feat: PM brain upgrade — macro/micro agents & memory split (#123)

* feat: introduce flow_id with timestamp-based report versioning

Replace run_id with flow_id as the primary grouping concept (one flow =
one user analysis intent spanning scan + pipeline + portfolio). Reports
are now written as {timestamp}_{name}.json so load methods always return
the latest version by lexicographic sort, eliminating the latest.json
pointer pattern for new flows.

Key changes:
- report_paths.py: add generate_flow_id(), ts_now() (ms precision),
  flow_id kwarg on all path helpers; keep run_id / pointer helpers for
  backward compatibility
- ReportStore: dual-mode save/load — flow_id uses timestamped layout,
  run_id uses legacy runs/{id}/ layout with latest.json
- MongoReportStore: add flow_id field and index; run_id stays for compat
- DualReportStore: expose flow_id property
- store_factory: accept flow_id as primary param, run_id as alias
- runs.py / langgraph_engine.py: generate and thread flow_id through all
  trigger endpoints and run methods
- Tests: add flow_id coverage for all layers; 905 tests pass

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat: PM brain upgrade — macro/micro summary agents, memory split, forensic dashboard

Replaces the PM's raw-JSON context (~6,800 tokens on deep_think) with a
MAP-REDUCE compression layer using two parallel mid_think summary agents,
achieving ~70% cost reduction at the PM tier.

Architecture:
- MacroMemory: new regime-level memory class (MongoDB/JSON, separate from
  per-ticker reflexion memory) with record_macro_state/build_macro_context
- ReflexionMemory: extended with collection_name param to isolate
  micro_reflexion from the pipeline reflexion collection (with distinct
  local JSON fallback path to prevent file collision)
- Macro_Summary_Agent (mid_think): compresses scan_summary into a 1-page
  regime brief with memory injection; sentinel guard prevents LLM call on
  empty/error scan data ("NO DATA AVAILABLE - ABORT MACRO")
- Micro_Summary_Agent (mid_think): compresses holding_reviews + candidates
  into a markdown table brief with per-ticker memory injection
- Portfolio graph: parallel fan-out (prioritize_candidates → macro_summary
  ‖ micro_summary → make_pm_decision) using _last_value reducers for safe
  concurrent state writes (ADR-005 pattern)
- PM refactor: Pydantic PMDecisionSchema enforces Forensic Execution
  Dashboard output (macro_regime, forensic_report, per-trade
  macro_alignment/memory_note/position_sizing_logic); with_structured_output
  as primary path, extract_json fallback for non-conforming providers
- PM sentinel handling: "NO DATA AVAILABLE" in macro_brief substituted
  with actionable conservative guidance before LLM sees it

62 new unit tests across 4 test files covering all new components.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix: address code review — relaxed error guard, ticker_analyses, PM memory wiring

1. macro_summary_agent: relaxed error guard to only abort when scan_summary's
   sole key is "error" (partial failures with real data are now processed)
2. micro_summary_agent: now reads ticker_analyses from state and enriches
   the per-ticker table with trading graph analysis data
3. portfolio_graph: wires macro_memory and micro_memory to PM factory call
4. test_empty_state: updated test for new partial-failure behavior

