diff --git a/src/context_engine/integration/mcp_server.py b/src/context_engine/integration/mcp_server.py
index d359223..befb549 100644
--- a/src/context_engine/integration/mcp_server.py
+++ b/src/context_engine/integration/mcp_server.py
@@ -469,6 +469,12 @@ def __init__(self, retriever, backend, compressor, embedder, config) -> None:
         # Lazy indexing flag — triggers on first context_search if index is empty.
         self._lazy_indexed = False
 
+        # Memory nudge state — tracks tool activity so context_search can
+        # append reminders when the agent hasn't been recording (#nudges).
+        # Reset on every record_decision / record_code_area call.
+        self._searches_since_last_decision = 0
+        self._has_recorded_decision = False
+
         self._register_tools()
         self._register_prompts()
 
@@ -709,6 +715,76 @@ def _apply_output_compression(self, body: str) -> str:
             )
         return out
 
+    # ── memory nudges ───────────────────────────────────────────────────
+
+    # Thresholds for triggering nudges. After the first record_decision,
+    # the search threshold doubles — the agent has shown it knows how, so
+    # we nudge less aggressively.
+    _DECISION_NUDGE_FIRST = 4
+    _DECISION_NUDGE_REPEAT = 8
+    _CODE_AREA_NUDGE_THRESHOLD = 3
+    _CODE_AREA_NUDGE_MAX_FILES = 5
+
+    def _build_nudge(self) -> str:
+        """Return a nudge string to append to context_search results, or "".
+
+        Two independent triggers:
+        1. Decision nudge: N searches with 0 decisions recorded.
+        2. Code area nudge: M files touched (via search/expand) but not
+           explicitly record_code_area'd this session.
+
+        Both can fire in the same response. The "skip if" clause gives the
+        agent permission to ignore the nudge when it's just reading.
+        """
+        parts: list[str] = []
+
+        # ── decision nudge ────────────────────────────────────────────
+        threshold = (
+            self._DECISION_NUDGE_REPEAT
+            if self._has_recorded_decision
+            else self._DECISION_NUDGE_FIRST
+        )
+        if self._searches_since_last_decision >= threshold:
+            n = self._searches_since_last_decision
+            snap = self._session_capture.get_session_snapshot(self._session_id)
+            n_decisions = len(snap["decisions"]) if snap else 0
+            parts.append(
+                f"[CCE] {n} context searches this session, "
+                f"{n_decisions} decision(s) recorded.\n"
+                f"If you chose a library, resolved an ambiguity, or "
+                f"established a convention, call "
+                f'record_decision(decision="...", reason="...").\n'
+                f"Skip if this session is just exploration/reading."
+            )
+
+        # ── code area nudge ───────────────────────────────────────────
+        snap = self._session_capture.get_session_snapshot(self._session_id)
+        if snap:
+            touched = set(snap.get("touched_files", {}).keys())
+            recorded = {
+                ca["file_path"] for ca in snap.get("code_areas", [])
+            }
+            unrecorded = sorted(touched - recorded)
+            if len(unrecorded) >= self._CODE_AREA_NUDGE_THRESHOLD:
+                shown = unrecorded[:self._CODE_AREA_NUDGE_MAX_FILES]
+                suffix = (
+                    f" (+{len(unrecorded) - len(shown)} more)"
+                    if len(unrecorded) > len(shown) else ""
+                )
+                file_list = ", ".join(shown) + suffix
+                parts.append(
+                    f"[CCE] {len(unrecorded)} files explored but not "
+                    f"recorded: {file_list}\n"
+                    f"If you modified or traced a non-obvious flow, call "
+                    f'record_code_area(file_path="...", '
+                    f'description="...").\n'
+                    f"Skip for files you only glanced at."
+                )
+
+        if not parts:
+            return ""
+        return "\n\n---\n" + "\n\n".join(parts)
+
     def get_tool_names(self) -> list[str]:
         return list(self.TOOL_NAMES)
 
