AI AgentComing soon · Early access

Chat with your second brain.
On your device.

An optional AI agent that lives inside memrynote. Turn it on or off anytime. Local-first, BYOK, MCP-native. Your vault never leaves your machine with local models, and every write needs your nod.

memrynote notes page showing a markdown note with properties, linked context, and vault structure
Architecture, not magic

Built for trust.
Not for autoplay.

We are designing the agent the way we would want one. AI stays optional, open backends and local options are first-class, and there is a stop button before any write.

Chat with citations

Every answer points back to the source notes. Click a citation, jump straight to the line.

Coming soon

Approval-gated writes

The agent proposes the edit. You approve, decline, or always-allow for the conversation.

Coming soon

BYOK providers

Bring your own Claude, Codex, Ollama, or OpenAI-compatible backend only if you want AI on. Keys stay in the OS keychain.

Coming soon

MCP-native vault

One localhost Vault MCP server runs in memrynote. External MCP clients can read it. Writes still need you.

Backends · MCP · Permissions

Your model.
Your server. Your audit log.

Pick a backend. Plug into the local Vault MCP server. Watch every permission decision in a visible audit log.

BYOK backends

Bring your own model.

Switch providers per conversation. Settings persist with the chat, not the composer.

  • Claude CLI

    Anthropic, via local CLI

  • Codex CLI

    OpenAI Codex, first-class

  • Ollama

    Local models, zero cloud

  • OpenAI-compatible

    vLLM, LM Studio, custom

Keys live in the OS keychain. Never bundled, never synced.

Vault MCP server

One server. Many clients.

memrynote runs a localhost MCP server. Claude CLI, Codex CLI, and your own tools can connect.

Running · localhost:7421Read-only default
  • read_noteallowed
  • search_vaultallowed
  • write_noteneeds approval
  • delete_noteneeds approval

External MCP clients are read-only by default. Writes route through an active memrynote Agent conversation.

Permission audit log

Every read. Every write.

A persistent log of what the agent and connected clients did, with one-tap revoke.

  • 14:02Claude CLI read [[summary-2026]]Allowed
  • 14:02memrynote Agent wrote [[summary-2026]]Allowed
  • 13:58Codex CLI searched "pkm"Allowed
  • 13:51You declined delete [[draft-old]]Declined
  • 13:44memrynote Agent read [[journal/2026-05-15]]Allowed
Approval flow

Diff first.
Write only on approval.

When the agent wants to edit a note, you see the exact diff first. Approve, decline, or hand it the keys for the conversation.

Pending writeconversation · pkm-summary

Update [[reading-2026]]

The agent proposes three line changes. Nothing has touched disk.

reading-2026.md+2 / -1
-3. Read more PKM books this year
+3. Build a personal slip-box from PKM reading
+4. Linked: [[zettelkasten-method]], [[linked-thought]]
Conversation settings

Persist with the chat.

Provider, model, reasoning level, and temperature live on the conversation, not the composer.

  • ProviderClaude CLI
  • Modelclaude-opus-4-7
  • Reasoningmedium
  • Temperature0.35
Connected MCP clients

Three agents, scoped.

  • Claude CLIread · write (this convo)
  • Codex CLIread only
  • Ollama (local)read only
Local-only mode

Run a local model.
Vault never leaves your machine.

Pair memrynote with Ollama, llama.cpp, MLX, or vLLM when you want AI on. Switch one toggle and the agent stops talking to any network. Inference is yours.

Ollamallama.cppMLXvLLM
Local mode active0 outbound

Inference

localhost:11434

ollama · llama3.1:70b

Vault

~/memrynote/Vault

on-device only

$ netstat -an | grep ESTABLISHED
— no remote endpoints —

The same approval gate still guards every write. Local does not mean unattended.

Structure

Six guarantees.

The hard architectural commitments behind the chat window.

Settings persistence

Provider, model, reasoning level, and temperature stay on the conversation. Pick up where you left off.

Coming soon

MCP discoverability

The Vault MCP server exposes a clear capability list. Any compliant client can list, scope, and call.

Read-only by default

External MCP clients get read access only. Writes require an active memrynote Agent conversation and your approval.

Approval audit log

Every read, write, and decline lands in a persistent log. Revoke a client in one tap.

Coming soon

Cite-with-context

Answers ship with [[wiki-link]] citations and the surrounding line. No hallucinated sources.

Coming soon

Per-vault permissions

Different vaults can grant different scopes. Work vault stays sealed off from your personal one.

Coming soon
One workspace

Works with the rest of memrynote.

The agent is not a separate app. It is a layer over the same vault that powers your inbox, journal, notes, and tasks.

Reads the vault

Inbox, journal, notes, tasks — all reachable through the local Vault MCP server.

Writes with approval

Drafts edits as diffs you can accept. Nothing touches a file before you say yes.

Never leaves your device

With AI on and a local model selected, your vault and prompts stay on-machine. No cloud round-trips.

Use cases

Built for the way you actually work.

Drafters

Outline, rewrite, polish. The agent reads your prior notes and matches your voice.

Researchers

Ask a question. Get cited [[wiki-links]] back. Jump straight to the source line.

Builders

Draft specs, ADRs, runbooks. The agent proposes a diff; you approve the change.

ADHD brains

Capture now, organize later. The agent suggests where it belongs after the fact.

And more

Wait — there's more.

Citation hyperlinks

Every cited [[wiki-link]] jumps to the source note, scrolled to the line.

Write preview diff

See every proposed line change before disk hears a thing.

Conversation history

Threads persist with their provider, model, and reasoning settings.

Per-vault permissions

Different vaults, different scopes. Work and personal stay separate.

MCP server logs

Inspect what each connected client called, when, and how.

BYOK encryption

Provider keys live in the OS keychain. Never bundled, never synced to our servers.

FAQ

Agent, answered.

Coming soon · Early access

Your second brain. Your model. Your approval.

Local-first. BYOK. MCP-native. The agent ships when the architecture is right.

Believers get early access · Late 2026