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

tests/unit/test_empty_state.py

@@ -0,0 +1,120 @@
+"""Tests for empty/error state handling across agents.
+
+Validates that agents handle missing/empty/error data gracefully without
+hallucinating — particularly the NO-DATA guard in MacroSummaryAgent that
+must short-circuit before invoking the LLM.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+from tradingagents.agents.portfolio.macro_summary_agent import (
+    create_macro_summary_agent,
+)
+
+
+class TestEmptyStateGuards:
+    """Validate that agents handle missing/empty data gracefully without hallucinating."""
+
+    def test_macro_agent_empty_dict(self):
+        """Empty scan_summary dict triggers NO DATA sentinel; LLM not invoked."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({"scan_summary": {}, "messages": [], "analysis_date": ""})
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+        # LLM must NOT be invoked
+        mock_llm.invoke.assert_not_called()
+        mock_llm.with_structured_output.assert_not_called()
+
+    def test_macro_agent_none_scan(self):
+        """None scan_summary triggers NO DATA sentinel."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({"scan_summary": None, "messages": [], "analysis_date": ""})
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+    def test_macro_agent_error_key(self):
+        """scan_summary with 'error' key triggers NO DATA sentinel."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({
+            "scan_summary": {"error": "rate limit exceeded"},
+            "messages": [],
+            "analysis_date": "",
+        })
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+    def test_macro_agent_missing_scan_key(self):
+        """State dict with no scan_summary key at all triggers NO DATA sentinel.
+
+        state.get('scan_summary') returns None → should trigger guard.
+        """
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({"messages": [], "analysis_date": ""})
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+    def test_macro_agent_no_data_path_does_not_invoke_llm(self):
+        """All NO-DATA guard paths must leave the LLM untouched."""
+        no_data_states = [
+            {"scan_summary": {}, "messages": [], "analysis_date": ""},
+            {"scan_summary": None, "messages": [], "analysis_date": ""},
+            {"scan_summary": {"error": "timeout"}, "messages": [], "analysis_date": ""},
+            {"messages": [], "analysis_date": ""},
+        ]
+        for state in no_data_states:
+            mock_llm = MagicMock()
+            agent = create_macro_summary_agent(mock_llm)
+            agent(state)
+            mock_llm.invoke.assert_not_called()
+            mock_llm.__ror__.assert_not_called()
+
+    def test_macro_agent_no_data_returns_correct_sender(self):
+        """Sender is always 'macro_summary_agent' even on the NO-DATA path."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({"scan_summary": {}, "messages": [], "analysis_date": ""})
+        assert result["sender"] == "macro_summary_agent"
+
+    def test_macro_agent_no_data_macro_memory_context_empty_string(self):
+        """macro_memory_context is an empty string on the NO-DATA path."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({"scan_summary": {}, "messages": [], "analysis_date": ""})
+        assert result["macro_memory_context"] == ""
+
+    def test_macro_agent_error_only_key_triggers_sentinel(self):
+        """scan_summary that ONLY contains 'error' (no other keys) triggers guard."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({
+            "scan_summary": {"error": "vendor offline"},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        })
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+    def test_macro_agent_scan_with_data_and_error_key_proceeds(self):
+        """scan_summary with real data AND an 'error' key is NOT discarded.
+
+        Only scan_summary whose *only* key is 'error' triggers the guard.
+        Partial failures with usable data should still be compressed.
+        """
+        from langchain_core.messages import AIMessage
+        from langchain_core.runnables import RunnableLambda
+
+        mock_llm = RunnableLambda(lambda _: AIMessage(content="MACRO REGIME: neutral\nPartial data processed"))
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({
+            "scan_summary": {
+                "executive_summary": "Partial data",
+                "error": "partial failure",
+            },
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        })
+        # Should NOT be sentinel — the LLM was invoked
+        assert result["macro_brief"] != "NO DATA AVAILABLE - ABORT MACRO"

