diff --git a/Deployment.md b/Deployment.md
new file mode 100644
index 0000000..5e4c417
--- /dev/null
+++ b/Deployment.md
@@ -0,0 +1,56 @@
+# Deployment
+
+SecondBrain is one file + one SQLite db. It runs anywhere Python and Ollama reach.
+
+## 1. Local (any OS)
+```bash
+ollama pull nomic-embed-text          # embedding backend
+bash deploy/install.sh                # venv + deps + schema
+. .venv/bin/activate
+python3 brain.py ingest-files <dir>
+python3 brain.py embed
+python3 brain.py recall "q" --hybrid
+```
+
+## 2. Continuous refresh (systemd timer, Linux)
+Set your sources in `.env`:
+```
+SB_FILE_DIRS=/home/me/notes:/home/me/code
+SB_TRANSCRIPTS=/home/me/.claude/projects
+SB_EVENTS_DB=/home/me/events.db
+SB_EMBED_LIMIT=6000
+```
+Then:
+```bash
+sudo cp deploy/secondbrain-refresh.* /etc/systemd/system/
+# edit WorkingDirectory/EnvironmentFile in the .service to match your install
+sudo systemctl enable --now secondbrain-refresh.timer
+```
+`refresh.sh` is locked (flock), niced, and idempotent — safe to run every 15 min.
+
+## 3. Docker
+```bash
+docker build -t secondbrain -f deploy/Dockerfile .
+docker run --rm -v sbdata:/data \
+  -e OLLAMA_URL=http://host.docker.internal:11434 \
+  secondbrain recall "q" --hybrid
+```
+The db lives on the `/data` volume and survives rebuilds.
+
+## 4. GPU offload (keep a laptop's brain current from a desktop GPU)
+Embeddings are the only heavy step. Point `OLLAMA_URL` at any reachable Ollama —
+a desktop/GPU box on your LAN or tailnet:
+```
+OLLAMA_URL=http://100.x.y.z:11434
+```
+SecondBrain embeds *there* and stores the vectors *locally*. No extra infrastructure,
+no tunnels — if you can curl the Ollama endpoint, it works. When the GPU box is
+asleep, embedding simply pauses; lexical (FTS5) recall keeps working meanwhile.
+
+## Moving / backing up a brain
+It's one file. `cp brain.db elsewhere`. That's the whole backup and migration story.
+
+## Sizing
+- ~768 floats × 4 bytes ≈ 3 KB of vector per chunk, plus the text.
+- Embedding rate depends on your Ollama backend (CPU ~5–20/s, small GPU ~20–60/s).
+- `embed --limit N` bounds a run; the rest continues next run.