diff --git a/anton/core/session.py b/anton/core/session.py
index 77eece9a..11bbf44d 100644
--- a/anton/core/session.py
+++ b/anton/core/session.py
@@ -353,6 +353,11 @@ def __init__(self, config: ChatSessionConfig) -> None:
         self._history: list[dict] = (
             list(config.initial_history) if config.initial_history else []
         )
+        # Seed length + how many leading history messages the last compaction
+        # folded into its summary — lets a host map the compaction back onto
+        # its own message list (see `last_compaction`).
+        self._seed_len = len(self._history)
+        self._last_compacted_count: int | None = None
         self._pending_memory_confirmations: list = []
         self._turn_count = (
             sum(1 for m in self._history if is_user_turn(m))
@@ -484,6 +489,23 @@ def _acc_has_similar(rule: str) -> bool:
     def history(self) -> list[dict]:
         return self._history
 
+    @property
+    def last_compaction(self) -> dict | None:
+        """Result of the last history compaction this session ran, or None.
+
+        `summary`: the compacted message content, verbatim (re-seed it as
+        history[0] next turn so it gets recognized and extended, not
+        resummarized). `covered_through`: how many of `initial_history`'s
+        messages are now folded into it — map this count onto your own
+        message list to find the cutoff.
+        """
+        if self._last_compacted_count is None:
+            return None
+        return {
+            "summary": self._history[0]["content"],
+            "covered_through": min(self._last_compacted_count, self._seed_len),
+        }
+
     def _apply_error_tracking(
         self,
         result_text: str,
@@ -983,17 +1005,19 @@ async def _summarize_history(self) -> None:
             return  # Too short to summarize
 
         min_recent = 4
-        split = max(int(len(self._history) * 0.6), 1)
+        # Number of leading messages to fold into the summary; doubles as the
+        # cut index (old = history[:compacted_count], recent = the rest).
+        compacted_count = max(int(len(self._history) * 0.6), 1)
         # Ensure we keep at least min_recent turns
-        split = min(split, len(self._history) - min_recent)
-        if split < 2:
+        compacted_count = min(compacted_count, len(self._history) - min_recent)
+        if compacted_count < 2:
             return
 
-        # Walk split backward to avoid breaking tool_use / tool_result pairs.
-        # A user message containing tool_result blocks must stay with the
-        # preceding assistant message that contains the matching tool_use.
-        while split > 1:
-            msg = self._history[split]
+        # Walk the cut backward to avoid breaking tool_use / tool_result
+        # pairs. A user message containing tool_result blocks must stay with
+        # the preceding assistant message that contains the matching tool_use.
+        while compacted_count > 1:
+            msg = self._history[compacted_count]
             if msg.get("role") != "user":
                 break
             content = msg.get("content")
@@ -1006,17 +1030,43 @@ async def _summarize_history(self) -> None:
                 break
             # This user message has tool_results — keep it (and its paired
             # assistant message) in the recent portion.
-            split -= 1
+            compacted_count -= 1
             # Also pull back over the preceding assistant message so the
             # pair stays together.
-            if split > 1 and self._history[split].get("role") == "assistant":
-                split -= 1
+            if compacted_count > 1 and self._history[compacted_count].get("role") == "assistant":
+                compacted_count -= 1
 
-        if split < 2:
+        if compacted_count < 2:
             return
 
-        old_turns = self._history[:split]
-        recent_turns = self._history[split:]
+        old_turns = self._history[:compacted_count]
+        recent_turns = self._history[compacted_count:]
+
+        # A prior summary carried forward from an earlier compaction isn't
+        # new material — exclude it so re-summarizing a barely-grown summary
+        # doesn't look worthwhile.
+        new_old_turns = old_turns
+        if (
+            old_turns
+            and isinstance(old_turns[0].get("content"), str)
+            and old_turns[0]["content"].lstrip().startswith(_COMPACTED_MARKER)
+        ):
+            new_old_turns = old_turns[1:]
+
+        def _approx_len(msgs: list[dict]) -> int:
+            total = 0
+            for m in msgs:
+                content = m.get("content", "")
+                total += len(content) if isinstance(content, str) else len(str(content))
+            return total
+
+        # Skip the LLM round-trip when the genuinely-new old turns are under
+        # ~10% of what we'd re-send — folding them in wouldn't shrink history
+        # enough to be worth the summarization call.
+        new_old_len = _approx_len(new_old_turns)
+        recent_len = _approx_len(recent_turns)
+        if new_old_len < 0.10 * (new_old_len + recent_len):
+            return
 
         # Serialize old turns. Pull out any prior compacted summary so we
         # UPDATE it in place rather than summarize a summary (which compounds
@@ -1107,6 +1157,7 @@ async def _summarize_history(self) -> None:
             f"{summary}"
         )
         summary_msg = {"role": "user", "content": summary_body}
+        self._last_compacted_count = compacted_count
 