tests/unit/test_macro_memory.py

@@ -0,0 +1,188 @@
+"""Tests for MacroMemory — regime-level learning memory.
+
+Covers:
+- record_macro_state + get_recent round-trip (local JSON fallback)
+- build_macro_context formatting
+- record_outcome feedback loop
+- Ordering guarantees (newest-first)
+- Persistence across instances
+"""
+
+from __future__ import annotations
+
+import json
+
+import pytest
+
+from tradingagents.memory.macro_memory import MacroMemory
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def mem(tmp_path):
+    """Return a MacroMemory using local JSON fallback in a temp directory."""
+    return MacroMemory(fallback_path=tmp_path / "macro.json")
+
+
+# ---------------------------------------------------------------------------
+# record_macro_state + get_recent
+# ---------------------------------------------------------------------------
+
+
+class TestMacroMemoryLocalFallback:
+    """Tests using local JSON fallback (no MongoDB needed)."""
+
+    def test_record_and_retrieve(self, tmp_path):
+        """record_macro_state() stores and get_recent() retrieves."""
+        m = MacroMemory(fallback_path=tmp_path / "macro.json")
+        m.record_macro_state(
+            date="2026-03-26",
+            vix_level=25.3,
+            macro_call="risk-off",
+            sector_thesis="Energy under pressure",
+            key_themes=["rate hikes", "oil volatility"],
+        )
+        records = m.get_recent(limit=5)
+        assert len(records) == 1
+        assert records[0]["macro_call"] == "risk-off"
+        assert records[0]["vix_level"] == 25.3
+
+    def test_build_macro_context_no_history(self, tmp_path):
+        """build_macro_context() returns a message when no history."""
+        m = MacroMemory(fallback_path=tmp_path / "macro.json")
+        ctx = m.build_macro_context()
+        assert isinstance(ctx, str)
+        assert len(ctx) > 0
+
+    def test_build_macro_context_with_history(self, tmp_path):
+        """build_macro_context() includes date, macro_call, vix."""
+        m = MacroMemory(fallback_path=tmp_path / "macro.json")
+        m.record_macro_state("2026-03-20", 28.0, "risk-off", "hawkish Fed", ["rates"])
+        ctx = m.build_macro_context(limit=1)
+        assert "2026-03-20" in ctx
+        assert "risk-off" in ctx or "28" in ctx  # either VIX or call shows up
+
+    def test_record_outcome(self, tmp_path):
+        """record_outcome() attaches an outcome dict to the matching record."""
+        m = MacroMemory(fallback_path=tmp_path / "macro.json")
+        m.record_macro_state("2026-03-20", 25.0, "neutral", "mixed signals", [])
+        ok = m.record_outcome("2026-03-20", {"correct": True, "note": "regime held"})
+        assert ok is True
+        records = m.get_recent()
+        assert records[0]["outcome"] is not None
+
+    def test_get_recent_newest_first(self, tmp_path):
+        """get_recent() returns records sorted newest-first."""
+        m = MacroMemory(fallback_path=tmp_path / "macro.json")
+        m.record_macro_state("2026-03-01", 20.0, "risk-on", "", [])
+        m.record_macro_state("2026-03-26", 25.0, "risk-off", "", [])
+        records = m.get_recent(limit=2)
+        assert records[0]["regime_date"] == "2026-03-26"
+        assert records[1]["regime_date"] == "2026-03-01"
+
+
+# ---------------------------------------------------------------------------
+# Additional coverage
+# ---------------------------------------------------------------------------
+
+
+def test_macro_call_normalized_to_lowercase(mem):
+    """macro_call is stored in lowercase regardless of input casing."""
+    mem.record_macro_state("2026-03-26", 20.0, "Risk-Off", "mixed", [])
+    records = mem.get_recent()
+    assert records[0]["macro_call"] == "risk-off"
+
+
+def test_vix_stored_as_float(mem):
+    """vix_level is always stored as a float."""
+    mem.record_macro_state("2026-03-26", 22, "neutral", "flat market", [])
+    records = mem.get_recent()
+    assert isinstance(records[0]["vix_level"], float)
+
+
+def test_key_themes_stored_as_list(mem):
+    """key_themes is persisted as a list."""
+    themes = ["inflation", "rate hikes"]
+    mem.record_macro_state("2026-03-26", 20.0, "risk-off", "Fed hawkish", themes)
+    records = mem.get_recent()
+    assert records[0]["key_themes"] == themes
+
+
+def test_get_recent_limit_respected(mem):
+    """get_recent() returns at most *limit* records."""
+    for i in range(5):
+        mem.record_macro_state(f"2026-03-{i + 1:02d}", float(i), "neutral", "", [])
+    records = mem.get_recent(limit=3)
+    assert len(records) == 3
+
+
+def test_record_outcome_returns_false_for_unknown_date(mem):
+    """record_outcome() returns False when no matching date exists."""
+    result = mem.record_outcome("9999-01-01", {"correct": True})
+    assert result is False
+
+
+def test_record_outcome_only_fills_null_outcome(mem):
+    """record_outcome() will not overwrite a record that already has an outcome."""
+    mem.record_macro_state("2026-03-26", 25.0, "risk-off", "test", [])
+    mem.record_outcome("2026-03-26", {"correct": True})
+
+    # Second call should return False — outcome already set
+    result = mem.record_outcome("2026-03-26", {"correct": False})
+    assert result is False
+
+    records = mem.get_recent()
+    assert records[0]["outcome"]["correct"] is True
+
+
+def test_build_macro_context_no_prior_history_message(mem):
+    """build_macro_context() returns informative text when no records exist."""
+    ctx = mem.build_macro_context()
+    assert "No prior" in ctx
+
+
+def test_build_macro_context_shows_outcome_pending(mem):
+    """build_macro_context() shows 'pending' for records with no outcome."""
+    mem.record_macro_state("2026-03-26", 25.0, "risk-off", "test", [])
+    ctx = mem.build_macro_context()
+    assert "pending" in ctx
+
+
+def test_build_macro_context_shows_outcome_confirmed(mem):
+    """build_macro_context() shows outcome notes when outcome is set."""
+    mem.record_macro_state("2026-03-26", 25.0, "risk-off", "test", [])
+    mem.record_outcome(
+        "2026-03-26",
+        {"regime_confirmed": True, "notes": "Bear market held"},
+    )
+    ctx = mem.build_macro_context()
+    assert "Bear market held" in ctx
+
+
+def test_persistence_across_instances(tmp_path):
+    """Records written by one MacroMemory instance are visible to another."""
+    fb = tmp_path / "macro.json"
+
+    m1 = MacroMemory(fallback_path=fb)
+    m1.record_macro_state("2026-03-26", 25.0, "risk-off", "thesis", ["theme"])
+
+    m2 = MacroMemory(fallback_path=fb)
+    records = m2.get_recent()
+    assert len(records) == 1
+
+
+def test_local_file_created_on_first_write(tmp_path):
+    """The fallback JSON file is created automatically on first write."""
+    fb = tmp_path / "subdir" / "macro.json"
+    assert not fb.exists()
+
+    m = MacroMemory(fallback_path=fb)
+    m.record_macro_state("2026-03-26", 20.0, "neutral", "test", [])
+
+    assert fb.exists()
+    data = json.loads(fb.read_text())
+    assert len(data) == 1

