- Auto-resume: If you close and reopen NoteTalker within 4 hours, it picks up where you left off in the same session
- Topics: After a few messages, NoteTalker generates a topic title for the session so you can find it later
- Summaries: When a session ends, NoteTalker writes a summary to long-term memory so important things aren’t lost
- History: The full transcript of each session is saved in
atlas/memory/sessions/as a JSONL file
You can browse past sessions and review what was discussed. Sessions are part of your vault — plain files you can read anytime.
RAG (Retrieval-Augmented Generation)
Section titled “RAG (Retrieval-Augmented Generation)”RAG is the mechanism that makes NoteTalker’s AI responses relevant to your specific notes rather than generic.
Here’s what happens every time you send a message:
- NoteTalker searches your vault for notes related to your message
- The most relevant notes are included as context in the prompt sent to the AI
- The AI answers based on both its training and your actual notes
This search uses a hybrid approach: semantic search (understanding meaning) combined with keyword search, fused together for the best results. It uses a local embedding model — no internet required for search.
The practical effect: if you’ve written a note about something, NoteTalker knows about it and can reference it accurately. The more notes you have, the more personalized and accurate NoteTalker becomes.
How Local AI Works
Section titled “How Local AI Works”NoteTalker’s AI runs entirely on your own computer. The language model behind chat, the speech models for voice input and output, and the embedding model for search all execute on your hardware — nothing you say, write, or ask is sent to a server.
- Runs on your GPU: On Windows and Linux the language model runs on an NVIDIA GPU (CUDA); on macOS it runs on Apple Silicon (MLX/Metal). See Installation for the hardware details.
- Lazy-loaded: The model loads into memory the first time you use it, so launching NoteTalker stays quick.
- Auto-unloads when idle: After about 10 minutes with no activity, NoteTalker unloads the model to free memory, then reloads it on your next message. The first reply after an idle stretch takes a little longer while the model loads back in.
- Nothing leaves the device: Chat, voice input, voice output, and search are all processed locally. The only time NoteTalker needs the internet is a one-time model download, a one-time license activation, and any optional extras you turn on yourself (like weather or web search).
Everything NoteTalker does is covered by a single one-time purchase — see What’s Included.