diff --git a/skills/llmdoc/SKILL.md b/skills/llmdoc/SKILL.md
index 2c65ae1..d7e3f21 100644
--- a/skills/llmdoc/SKILL.md
+++ b/skills/llmdoc/SKILL.md
@@ -24,6 +24,7 @@ Read these references in order:
 2. `references/operating-protocol.md`
 3. `references/doc-structure.md`
 4. `references/update-and-memory.md`
+5. `references/lessons-learned.md`
 
 Then load only the specific extras you need:
 
diff --git a/skills/llmdoc/references/lessons-learned.md b/skills/llmdoc/references/lessons-learned.md
new file mode 100644
index 0000000..36a3e5f
--- /dev/null
+++ b/skills/llmdoc/references/lessons-learned.md
@@ -0,0 +1,70 @@
+# Lessons Learned And Memory Archive
+
+This reference is the single source of truth for the optional curated memory file:
+
+- `llmdoc/memory/lessons-learned.md`
+
+The file is a compact index of reusable behavioral rules distilled from accumulated memory files. It is useful when a project has accumulated enough memory that future sessions should read a summary instead of every raw note.
+
+## Trigger
+
+During `/llmdoc:update`, automatically run a memory summary/archive pass when active memory files exceed 5.
+
+Active memory files are files under `llmdoc/memory/` excluding:
+
+- `llmdoc/memory/lessons-learned.md`
+- `llmdoc/memory/doc-gaps.md`
+- anything under `llmdoc/memory/archive/`
+
+The threshold check should happen after the new task reflection is written, so the newest lesson is included.
+
+## What Belongs Here
+
+Use `memory/lessons-learned.md` for:
+
+- cross-task rules that should change future agent behavior
+- recurring engineering discipline
+- repeated workflow mistakes and their corrected rule
+- source-linked lessons extracted from one or more reflections
+
+Do not use it for:
+
+- raw task narratives, which belong in `memory/reflections/`
+- durable decisions, which belong in `memory/decisions/`
+- missing-doc inventory, which belongs in `memory/doc-gaps.md`
+- full workflow instructions, which belong in `guides/`
+
+## Archive Rule
+
+When the threshold is crossed:
+
+1. Read the active memory files.
+2. Extract compact, actionable, recurring rules.
+3. Add or revise matching entries in `memory/lessons-learned.md`.
+4. Link each entry back to the source memory file before moving it.
+5. Move summarized raw memory files into `llmdoc/memory/archive/YYYY-MM-DD/`.
+6. Leave unsummarized active memory files in place if they still need direct follow-up.
+7. Do not duplicate the same instruction across prompts, agents, README files, and stable docs.
+
+This is a memory compaction step, not a stable-doc rewrite. Promote a lesson into `must/`, `guides/`, `architecture/`, or `reference/` only when it is mature enough to become stable project guidance.
+
+If the count is 5 or lower, do not update `lessons-learned.md` just because a single task produced a useful rule. Let raw reflections accumulate until the archive threshold is crossed.
+
+## Template
+
+```md
+# Lessons Learned
+
+Curated cross-task rules distilled from archived memory.
+
+## <Theme>
+
+### <Rule Name>
+**Rule**: One actionable sentence.
+**Why**: One sentence naming the failure, correction, or discovery that produced the rule.
+**Source**: `llmdoc/memory/archive/YYYY-MM-DD/<source-file>.md`
+```
+
+## Hook Reinforcement
+
+Hooks may remind the assistant that the active-memory count crossed the archive threshold, but the detailed behavior should stay here.
diff --git a/skills/llmdoc/templates/session-start.sh b/skills/llmdoc/templates/session-start.sh
index 134d442..6bbe850 100755
--- a/skills/llmdoc/templates/session-start.sh
+++ b/skills/llmdoc/templates/session-start.sh
@@ -8,7 +8,7 @@ if [ -d "$project_dir/llmdoc" ]; then
 {
   "hookSpecificOutput": {
     "hookEventName": "SessionStart",
-    "additionalContext": "This project has llmdoc enabled. Load the llmdoc skill. Read llmdoc/index.md, then llmdoc/startup.md, then the MUST docs listed there. Before non-trivial edits, proactively read relevant guides and reflections and align with the user on approach."
+    "additionalContext": "This project has llmdoc enabled. Load the llmdoc skill. Read llmdoc/index.md, then llmdoc/startup.md, then the MUST docs listed there. If llmdoc/memory/lessons-learned.md exists, read it as the archived memory summary. Before non-trivial edits, proactively read relevant guides and reflections and align with the user on approach."
   }
 }
 EOF
diff --git a/skills/llmdoc/templates/stop.sh b/skills/llmdoc/templates/stop.sh
index 1ea67d4..425ab55 100755
--- a/skills/llmdoc/templates/stop.sh
+++ b/skills/llmdoc/templates/stop.sh
@@ -8,7 +8,29 @@ timestamp="$(date -u +%Y%m%dT%H%M%SZ)"
 mkdir -p "$tmp_dir"
 cat > "$tmp_dir/stop-$timestamp.json"
 
-cat <<'EOF'
+active_memory_count=0
+if [ -d "$project_dir/llmdoc/memory" ]; then
+  active_memory_count="$(
+    find "$project_dir/llmdoc/memory" -type f \
+      ! -path "$project_dir/llmdoc/memory/archive/*" \
+      ! -name "lessons-learned.md" \
+      ! -name "doc-gaps.md" \
+      | wc -l \
+      | tr -d '[:space:]'
+  )"
+fi
+
+if [ "$active_memory_count" -gt 5 ]; then
+  cat <<EOF
+{
+  "hookSpecificOutput": {
+    "hookEventName": "Stop",
+    "additionalContext": "llmdoc active memory has ${active_memory_count} files, which exceeds the archive threshold of 5. During /llmdoc:update, summarize recurring lessons into llmdoc/memory/lessons-learned.md and archive summarized raw memory per skills/llmdoc/references/lessons-learned.md."
+  }
+}
+EOF
+else
+  cat <<'EOF'
 {
   "hookSpecificOutput": {
     "hookEventName": "Stop",
@@ -16,3 +38,4 @@ cat <<'EOF'
   }
 }
 EOF
+fi