         # If the recent portion starts with a user message, insert a minimal
         # assistant separator to avoid consecutive user messages (API error).
diff --git a/tests/test_history_compaction.py b/tests/test_history_compaction.py
new file mode 100644
index 00000000..c9c3ae84
--- /dev/null
+++ b/tests/test_history_compaction.py
@@ -0,0 +1,104 @@
+"""Tests for the persisted-summary compaction result (ENG-664).
+
+Covers `ChatSession.last_compaction` (what a host reads to persist the
+summary + cutoff) and the "skip if there's little new material" guard in
+`_summarize_history`. The once-per-turn guard (`_compacted_this_turn`) is
+already covered by `TestContextCompaction` in test_chat.py.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock
+
+from anton.core.session import ChatSession, ChatSessionConfig, _COMPACTED_MARKER
+from anton.core.llm.provider import LLMResponse, Usage
+from tests.conftest import make_mock_llm
+
+
+def _msg(role: str, text: str) -> dict:
+    return {"role": role, "content": text}
+
+
+def _summarize_response(text: str) -> LLMResponse:
+    return LLMResponse(content=text, usage=Usage(input_tokens=10, output_tokens=20))
+
+
+def _alternating_history(n: int, body: str) -> list[dict]:
+    return [_msg("user" if i % 2 == 0 else "assistant", body) for i in range(n)]
+
+
+class TestLastCompaction:
+    async def test_none_before_any_compaction(self):
+        session = ChatSession(ChatSessionConfig(
+            llm_client=make_mock_llm(), initial_history=_alternating_history(10, "hi"),
+        ))
+        assert session.last_compaction is None
+
+    async def test_populated_after_compaction(self):
+        original = _alternating_history(10, "x" * 50)
+        mock_llm = make_mock_llm()
+        mock_llm.summarize = AsyncMock(return_value=_summarize_response("## Goal\nTest goal"))
+
+        session = ChatSession(ChatSessionConfig(llm_client=mock_llm, initial_history=original))
+        await session._summarize_history()
+
+        # split = min(int(10*0.6), 10-4) = 6 — matches the plain 60/40 cut since
+        # none of these messages carry tool_result blocks.
+        compaction = session.last_compaction
+        assert compaction is not None
+        assert compaction["covered_through"] == 6
+        assert compaction["summary"].startswith(_COMPACTED_MARKER)
+        assert "## Goal\nTest goal" in compaction["summary"]
+        # history[0] is exactly the summary reported by last_compaction.
+        assert session.history[0]["content"] == compaction["summary"]
+        # The last 4 (uncompacted) turns survive verbatim, in order.
+        assert session.history[-4:] == original[6:]
+
+
+class TestSkipWhenLittleNewMaterial:
+    async def test_skips_llm_call_when_old_turns_are_negligible(self):
+        """Old turns tiny, recent turns huge — folding them in wouldn't
+        meaningfully shrink history, so don't pay for the LLM round-trip."""
+        history = _alternating_history(6, "ok") + _alternating_history(4, "y" * 5000)
+        mock_llm = make_mock_llm()
+        mock_llm.summarize = AsyncMock(return_value=_summarize_response("summary"))
+
+        session = ChatSession(ChatSessionConfig(llm_client=mock_llm, initial_history=history))
+        await session._summarize_history()
+
+        mock_llm.summarize.assert_not_called()
+        assert session.last_compaction is None
+        assert session.history == history
+
+    async def test_prior_summary_excluded_from_new_material_measurement(self):
+        """A leading prior-summary entry doesn't count as "new" — if the
+        actually-new old turns are negligible next to what stays verbatim,
+        skip even though the prior summary itself is large."""
+        prior_summary = _msg("user", f"{_COMPACTED_MARKER}\n" + ("z" * 5000))
+        new_old_turns = _alternating_history(4, "ok")[1:]  # tiny "new" material
+        recent_turns = _alternating_history(4, "y" * 5000)
+        history = [prior_summary, *new_old_turns, *recent_turns]
+        mock_llm = make_mock_llm()
+        mock_llm.summarize = AsyncMock(return_value=_summarize_response("summary"))
+
+        session = ChatSession(ChatSessionConfig(llm_client=mock_llm, initial_history=history))
+        await session._summarize_history()
+
+        mock_llm.summarize.assert_not_called()
+        assert session.last_compaction is None
+
+    async def test_runs_when_new_material_is_substantial(self):
+        """Sanity check for the guard's threshold: same shape as the skip
+        test above, but the new old turns are large enough to be worth it."""
+        prior_summary = _msg("user", f"{_COMPACTED_MARKER}\nprior")
+        new_old_turns = _alternating_history(4, "x" * 5000)[1:]
+        recent_turns = _alternating_history(4, "y" * 100)
+        history = [prior_summary, *new_old_turns, *recent_turns]
+        mock_llm = make_mock_llm()
+        mock_llm.summarize = AsyncMock(return_value=_summarize_response("updated summary"))
+
+        session = ChatSession(ChatSessionConfig(llm_client=mock_llm, initial_history=history))
+        await session._summarize_history()
+
+        mock_llm.summarize.assert_called_once()
+        assert session.last_compaction is not None