tests/unit/test_pydantic_schema.py

@@ -0,0 +1,272 @@
+"""Tests for PMDecisionSchema Pydantic structured output model.
+
+Covers:
+- Valid payload parses correctly
+- Invalid enum values raise ValidationError
+- Required fields enforce presence
+- JSON round-trip fidelity
+- Type coercion behaviour (SellOrder.macro_driven bool coercion)
+"""
+
+from __future__ import annotations
+
+import json
+
+import pytest
+from pydantic import ValidationError
+
+from tradingagents.agents.portfolio.pm_decision_agent import (
+    BuyOrder,
+    ForensicReport,
+    HoldOrder,
+    PMDecisionSchema,
+    SellOrder,
+)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _valid_payload() -> dict:
+    """Return a fully valid PMDecisionSchema payload."""
+    return {
+        "macro_regime": "risk-off",
+        "regime_alignment_note": "Elevated VIX supports defensive posture",
+        "sells": [
+            {
+                "ticker": "AAPL",
+                "shares": 10.0,
+                "rationale": "Overvalued",
+                "macro_driven": True,
+            }
+        ],
+        "buys": [
+            {
+                "ticker": "XOM",
+                "shares": 5.0,
+                "price_target": 120.0,
+                "stop_loss": 108.0,
+                "take_profit": 138.0,
+                "sector": "Energy",
+                "rationale": "Energy tailwind",
+                "thesis": "Oil cycle upswing",
+                "macro_alignment": "Fits risk-off energy play",
+                "memory_note": "XOM held well in last risk-off",
+                "position_sizing_logic": "2% position; below 15% cap",
+            }
+        ],
+        "holds": [{"ticker": "MSFT", "rationale": "Thesis intact"}],
+        "cash_reserve_pct": 0.1,
+        "portfolio_thesis": "Defensive tilt with energy exposure",
+        "risk_summary": "Moderate risk; elevated VIX",
+        "forensic_report": {
+            "regime_alignment": "risk-off favours energy and cash",
+            "key_risks": ["oil demand drop", "rate surprise"],
+            "decision_confidence": "high",
+            "position_sizing_rationale": "All positions within 15% cap",
+        },
+    }
+
+
+# ---------------------------------------------------------------------------
+# TestPMDecisionSchema — valid payloads
+# ---------------------------------------------------------------------------
+
+
+class TestPMDecisionSchema:
+    def test_valid_payload_parses(self):
+        """A valid payload produces a PMDecisionSchema instance."""
+        obj = PMDecisionSchema(**_valid_payload())
+        assert obj.macro_regime == "risk-off"
+        assert len(obj.buys) == 1
+        assert obj.forensic_report.decision_confidence == "high"
+
+    def test_macro_regime_all_valid_values(self):
+        """Each of the four valid macro_regime values parses correctly."""
+        for regime in ("risk-on", "risk-off", "neutral", "transition"):
+            payload = _valid_payload()
+            payload["macro_regime"] = regime
+            obj = PMDecisionSchema(**payload)
+            assert obj.macro_regime == regime
+
+    def test_invalid_macro_regime_raises(self):
+        """Invalid macro_regime value raises ValidationError."""
+        payload = _valid_payload()
+        payload["macro_regime"] = "unknown"
+        with pytest.raises(ValidationError):
+            PMDecisionSchema(**payload)
+
+    def test_empty_string_macro_regime_raises(self):
+        """Empty string macro_regime raises ValidationError."""
+        payload = _valid_payload()
+        payload["macro_regime"] = ""
+        with pytest.raises(ValidationError):
+            PMDecisionSchema(**payload)
+
+    def test_invalid_decision_confidence_raises(self):
+        """Invalid decision_confidence in forensic_report raises ValidationError."""
+        payload = _valid_payload()
+        payload["forensic_report"]["decision_confidence"] = "very high"
+        with pytest.raises(ValidationError):
+            PMDecisionSchema(**payload)
+
+    def test_decision_confidence_all_valid_values(self):
+        """Each of the three valid decision_confidence values parses correctly."""
+        for level in ("high", "medium", "low"):
+            payload = _valid_payload()
+            payload["forensic_report"]["decision_confidence"] = level
+            obj = PMDecisionSchema(**payload)
+            assert obj.forensic_report.decision_confidence == level
+
+    def test_missing_forensic_report_raises(self):
+        """Missing forensic_report field raises ValidationError."""
+        payload = _valid_payload()
+        del payload["forensic_report"]
+        with pytest.raises(ValidationError):
+            PMDecisionSchema(**payload)
+
+    def test_missing_macro_regime_raises(self):
+        """Missing macro_regime field raises ValidationError."""
+        payload = _valid_payload()
+        del payload["macro_regime"]
+        with pytest.raises(ValidationError):
+            PMDecisionSchema(**payload)
+
+    def test_model_dump_json_roundtrip(self):
