diff --git a/notes/bf-2q9r.md b/notes/bf-2q9r.md
new file mode 100644
index 0000000..3c750cc
--- /dev/null
+++ b/notes/bf-2q9r.md
@@ -0,0 +1,41 @@
+# Bead bf-2q9r: Per-needle-worker MemoryMax ceiling (4 GB)
+
+## Summary
+
+Added per-needle-worker MemoryMax ceiling (4 GB) via systemd-run scope wrapper so no single worker can exhaust the cgroup's available memory.
+
+## Implementation
+
+**Approach:** systemd transient scope with MemoryMax=4G
+
+Modified needle adapter configs to wrap worker invocation with:
+```bash
+systemd-run --user --scope -p MemoryMax=4G bash -c '...'
+```
+
+**Files modified:**
+- `/home/coding/.config/needle/adapters/claude-code-glm-4.7.yaml` (already had MemoryMax)
+- `/home/coding/.config/needle/adapters/claude-code-glm-5.yaml` (added MemoryMax)
+- `/home/coding/.config/needle/adapters/claude-code-glm-5-1.yaml` (added MemoryMax)
+
+## Verification
+
+Checked active systemd scopes:
+```bash
+$ systemctl --user show run-p857080-i9244556.scope -p MemoryMax
+MemoryMax=4294967296  # 4 GB
+
+$ systemctl --user show run-p857080-i9244556.scope -p MemoryCurrent
+MemoryCurrent=647168  # ~632 KB
+```
+
+## Rationale
+
+With only a cgroup-level soft limit (user-1001.slice MemoryHigh), one runaway worker could still consume all available memory before pressure kills it. The per-process MemoryMax bounds each Claude Code session at 4 GB RSS, ensuring 6 workers + fabric-web + VSCode stay well under 32 GB total.
+
+## Notes
+
+- MemoryMax is a hard limit — the process is OOM-killed if it exceeds 4 GB
+- This complements (not replaces) the user-1001.slice MemoryHigh soft limit
+- systemd-run scope automatically cleanup when process exits
+- Backups created: `.pre-memorycap.bak` files