Mnemo Cortex on macOS

July 4, 2026 · View on GitHub

Install guide for macOS 12+ on Apple Silicon (M1–M4) or Intel. Tested by resolving the full dependency tree as prebuilt macOS wheels (Python 3.12/3.13/3.14, arm64 + x86_64) — nothing compiles on your machine. This guide is part of our macOS beta: if anything below doesn't match what you see, please file an issue with the exact command and output.

⚠️ Read this first: which Python you use matters

Mnemo Cortex stores vectors in SQLite via the sqlite-vec extension. Loading a SQLite extension requires a Python whose sqlite3 module was built with extension support — and on macOS, two common Pythons were not:

Python	`sqlite-vec` works?
Homebrew (`brew install python`)	✅ yes
uv-managed (`uv python install`)	✅ yes
MacPorts	✅ yes
❌ Apple system Python (`/usr/bin/python3`, Xcode CLT)	no — extension loading compiled out
❌ python.org installer	no — same limitation

10-second preflight — run this with the Python you plan to use:

python3 -c "import sqlite3; c = sqlite3.connect(':memory:'); c.enable_load_extension(True); print('✅ this Python can load SQLite extensions')"

If you see AttributeError: 'sqlite3.Connection' object has no attribute 'enable_load_extension', that Python cannot run Mnemo Cortex. Install Homebrew Python and use it explicitly ($(brew --prefix)/bin/python3).

Prerequisites

# Homebrew itself: https://brew.sh
brew install python git
brew install ollama          # recommended: free, private, local models
brew install node            # only if you'll run an MCP-bridge integration
                             # (Claude Desktop, Claude Code, LM Studio, ...)

Start Ollama and pull the embedding model (the vector index is locked to nomic-embed-text, 768 dimensions):

brew services start ollama   # keeps Ollama running across reboots
ollama pull nomic-embed-text
ollama pull qwen3:8b         # or any tool-capable reasoning model you prefer

Step 1 — Install

git clone https://github.com/GuyMannDude/mnemo-cortex.git
cd mnemo-cortex
$(brew --prefix)/bin/python3 -m venv .venv    # explicitly Homebrew Python
source .venv/bin/activate
pip install -e .

Run the preflight from the venv to confirm the whole chain:

python -c "import sqlite3, sqlite_vec; c = sqlite3.connect(':memory:'); c.enable_load_extension(True); sqlite_vec.load(c); print('✅ sqlite-vec', c.execute('select vec_version()').fetchone()[0])"

Step 2 — Configure

mnemo-cortex init

The wizard writes ~/.config/agentb/agentb.yaml (data lands in ~/.agentb/). Two macOS notes on the result:

Bind address. The default binds 0.0.0.0 (all interfaces). On a laptop that joins coffee-shop Wi-Fi, set it to loopback unless you know you want LAN access:
```
server:
  host: 127.0.0.1
  port: 50001
```
If other machines need to reach it, prefer a Tailscale address over 0.0.0.0, and set server.auth_token.
Config lives at ~/.config/agentb/agentb.yaml on macOS too (not ~/Library/Application Support). The AGENTB_CONFIG env var overrides the path.

Step 3 — Run and verify

mnemo-cortex start        # background; or: mnemo-cortex start --foreground
curl -s http://127.0.0.1:50001/health
mnemo-cortex status

/health should report the version and green components. Save a memory and recall it:

mnemo-cortex test

macOS uses launchd where Linux uses systemd. A template LaunchAgent ships at deploy/macos/com.mnemo-cortex.server.plist. It runs the server in the foreground under launchd's supervision (do not point it at mnemo-cortex start, which daemonizes and would confuse launchd's process tracking).

# From the repo root, with the venv at .venv — substitutes your real paths:
sed -e "s|@REPO@|$(pwd)|g" -e "s|@HOME@|$HOME|g" \
    deploy/macos/com.mnemo-cortex.server.plist \
    > ~/Library/LaunchAgents/com.mnemo-cortex.server.plist

mkdir -p ~/Library/Logs/mnemo-cortex
launchctl bootstrap gui/$UID ~/Library/LaunchAgents/com.mnemo-cortex.server.plist

Verify:

launchctl print gui/$UID/com.mnemo-cortex.server | head -20
curl -s http://127.0.0.1:50001/health
tail -f ~/Library/Logs/mnemo-cortex/server.log

Stop / uninstall:

launchctl bootout gui/$UID/com.mnemo-cortex.server
rm ~/Library/LaunchAgents/com.mnemo-cortex.server.plist

If you use the LaunchAgent, don't also use mnemo-cortex start/stop — launchd restarts the server when it dies, so stop alone won't keep it down. Use launchctl bootout to stop it for real.

Step 5 — Connect your agent (MCP bridge)

The bridge is pure JavaScript (no native modules) and works unchanged on macOS:

cd integrations/mcp-bridge && npm install

Then point your MCP host at integrations/mcp-bridge/server.js — per-host walkthroughs (Claude Desktop .mcpb bundle, Claude Code, LM Studio, AnythingLLM, OpenClaw, llama.cpp) are in the main README. Claude Desktop config on macOS lives at ~/Library/Application Support/Claude/claude_desktop_config.json.

Client support: any MCP client speaking stdio transport works (Claude Code, Claude Desktop, OpenAI Codex CLI, LM Studio, and the rest of the README list). ChatGPT Desktop is not supported — it has no native MCP; its connectors would require exposing your memory server on a public HTTPS endpoint, and that security cost exceeds the value. We'll revisit if OpenAI ships native MCP support.

Known limitations on macOS (beta)

mnemo-cortex doctor / health service checks are systemd-aware only — the "Services" section reports SKIP: systemctl not available on macOS. That's expected, not a failure. The launchd equivalent is launchctl print gui/$UID/com.mnemo-cortex.server.
The OpenClaw session watcher service unit is Linux-only (systemd). The watcher itself runs fine in the foreground; a launchd template for it will follow after the server template survives beta.
No Homebrew formula yet. sqlite-vec publishes wheels but no sdist, and mnemo-cortex isn't on PyPI yet, so a classic formula can't build from source. git clone + venv (above) is the supported path for now.

Troubleshooting

AttributeError: ... no attribute 'enable_load_extension' — you're on Apple's or python.org's Python. See the warning at the top; rebuild the venv with Homebrew Python.

sqlite3.OperationalError mentioning vec0 or vec_version — the sqlite-vec extension didn't load. Rerun the Step 1 preflight inside the venv.

Embeddings fail / recall returns nothing — check Ollama: curl -s http://localhost:11434/api/tags should list nomic-embed-text. If you installed Ollama.app instead of Homebrew, it must be running (menu-bar icon).

Port 50001 already in use — lsof -nP -iTCP:50001 | grep LISTEN to find the owner. Change the port in agentb.yaml (server.port) or set MNEMO_PORT.

LaunchAgent won't start — check ~/Library/Logs/mnemo-cortex/server.log first, then launchctl print gui/$UID/com.mnemo-cortex.server (look at last exit code). The most common cause is a stale @REPO@/@HOME@ placeholder from skipping the sed step.

"cannot be opened because the developer cannot be verified" — shouldn't happen for pip-installed wheels (Gatekeeper quarantine applies to browser downloads, not pip), but if you copied the repo from a downloaded zip: xattr -dr com.apple.quarantine .venv.