myMem

Local-first AI notes for Mac — your notes, your models. Sign in with the ChatGPT or Claude subscription you already have (OAuth, no API key required); everything is stored locally on your machine.

Features

• Markdown editor with wikilinks and a slash command menu
• Collections, plus AI auto-organize to file notes for you
• Full-text and semantic search (embeddings computed locally, fully offline)
• Version history for every note
• Quick capture global hotkey
• AI chat over your knowledge base with citations — it can also create and edit notes for you
• AI Clean Up with a diff preview before anything is applied
• Dark mode

Architecture

myMem is an Electron app. Notes live in a local SQLite database — FTS5 for keyword search and sqlite-vec for vector search — with embeddings computed on-device by transformers.js (MiniLM). All AI features run through @earendil-works/pi-ai, a provider-agnostic LLM client that supports OpenAI Codex OAuth and Claude Pro/Max OAuth, so chat works on your existing subscription. Nothing leaves your machine except the model calls you explicitly make.

Install

Homebrew (recommended)

brew tap jgwesterlund/tap
brew install --cask mymem           # the app
brew install jgwesterlund/tap/mym   # optional: the CLI for agents

Builds are currently unsigned, so macOS Gatekeeper blocks the first launch: right-click /Applications/myMem.app and choose "Open" (once), or run xattr -dr com.apple.quarantine /Applications/myMem.app.

Download

Grab the .dmg from the latest release and drag myMem to Applications (same Gatekeeper note applies).

Build from source

npm install
npm run dist     # packaged .app/.dmg/.zip in release/ (arm64)

Requirements: macOS (Apple Silicon / arm64), Node >= 22.19. Go is only needed if you want the mym CLI. See the comments in electron-builder.yml for enabling signing/notarization with your own Developer ID.

Develop

npm run dev        # HMR dev server + Electron
npm run typecheck  # tsc over main/preload, renderer, and tests
npm test           # vitest unit/component suite
npm run smoke      # builds, then verifies the whole backend stack inside Electron
npm run e2e        # builds, then drives the real app UI with Playwright (_electron)

MYMEM_SMOKE=1 MYMEM_SMOKE_EMBED=1 electron . additionally exercises the real embeddings worker (downloads the ~23 MB MiniLM model — network required). To sanity-check a packaged build, run the same smoke against ./release/mac-arm64/myMem.app/Contents/MacOS/myMem.

The app icon is generated — npm run icon rebuilds it from scripts/make-icon.swift (see resources/README.md). Design notes and known issues live in docs/.

CLI and agent skill

mym is a small Go CLI that talks to the running app over a local unix socket, so any shell — or any coding agent — can search, read, and write your notes. An agent skill (skills/mymem/SKILL.md) teaches agents the workflow: search first, cite note titles, write markdown.

brew install jgwesterlund/tap/mym   # or: npm run build:cli (requires Go)
npm run install-skill                # installs the skill for every agent detected on your machine

install-skill symlinks skills/mymem into each agent's skill directory (idempotent; --agent claude|codex|pi|opencode|all to target one):

Agent	Skill location	Notes
Claude Code	~/.claude/skills/mymem
Codex CLI	~/.agents/skills/mymem	Codex's user scope; symlinks officially supported
pi	~/.pi/agent/skills/mymem	also reads ~/.agents/skills, so a Codex install already covers it
opencode	~/.config/opencode/skills/mymem	also reads ~/.claude/skills and ~/.agents/skills, so a Claude/Codex install already covers it

All four use the same SKILL.md convention (name + description frontmatter), so one skill directory serves every agent. The installer skips agents it can't find and never overwrites files it didn't create.

From the shell:

mym search "kubernetes notes" --json
printf '## Decisions\n- v2 endpoints stay REST\n' | mym create --title "API design review" --collection Work -
mym append 9f3ac2d4 -- '- follow up on X'
mym related 9f3ac2d4

Also: status, get, list, update, collections, trash — run ./cli/mym help. cd cli && go test ./... runs its test suite.

From an agent, just talk:

• "Check my notes for what we decided about the API design" — the agent runs mym search / mym get and answers with citations from your notes.
• "Save a note summarizing this conversation" — the agent writes a markdown note via mym create into a sensible collection.
• "Add 'rotate the staging certs' to my ops checklist note" — the agent finds the note and runs mym append.

Security

The API is a unix socket with 0600 permissions (~/Library/Application Support/mymem/api.sock; MYMEM_SOCKET overrides) — local-only, current user only, no network listener. There is no telemetry. AI provider credentials are encrypted with Electron safeStorage, backed by the macOS Keychain.

License

MIT © 2026 John Westerlund