+        """model_dump_json() produces valid JSON that round-trips back."""
+        obj = PMDecisionSchema(**_valid_payload())
+        json_str = obj.model_dump_json()
+        data = json.loads(json_str)
+        assert data["macro_regime"] == "risk-off"
+        assert data["forensic_report"]["decision_confidence"] == "high"
+
+    def test_sell_order_macro_driven_bool(self):
+        """SellOrder.macro_driven must be a boolean (Pydantic v2 coerces str)."""
+        payload = _valid_payload()
+        payload["sells"][0]["macro_driven"] = "yes"  # string — Pydantic v2 coerces
+        obj = PMDecisionSchema(**payload)
+        assert isinstance(obj.sells[0].macro_driven, bool)
+
+    def test_empty_sells_buys_holds_allowed(self):
+        """Sells, buys, and holds can all be empty lists."""
+        payload = _valid_payload()
+        payload["sells"] = []
+        payload["buys"] = []
+        payload["holds"] = []
+        obj = PMDecisionSchema(**payload)
+        assert obj.sells == []
+        assert obj.buys == []
+        assert obj.holds == []
+
+    def test_multiple_buys_parsed(self):
+        """Multiple BuyOrder entries in buys list all parse correctly."""
+        payload = _valid_payload()
+        extra_buy = {
+            "ticker": "CVX",
+            "shares": 3.0,
+            "price_target": 160.0,
+            "stop_loss": 144.0,
+            "take_profit": 184.0,
+            "sector": "Energy",
+            "rationale": "CVX undervalued",
+            "thesis": "Same oil cycle thesis",
+            "macro_alignment": "Energy fits risk-off",
+            "memory_note": "CVX volatile in past cycles",
+            "position_sizing_logic": "1.5% position",
+        }
+        payload["buys"].append(extra_buy)
+        obj = PMDecisionSchema(**payload)
+        assert len(obj.buys) == 2
+        assert obj.buys[1].ticker == "CVX"
+
+    def test_cash_reserve_pct_stored_as_float(self):
+        """cash_reserve_pct is preserved as a float."""
+        payload = _valid_payload()
+        payload["cash_reserve_pct"] = 0.15
+        obj = PMDecisionSchema(**payload)
+        assert obj.cash_reserve_pct == 0.15
+
+
+# ---------------------------------------------------------------------------
+# TestForensicReport
+# ---------------------------------------------------------------------------
+
+
+class TestForensicReport:
+    def test_valid_forensic_report(self):
+        """ForensicReport validates correctly with all required fields."""
+        report = ForensicReport(
+            regime_alignment="risk-off supports cash",
+            key_risks=["rate spike", "credit crunch"],
+            decision_confidence="medium",
+            position_sizing_rationale="All within 10% cap",
+        )
+        assert report.decision_confidence == "medium"
+        assert len(report.key_risks) == 2
+
+    def test_key_risks_can_be_empty(self):
+        """key_risks list can be empty."""
+        report = ForensicReport(
+            regime_alignment="aligned",
+            key_risks=[],
+            decision_confidence="low",
+            position_sizing_rationale="cautious",
+        )
+        assert report.key_risks == []
+
+
+# ---------------------------------------------------------------------------
+# TestBuyOrder
+# ---------------------------------------------------------------------------
+
+
+class TestBuyOrder:
+    def test_valid_buy_order(self):
+        """BuyOrder validates with all required fields."""
+        order = BuyOrder(
+            ticker="NVDA",
+            shares=2.0,
+            price_target=900.0,
+            stop_loss=810.0,
+            take_profit=1080.0,
+            sector="Technology",
+            rationale="AI demand surge",
+            thesis="GPU dominance continues",
+            macro_alignment="Neutral regime allows tech exposure",
+            memory_note="NVDA strong in prior neutral regimes",
+            position_sizing_logic="1% position",
+        )
+        assert order.ticker == "NVDA"
+        assert order.price_target == 900.0
+
+
+# ---------------------------------------------------------------------------
+# TestSellOrder
+# ---------------------------------------------------------------------------
+
+
+class TestSellOrder:
+    def test_valid_sell_order(self):
+        """SellOrder validates with all required fields."""
+        order = SellOrder(
+            ticker="TSLA",
+            shares=5.0,
+            rationale="Overextended rally",
+            macro_driven=False,
+        )
+        assert order.ticker == "TSLA"
+        assert order.macro_driven is False
+
+
+# ---------------------------------------------------------------------------
+# TestHoldOrder
+# ---------------------------------------------------------------------------
+
+
+class TestHoldOrder:
+    def test_valid_hold_order(self):
+        """HoldOrder validates with required fields."""
+        order = HoldOrder(ticker="AMZN", rationale="Cloud thesis intact")
+        assert order.ticker == "AMZN"

