diff --git a/src/resources/extensions/sf/templates/PREFERENCES.md b/src/resources/extensions/sf/templates/PREFERENCES.md
index 10925ff4e..60b0827bf 100644
--- a/src/resources/extensions/sf/templates/PREFERENCES.md
+++ b/src/resources/extensions/sf/templates/PREFERENCES.md
@@ -117,3 +117,16 @@ pre_dispatch_hooks: []
 # SF Skill Preferences
 
 See `~/.sf/agent/extensions/sf/docs/preferences-reference.md` for full field documentation and examples.
+
+## Environment variables (not in YAML — set via shell)
+
+These are runtime-only; SF reads them at startup, never persists them, never logs the key.
+
+| Variable | Purpose | Default |
+|---|---|---|
+| `SF_LLM_GATEWAY_KEY` | Bearer token for the inference-fabric llm-gateway. **When unset, embeddings are disabled** and `getRelevantMemoriesRanked` falls back to static (confidence × hit-count) ranking. | (unset) |
+| `SF_LLM_GATEWAY_URL` | OpenAI-compatible endpoint base, including `/v1`. | `https://llm-gateway.centralcloud.com/v1` |
+| `SF_LLM_GATEWAY_EMBED_MODEL` | Embedding model id served by the gateway. | `qwen/qwen3-embedding-4b` |
+| `SF_LLM_GATEWAY_RERANK_MODEL` | Rerank model id. When unset OR no rerank worker is online, rerank silently degrades and the cosine pass alone ranks results. | (unset) |
+
+Once `SF_LLM_GATEWAY_KEY` is set, the agent_end hook opportunistically backfills embeddings for any memories without vectors (50 per turn, 16 per batch). `/sf memory search "<query>"` lights up the embedding-ranked path; without the key it shows static rank.