diff --git a/install.sh b/install.sh
index 1db7f3b..5549f35 100755
--- a/install.sh
+++ b/install.sh
@@ -190,6 +190,40 @@ fi
 echo "-> Step 5: Setting up dashboard service..."
 "$DASHBOARD_SERVICE" install "$PY_PATH" "$INSTALL_DIR" "$MEMORY_DIR"
 
+# -- 5b. Linux: systemd auto-drain (equivalent of macOS LaunchAgent WatchPaths) --
+# On macOS, the reflection LaunchAgent in launchagents/ picks up
+# `touch ~/.claude-memory/.reflect-pending` and runs run_reflection.py.
+# On Linux there's no LaunchAgent — this block installs a systemd.path +
+# oneshot service pair that gives the same behavior via inotify.
+if [ "$(uname)" = "Linux" ] && [ -d "$INSTALL_DIR/systemd" ]; then
+    echo "-> Step 5b: Installing Linux systemd auto-drain unit..."
+    SYSTEMD_USER_DIR="$HOME/.config/systemd/user"
+    mkdir -p "$SYSTEMD_USER_DIR"
+
+    for unit in claude-memory-reflection.service claude-memory-reflection.path; do
+        sed \
+            -e "s|@INSTALL_DIR@|$INSTALL_DIR|g" \
+            -e "s|@MEMORY_DIR@|$MEMORY_DIR|g" \
+            "$INSTALL_DIR/systemd/$unit" > "$SYSTEMD_USER_DIR/$unit"
+    done
+
+    # Ensure the trigger file exists so systemd.path can watch it from the start
+    touch "$MEMORY_DIR/.reflect-pending"
+
+    if command -v systemctl &>/dev/null; then
+        systemctl --user daemon-reload
+        systemctl --user enable --now claude-memory-reflection.path >/dev/null 2>&1 || {
+            echo "  WARN: could not enable the .path unit via systemctl --user."
+            echo "        Run manually: systemctl --user enable --now claude-memory-reflection.path"
+        }
+        if systemctl --user is-active claude-memory-reflection.path >/dev/null 2>&1; then
+            echo "  OK: Reflection auto-drain active (watch: $MEMORY_DIR/.reflect-pending)"
+        fi
+    else
+        echo "  WARN: systemctl not found — units copied to $SYSTEMD_USER_DIR but not activated"
+    fi
+fi
+
 # -- 6. Verify --
 echo ""
 echo "-> Step 6: Verifying installation..."
diff --git a/systemd/README.md b/systemd/README.md
new file mode 100644
index 0000000..8bf4a6a
--- /dev/null
+++ b/systemd/README.md
@@ -0,0 +1,29 @@
+# systemd units (Linux)
+
+Linux equivalent of the macOS LaunchAgents in `../launchagents/`.
+
+`install.sh` auto-installs these when running on Linux — it substitutes
+`@INSTALL_DIR@` and `@MEMORY_DIR@`, copies the files to
+`~/.config/systemd/user/`, and enables them via `systemctl --user`.
+
+## Units
+
+- **`claude-memory-reflection.path`** — watches `~/.claude-memory/.reflect-pending`
+  and fires the service whenever `memory_save` touches the trigger file.
+- **`claude-memory-reflection.service`** — `Type=oneshot`, runs
+  `src/tools/run_reflection.py`, drains `triple_extraction_queue`,
+  `deep_enrichment_queue`, and `representations_queue` through Ollama.
+
+## Manual management
+
+```bash
+systemctl --user status claude-memory-reflection.path
+systemctl --user status claude-memory-reflection.service
+journalctl --user -u claude-memory-reflection.service -f
+
+# disable auto-drain entirely
+systemctl --user disable --now claude-memory-reflection.path
+
+# run a one-off drain
+systemctl --user start claude-memory-reflection.service
+```
diff --git a/systemd/claude-memory-reflection.path b/systemd/claude-memory-reflection.path
new file mode 100644
index 0000000..63cee9d
--- /dev/null
+++ b/systemd/claude-memory-reflection.path
@@ -0,0 +1,11 @@
+[Unit]
+Description=Watch .reflect-pending → trigger reflection drain
+Documentation=https://github.com/vbcherepanov/total-agent-memory
+
+[Path]
+# Fires on IN_CLOSE_WRITE — `touch` from memory_save triggers it
+PathChanged=@MEMORY_DIR@/.reflect-pending
+Unit=claude-memory-reflection.service
+
+[Install]
+WantedBy=default.target
diff --git a/systemd/claude-memory-reflection.service b/systemd/claude-memory-reflection.service
new file mode 100644
index 0000000..0ab6781
--- /dev/null
+++ b/systemd/claude-memory-reflection.service
@@ -0,0 +1,15 @@
+[Unit]
+Description=Claude Memory — drain reflection queues via Ollama (one-shot)
+Documentation=https://github.com/vbcherepanov/total-agent-memory
+After=network-online.target
+
+[Service]
+Type=oneshot
+WorkingDirectory=@INSTALL_DIR@
+ExecStart=@INSTALL_DIR@/.venv/bin/python @INSTALL_DIR@/src/tools/run_reflection.py
+Environment=CLAUDE_MEMORY_DIR=@MEMORY_DIR@
+# Cold Ollama load ~30s first time, then ~10s per queue item
+TimeoutStartSec=1200
+Nice=10
+StandardOutput=journal
+StandardError=journal