@@ -1048,6 +1124,14 @@ async def _handle_context_search(self, args):
         self._persist_current_session()
 
         body = _format_results_with_overflow(inline_chunks, overflow_chunks)
+        # Memory nudge — appended before output compression so the agent
+        # sees it as part of the tool result. Fires only when thresholds
+        # are met (see _build_nudge). Increment BEFORE building the nudge
+        # so the current search counts toward the threshold.
+        self._searches_since_last_decision += 1
+        nudge = self._build_nudge()
+        if nudge:
+            body = body + nudge
         body = self._apply_output_compression(body)
         # Only note omissions the retriever actually made (threshold or
         # marginal stop) — chunk-count heuristics false-positive on every
@@ -1246,6 +1330,10 @@ def _handle_record_decision(self, args):
         reason = memory_db.scrub_pii(reason)
         self._session_capture.record_decision(self._session_id, decision, reason)
         self._persist_current_session()
+        # Reset nudge state — the agent recorded something, give it a
+        # longer leash before the next nudge.
+        self._searches_since_last_decision = 0
+        self._has_recorded_decision = True
         # Dual-write into memory.db. `decision` and `reason` are compressed
         # via the grammar module before INSERT — structured tokens (paths,
         # versions, identifiers) are preserved byte-for-byte; only prose
diff --git a/src/context_engine/memory/hook_installer.py b/src/context_engine/memory/hook_installer.py
index 7e657be..eb20ed6 100644
--- a/src/context_engine/memory/hook_installer.py
+++ b/src/context_engine/memory/hook_installer.py
@@ -97,7 +97,7 @@ def _is_windows() -> bool:
     bash -c "exec 3<>/dev/tcp/127.0.0.1/${PORT}" 2>/dev/null || exit 0
 fi
 
-if [ "${HOOK_NAME}" = "SessionStart" ]; then
+if [ "${HOOK_NAME}" = "SessionStart" ] || [ "${HOOK_NAME}" = "Stop" ]; then
     RESPONSE="$(curl -sf -m 2 -X POST -H "Content-Type: application/json" \\
         --data-binary @- "http://127.0.0.1:${PORT}/hooks/${HOOK_NAME}" \\
         2>/dev/null || true)"
@@ -154,17 +154,19 @@ def _is_windows() -> bool:
 powershell -NoProfile -Command "$ErrorActionPreference='Stop'; try { (New-Object Net.Sockets.TcpClient).Connect('127.0.0.1',%PORT%); exit 0 } catch { exit 1 }" >nul 2>&1
 if errorlevel 1 exit /b 0
 
-if /i "%HOOK_NAME%"=="SessionStart" (
-    set "TMP_RESP=%TEMP%\\cce_hook_resp_%RANDOM%.txt"
-    curl -sf -m 2 -X POST -H "Content-Type: application/json" ^
-        --data-binary @- "http://127.0.0.1:%PORT%/hooks/%HOOK_NAME%" ^
-        > "%TMP_RESP%" 2>nul
-    if exist "%TMP_RESP%" type "%TMP_RESP%"
-    if exist "%TMP_RESP%" del "%TMP_RESP%" >nul 2>&1
-) else (
-    curl -sf -m 1 -X POST -H "Content-Type: application/json" ^
-        --data-binary @- "http://127.0.0.1:%PORT%/hooks/%HOOK_NAME%" >nul 2>&1
-)
+if /i "%HOOK_NAME%"=="SessionStart" goto :capture_response
+if /i "%HOOK_NAME%"=="Stop" goto :capture_response
+curl -sf -m 1 -X POST -H "Content-Type: application/json" ^
+    --data-binary @- "http://127.0.0.1:%PORT%/hooks/%HOOK_NAME%" >nul 2>&1
+goto :eof
+
+:capture_response
+set "TMP_RESP=%TEMP%\\cce_hook_resp_%RANDOM%.txt"
+curl -sf -m 2 -X POST -H "Content-Type: application/json" ^
+    --data-binary @- "http://127.0.0.1:%PORT%/hooks/%HOOK_NAME%" ^
+    > "%TMP_RESP%" 2>nul
+if exist "%TMP_RESP%" type "%TMP_RESP%"
+if exist "%TMP_RESP%" del "%TMP_RESP%" >nul 2>&1
 exit /b 0
 """
 