tests/unit/test_summary_agents.py

@@ -0,0 +1,339 @@
+"""Tests for Macro_Summary_Agent and Micro_Summary_Agent.
+
+Strategy:
+- Empty/error state paths skip the LLM entirely — test those directly.
+- LLM-invoked paths require the mock to be a proper LangChain Runnable so that
+  ``prompt | llm`` creates a working RunnableSequence.  LangChain's pipe operator
+  calls through its own Runnable machinery — a plain MagicMock is NOT invoked via
+  Python's raw ``__call__``.  We use ``RunnableLambda`` to wrap a lambda that
+  returns a fixed AIMessage, making it fully compatible with the chain.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+from langchain_core.messages import AIMessage
+from langchain_core.runnables import RunnableLambda
+
+from tradingagents.agents.portfolio.macro_summary_agent import (
+    create_macro_summary_agent,
+)
+from tradingagents.agents.portfolio.micro_summary_agent import (
+    create_micro_summary_agent,
+)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_runnable_llm(content: str = "MACRO REGIME: risk-off\nKEY NUMBERS: VIX=25"):
+    """Build a LangChain-compatible LLM stub via RunnableLambda.
+
+    ``ChatPromptTemplate | llm`` creates a ``RunnableSequence``.  LangChain
+    dispatches through its own Runnable protocol — the LLM must implement
+    ``.invoke()`` as a Runnable, not just as a Python callable.
+    ``RunnableLambda`` satisfies that contract.
+
+    Returns:
+        A ``RunnableLambda`` that always returns ``AIMessage(content=content)``.
+    """
+    ai_msg = AIMessage(content=content)
+    return RunnableLambda(lambda _: ai_msg)
+
+
+# Keep backward-compatible alias used by some tests that destructure a tuple
+def _make_chain_mock(content: str = "MACRO REGIME: risk-off\nKEY NUMBERS: VIX=25"):
+    """Return (llm_runnable, None) — second element kept for API compatibility."""
+    return _make_runnable_llm(content), None
+
+
+# ---------------------------------------------------------------------------
+# MacroSummaryAgent — NO-DATA guard paths (LLM never called)
+# ---------------------------------------------------------------------------
+
+
+class TestMacroSummaryAgentNoDataGuard:
+    """Verify the abort-early guard fires and LLM is not invoked."""
+
+    def test_empty_scan_summary_returns_sentinel(self):
+        """Empty scan_summary dict triggers NO DATA sentinel without LLM call."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        state = {"scan_summary": {}, "messages": [], "analysis_date": "2026-03-26"}
+        result = agent(state)
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+        mock_llm.invoke.assert_not_called()
+
+    def test_none_scan_summary_returns_sentinel(self):
+        """None scan_summary triggers NO DATA sentinel."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        state = {"scan_summary": None, "messages": [], "analysis_date": "2026-03-26"}
+        result = agent(state)
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+    def test_error_key_in_scan_returns_sentinel(self):
+        """scan_summary with 'error' key triggers NO DATA sentinel."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        state = {
+            "scan_summary": {"error": "vendor timeout"},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+    def test_missing_scan_key_returns_sentinel(self):
+        """State dict with no scan_summary key at all triggers NO DATA sentinel."""
+        mock_llm = MagicMock()
+        agent = create_macro_summary_agent(mock_llm)
+        result = agent({"messages": [], "analysis_date": "2026-03-26"})
+        assert result["macro_brief"] == "NO DATA AVAILABLE - ABORT MACRO"
+
+
+# ---------------------------------------------------------------------------
+# MacroSummaryAgent — required state keys returned
+# ---------------------------------------------------------------------------
+
+
+class TestMacroSummaryAgentReturnShape:
+    """Verify that every execution path returns the expected state keys."""
+
+    def test_no_data_path_returns_required_keys(self):
+        """NO-DATA guard path returns all required state keys."""
+        agent = create_macro_summary_agent(MagicMock())
+        result = agent({"scan_summary": {}, "messages": [], "analysis_date": ""})
+        assert "macro_brief" in result
+        assert "macro_memory_context" in result
+        assert "sender" in result
+        assert result["sender"] == "macro_summary_agent"
+
+    def test_no_data_path_messages_is_list(self):
+        """NO-DATA guard path returns messages as a list."""
+        agent = create_macro_summary_agent(MagicMock())
+        result = agent({"scan_summary": {}, "messages": [], "analysis_date": ""})
+        assert isinstance(result["messages"], list)
+
+    def test_llm_path_returns_required_keys(self):
+        """LLM-invoked path returns all required state keys."""
+        llm_mock, _ = _make_chain_mock("MACRO REGIME: neutral\nKEY NUMBERS: VIX=18")
+        agent = create_macro_summary_agent(llm_mock)
+        state = {
+            "scan_summary": {"executive_summary": "Flat markets"},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert "macro_brief" in result
+        assert "macro_memory_context" in result
+        assert "sender" in result
+        assert result["sender"] == "macro_summary_agent"
+
+    def test_llm_path_macro_brief_contains_llm_output(self):
+        """macro_brief contains the LLM's returned content."""
+        content = "MACRO REGIME: risk-on\nKEY NUMBERS: VIX=12"
+        llm_mock, _ = _make_chain_mock(content)
+        agent = create_macro_summary_agent(llm_mock)
+        state = {
+            "scan_summary": {"executive_summary": "Bull run"},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert result["macro_brief"] == content
+
+
+# ---------------------------------------------------------------------------
+# MacroSummaryAgent — macro_memory integration
+# ---------------------------------------------------------------------------
+
+
+class TestMacroSummaryAgentMemory:
+    """Verify macro_memory interaction without hitting MongoDB."""
+
+    def test_no_memory_context_is_empty_string_on_no_data_path(self):
+        """NO-DATA path returns empty string for macro_memory_context."""
+        agent = create_macro_summary_agent(MagicMock())
