Ollama setup (laptop-local embedding provider)¶

Topology B runs Ollama + bge-m3 natively on the laptop only (never on the server) — see DEPLOY.md and spec AC-44. This is the one piece of claude-memory's stack that is not this repo's own code; it is a Homebrew service you install and keep warm.

Install¶

brew install ollama
brew services start ollama

brew services start installs a launchd LaunchAgent at ~/Library/LaunchAgents/homebrew.mxcl.ollama.plist that keeps ollama serve running and restarts it on login/crash.

Keep the model loaded¶

Nothing to configure. claude-memory sends "keep_alive": -1 with every /api/embed request (internal/ollama/embedder.go), so after the first call the model stays loaded until Ollama restarts — the hook never pays the cold-start cost. (Editing OLLAMA_KEEP_ALIVE into ~/Library/LaunchAgents/homebrew.mxcl.ollama.plist does not persist: brew services restart and brew upgrade regenerate that plist.)

To warm it right after a reboot instead of on the first prompt:

curl -s http://127.0.0.1:11434/api/embed \
  -d '{"model":"bge-m3","input":"warmup","keep_alive":-1}' > /dev/null
ollama ps   # UNTIL: Forever

Pull the model¶

ollama pull bge-m3

Health check¶

curl -sS http://127.0.0.1:11434/api/embed \
  -d '{"model":"bge-m3","input":"health check","truncate":true,"options":{"num_ctx":2048,"num_batch":2048}}' \
  | python3 -m json.tool | head -5
# Expect a JSON body with an "embeddings" array of one 1024-length vector.

If this fails: - brew services list — confirm ollama shows started. - ollama list — confirm bge-m3 is pulled. - tail -f ~/Library/Logs/Homebrew/ollama/ollama.log (or brew services info ollama for the actual log path on your machine).