diff --git a/src/context_engine/memory/hooks.py b/src/context_engine/memory/hooks.py
index a76bf19..9ac659b 100644
--- a/src/context_engine/memory/hooks.py
+++ b/src/context_engine/memory/hooks.py
@@ -383,9 +383,75 @@ async def handle_stop(request: web.Request) -> web.Response:
     except Exception:
         log.exception("Stop enqueue failed")
         return web.json_response({"ok": False}, status=202)
+
+    # Memory nudge — summarise unrecorded activity so the agent can fire
+    # off quick record_decision / record_code_area calls before the session
+    # ends. The hook script captures Stop stdout (like SessionStart) and
+    # injects it into the model context.
+    nudge = _build_stop_nudge(conn, session_id)
+    if nudge:
+        return web.Response(text=nudge, content_type="text/plain")
     return web.json_response({"ok": True})
 
 
+def _build_stop_nudge(conn: sqlite3.Connection, session_id: str) -> str:
+    """Build a session-end nudge summarising unrecorded activity.
+
+    Queries memory.db for tool_events (context_search calls) and decisions
+    recorded during this session. Returns a short directive if activity
+    was significant but recording was low. Returns "" if nothing to nudge.
+    """
+    try:
+        # Count context_search tool calls this session
+        row = conn.execute(
+            "SELECT COUNT(*) AS n FROM tool_events "
+            "WHERE session_id = ? AND tool_name = 'context_search'",
+            (session_id,),
+        ).fetchone()
+        n_searches = int(row["n"]) if row else 0
+
+        # Count decisions recorded this session
+        row = conn.execute(
+            "SELECT COUNT(*) AS n FROM decisions WHERE session_id = ?",
+            (session_id,),
+        ).fetchone()
+        n_decisions = int(row["n"]) if row else 0
+
+        # Count code_areas recorded this session
+        row = conn.execute(
+            "SELECT COUNT(*) AS n FROM code_areas WHERE session_id = ?",
+            (session_id,),
+        ).fetchone()
+        n_code_areas = int(row["n"]) if row else 0
+
+        # Only nudge if there was real activity (4+ searches) and
+        # recording was sparse.
+        if n_searches < 4:
+            return ""
+        parts: list[str] = []
+        if n_decisions == 0:
+            parts.append(
+                "If you chose a library, resolved an ambiguity, or "
+                "established a convention, call "
+                'record_decision(decision="...", reason="...") now.'
+            )
+        if n_code_areas == 0:
+            parts.append(
+                "If you modified or traced non-obvious code, call "
+                'record_code_area(file_path="...", description="...") now.'
+            )
+        if not parts:
+            return ""
+        header = (
+            f"[CCE session end] {n_searches} searches, "
+            f"{n_decisions} decision(s), {n_code_areas} code area(s) recorded."
+        )
+        return header + "\n" + "\n".join(parts)
+    except Exception:
+        log.debug("Stop nudge query failed", exc_info=True)
+        return ""
+
+
 async def handle_session_end(request: web.Request) -> web.Response:
     data = await _read_json(request)
     session_id = data.get("session_id")
diff --git a/tests/integration/test_mcp_server.py b/tests/integration/test_mcp_server.py
index d95c800..93234d0 100644
--- a/tests/integration/test_mcp_server.py
+++ b/tests/integration/test_mcp_server.py
@@ -30,6 +30,9 @@ def _make_server(tmp_path):
     # _record_bucket already guards on this; tests don't need a real db.
     server._memory_conn = None
     server._storage_base = tmp_path
+    # Memory nudge state (see ContextEngineMCP.__init__)
+    server._searches_since_last_decision = 0
+    server._has_recorded_decision = False
     return server
 