+        result = agent({"scan_summary": {}, "messages": [], "analysis_date": ""})
+        assert result["macro_memory_context"] == ""
+
+    def test_memory_context_injected_into_result(self, tmp_path):
+        """When macro_memory is provided, macro_memory_context is populated."""
+        from tradingagents.memory.macro_memory import MacroMemory
+
+        mem = MacroMemory(fallback_path=tmp_path / "macro.json")
+        mem.record_macro_state("2026-03-20", 25.0, "risk-off", "hawkish", ["rates"])
+
+        llm_mock, _ = _make_chain_mock("MACRO REGIME: risk-off\nKEY NUMBERS: VIX=25")
+        agent = create_macro_summary_agent(llm_mock, macro_memory=mem)
+        state = {
+            "scan_summary": {"executive_summary": "Risk-off conditions persist"},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        # Past context built from the single recorded state should reference date
+        assert "2026-03-20" in result["macro_memory_context"]
+
+
+# ---------------------------------------------------------------------------
+# MicroSummaryAgent — return shape
+# ---------------------------------------------------------------------------
+
+
+class TestMicroSummaryAgentReturnShape:
+    """Verify the micro summary agent returns all required state keys."""
+
+    def test_result_has_required_keys(self):
+        """Agent returns all required state keys."""
+        llm_mock, _ = _make_chain_mock("HOLDINGS TABLE:\n| TICKER | ACTION |")
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": "{}",
+            "prioritized_candidates": "[]",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert "micro_brief" in result
+        assert "micro_memory_context" in result
+        assert "sender" in result
+        assert result["sender"] == "micro_summary_agent"
+
+    def test_micro_brief_contains_llm_output(self):
+        """micro_brief contains the LLM's returned content."""
+        content = "HOLDINGS TABLE:\n| AAPL | HOLD | 180 | green | good |"
+        llm_mock, _ = _make_chain_mock(content)
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": '{"AAPL": {"recommendation": "HOLD", "confidence": "high"}}',
+            "prioritized_candidates": "[]",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert result["micro_brief"] == content
+
+    def test_sender_always_set(self):
+        """sender key is always 'micro_summary_agent'."""
+        llm_mock, _ = _make_chain_mock("brief output")
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": "{}",
+            "prioritized_candidates": "[]",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "",
+        }
+        result = agent(state)
+        assert result["sender"] == "micro_summary_agent"
+
+
+# ---------------------------------------------------------------------------
+# MicroSummaryAgent — malformed input handling
+# ---------------------------------------------------------------------------
+
+
+class TestMicroSummaryAgentMalformedInput:
+    """Verify that malformed JSON in state fields does not raise exceptions."""
+
+    def test_invalid_holding_reviews_json_handled_gracefully(self):
+        """Malformed JSON in holding_reviews does not raise."""
+        llm_mock, _ = _make_chain_mock("brief")
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": "not valid json{{",
+            "prioritized_candidates": "[]",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert "micro_brief" in result
+
+    def test_invalid_candidates_json_handled_gracefully(self):
+        """Malformed JSON in prioritized_candidates does not raise."""
+        llm_mock, _ = _make_chain_mock("brief")
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": "{}",
+            "prioritized_candidates": "also broken",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert "micro_brief" in result
+
+    def test_both_inputs_malformed_does_not_raise(self):
+        """Both holding_reviews and prioritized_candidates malformed — no raise."""
+        llm_mock, _ = _make_chain_mock("brief")
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": "not valid json{{",
+            "prioritized_candidates": "also broken",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert "micro_brief" in result
+
+    def test_none_holding_reviews_handled(self):
+        """None holding_reviews falls back gracefully."""
+        llm_mock, _ = _make_chain_mock("brief")
+        agent = create_micro_summary_agent(llm_mock)
+        state = {
+            "holding_reviews": None,
+            "prioritized_candidates": None,
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        assert "micro_brief" in result
+
+    def test_missing_state_keys_handled(self):
+        """Missing optional keys in state do not cause a KeyError."""
+        llm_mock, _ = _make_chain_mock("brief")
+        agent = create_micro_summary_agent(llm_mock)
+        # Minimal state — only messages is truly required by the chain call
+        state = {"messages": [], "analysis_date": "2026-03-26"}
+        result = agent(state)
+        assert "micro_brief" in result
+
+
+# ---------------------------------------------------------------------------
+# MicroSummaryAgent — memory integration
+# ---------------------------------------------------------------------------
+
+
+class TestMicroSummaryAgentMemory:
+    """Verify micro_memory interaction."""
+
+    def test_micro_memory_context_includes_ticker_history(self, tmp_path):
+        """When micro_memory is provided with history, context string includes it."""
+        from tradingagents.memory.reflexion import ReflexionMemory
+
+        mem = ReflexionMemory(fallback_path=tmp_path / "reflexion.json")
+        mem.record_decision("AAPL", "2026-03-20", "BUY", "Strong momentum", "high")
+
+        llm_mock, _ = _make_chain_mock("brief")
+        agent = create_micro_summary_agent(llm_mock, micro_memory=mem)
+        state = {
+            "holding_reviews": '{"AAPL": {"recommendation": "HOLD", "confidence": "high"}}',
+            "prioritized_candidates": "[]",
+            "ticker_analyses": {},
+            "messages": [],
+            "analysis_date": "2026-03-26",
+        }
+        result = agent(state)
+        # micro_memory_context is JSON-serialised dict — AAPL should appear
+        assert "AAPL" in result["micro_memory_context"]

