Memory & MCP
ProGive your AI a memory you can check.
ADHD makes it hard to carry context from one conversation to the next. Connect Claude, Cursor or another MCP app to Squirrel Notes. It can read and write your notes and keep short memories of people, projects and preferences. You review what it saves, and everything stays end-to-end encrypted.
How it works
The AI app talks to a small server on your computer.
The Squirrel Notes MCP server runs on your machine. It unlocks your notes there with your passphrase, then gives your AI app tools to search, read and write. Our server stores the encrypted result and never sees the text.
Claude, Cursor or another MCP app
Asks for what it needs and saves memories.
Sends plain text, on your computer only
Squirrel Notes MCP
Encrypts and decrypts with your passphrase, and ranks memories with a local model. Runs with npx.
Sends ciphertext only
Squirrel Notes server
Stores what it can’t read. Holds no key.
Memory
Small facts, kept up to date.
A memory is one short statement about a person, a project or how you like to work. Your agent saves them as it goes, so the next conversation starts with context instead of a blank page.
- People and things
- Project state
- Preferences
- Decisions
- Mental models
- References
Your agent writes
It saves new memories, and when something changes it replaces the old one. The old memory is kept and linked, so you can see what changed and why.
You review
Everything an agent writes is marked for review until you check it. Mark it reviewed, edit it, or delete it. Only you can delete.
You pin what matters
Pinned memories form a short profile your agent reads at the start of a conversation. Only you can pin.
Search by meaning
Finds the memory even when you use different words.
We recommend the @squirrelnotes.app/mcp-embeddings package. It is the same server with the same tools, plus a small language model that ranks memories by meaning as well as by the words in them. The model runs on your computer.
Orders a flat white with oat milk.
People and things · Priya
Priya leads the 2.0 launch. Weekly sync is Thursday at 10.
People and things · Priya
Team prefers the café on George St for offsites.
Preferences
Word search alone misses the first result: it shares no words with the question.
| Package | mcp-embeddings | mcp |
|---|---|---|
| Search | Meaning and words | Words only |
| Download | About 17 MB, plus a 301 MB runtime | — |
| Intel Macs | Words only | Words only |
| Tools | Same | Same |
Switch any time by changing one line in your config. The vector file is rebuilt when needed.
No text is sent to be embedded
The model (all-MiniLM-L6-v2) ships inside the package and runs on your CPU. Nothing goes to Squirrel Notes, HuggingFace or an embedding API, and it works offline.
Vectors are encrypted too
Embeddings can hint at what a memory said, so they get the same care: kept in one file in your OS cache, encrypted with AES-256-GCM under your data key, never sent to the server.
Checked before it loads
Every start checks the model files against a pinned manifest. A file that doesn’t match is never loaded.
Falls back quietly
If the model can’t load, search uses words only, as the base package does. On Intel Macs the runtime isn’t available, so that is what happens there.
What your agent can do
Your whole library, not just Memory.
The same server covers notes, collections, tags, sharing and attachments, so your agent can file a meeting summary or turn a chat into a task list.
Notes
Search titles or full text, read, create, update, append, tag, pin and move notes.
search_notesget_notecreate_noteappend_to_noteMemory
Find memories, save new ones, replace outdated ones and read your pinned profile.
search_memoryremember_memorysupersede_memoryget_memory_profileCollections and tags
List, create, rename and nest collections; manage tags.
list_collectionscreate_collectionset_note_tagsSharing and files
Publish a note as a passphrase-protected link and attach local files.
share_noterevoke_shareupload_attachment_from_pathPrivacy
What stays where.
Your passphrase stays on your computer
The MCP server derives your key locally. The passphrase never crosses the network.
Our server stores ciphertext
Notes and memories are encrypted before they leave your machine, including a memory’s category, review state and history.
Your AI app sees what it reads
Text the agent reads goes to that app and its model provider, under their terms. Connect only apps you trust with those notes.
Protect your MCP config file
The config holds your passphrase. Anyone who can read it can unlock your notes. Revoking the API key in Settings ends a running session.
Runtime telemetry is off
The embeddings package turns off ONNX Runtime’s usage reporting on macOS and Linux. On Windows the runtime reports through Windows event tracing, which follows your Windows diagnostic-data settings.
Set up
Three steps, about five minutes.
Upgrade to Pro
Memory and MCP access are part of Pro.
Create an API key
In the app, open Settings → API & Data.
Add Squirrel Notes to your AI app
Paste this into your MCP config, for example Claude Desktop’s, and restart the app. ONNXRUNTIME_NODE_INSTALL stops the runtime downloading GPU add-ons it doesn’t use.
{
"mcpServers": {
"squirrel-notes": {
"command": "npx",
"args": ["-y", "@squirrelnotes.app/mcp-embeddings"],
"env": {
"SQUIRREL_API_BASE_URL": "https://api.squirrelnotes.app",
"SQUIRREL_API_KEY": "sqn_<your_api_key>",
"SQUIRREL_PASSPHRASE": "<your_passphrase>",
"ONNXRUNTIME_NODE_INSTALL": "skip"
}
}
}
}