 
@@ -176,6 +179,9 @@ async def fake_retrieve(query, top_k=10, confidence_threshold=0.0,
     server._compressor.compress = AsyncMock(return_value=[stub_chunk])
     server._session_capture = MagicMock()
     server._session_capture.touch_files = MagicMock()
+    server._session_capture.get_session_snapshot = MagicMock(return_value={
+        "decisions": [], "code_areas": [], "touched_files": {},
+    })
     server._persist_current_session = MagicMock()
     server._record = MagicMock()
     server._append_audit_log = MagicMock()
@@ -208,3 +214,62 @@ async def test_context_search_no_note_when_nothing_dropped(tmp_path):
     result = await server._handle_context_search({"query": "find something", "top_k": 5})
     text = result[0].text
     assert "lower-confidence results omitted" not in text
+
+
+@pytest.mark.asyncio
+async def test_decision_nudge_fires_after_threshold(tmp_path):
+    """After 4 context_search calls with no record_decision, the nudge appears."""
+    server = _make_search_server(tmp_path, dropped_low_value=0)
+    # First 3 searches: no nudge
+    for _ in range(3):
+        result = await server._handle_context_search({"query": "q", "top_k": 5})
+        assert "[CCE]" not in result[0].text
+    # 4th search: nudge fires
+    result = await server._handle_context_search({"query": "q", "top_k": 5})
+    assert "[CCE] 4 context searches" in result[0].text
+    assert "record_decision" in result[0].text
+
+
+@pytest.mark.asyncio
+async def test_decision_nudge_resets_after_record(tmp_path):
+    """Recording a decision resets the search counter."""
+    server = _make_search_server(tmp_path, dropped_low_value=0)
+    # Trigger nudge
+    for _ in range(4):
+        await server._handle_context_search({"query": "q", "top_k": 5})
+    # Simulate record_decision resetting state
+    server._searches_since_last_decision = 0
+    server._has_recorded_decision = True
+    # Next search: no nudge (counter reset, higher threshold now)
+    result = await server._handle_context_search({"query": "q", "top_k": 5})
+    assert "[CCE]" not in result[0].text
+
+
+@pytest.mark.asyncio
+async def test_code_area_nudge_fires_with_unrecorded_files(tmp_path):
+    """Nudge appears when 3+ files are touched but none recorded as code areas."""
+    server = _make_search_server(tmp_path, dropped_low_value=0)
+    # Simulate 3 touched files with no code_areas recorded
+    server._session_capture.get_session_snapshot = MagicMock(return_value={
+        "decisions": [], "code_areas": [],
+        "touched_files": {"a.py": 1, "b.py": 2, "c.py": 1},
+    })
+    result = await server._handle_context_search({"query": "q", "top_k": 5})
+    assert "files explored but not recorded" in result[0].text
+    assert "record_code_area" in result[0].text
+
+
+@pytest.mark.asyncio
+async def test_code_area_nudge_excludes_recorded_files(tmp_path):
+    """Files that were record_code_area'd don't count toward the threshold."""
+    server = _make_search_server(tmp_path, dropped_low_value=0)
+    server._session_capture.get_session_snapshot = MagicMock(return_value={
+        "decisions": [], "touched_files": {"a.py": 1, "b.py": 2, "c.py": 1},
+        "code_areas": [
+            {"file_path": "a.py", "description": "x", "timestamp": 0},
+            {"file_path": "b.py", "description": "y", "timestamp": 0},
+        ],
+    })
+    result = await server._handle_context_search({"query": "q", "top_k": 5})
+    # Only 1 unrecorded file (c.py) — below threshold
+    assert "[CCE]" not in result[0].text or "files explored" not in result[0].text