tradingagents/agents/portfolio/__init__.py

@@ -4,8 +4,12 @@ from __future__ import annotations
 
 from tradingagents.agents.portfolio.holding_reviewer import create_holding_reviewer
 from tradingagents.agents.portfolio.pm_decision_agent import create_pm_decision_agent
+from tradingagents.agents.portfolio.macro_summary_agent import create_macro_summary_agent
+from tradingagents.agents.portfolio.micro_summary_agent import create_micro_summary_agent
 
 __all__ = [
     "create_holding_reviewer",
     "create_pm_decision_agent",
+    "create_macro_summary_agent",
+    "create_micro_summary_agent",
 ]

tradingagents/agents/portfolio/macro_summary_agent.py

@@ -0,0 +1,200 @@
+"""Macro Summary Agent.
+
+Pure-reasoning LLM node (no tools). Reads the macro scan output and compresses
+it into a concise 1-page regime brief, injecting past macro regime memory.
+
+Pattern: ``create_macro_summary_agent(llm, macro_memory)`` → closure
+(mirrors macro_synthesis pattern).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import re
+
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+
+from tradingagents.memory.macro_memory import MacroMemory
+
+logger = logging.getLogger(__name__)
+
+
+def create_macro_summary_agent(llm, macro_memory: MacroMemory | None = None):
+    """Create a macro summary agent node.
+
+    Args:
+        llm: A LangChain chat model instance (deep_think recommended).
+        macro_memory: Optional MacroMemory instance for regime history injection
+            and post-call persistence. When None, memory features are skipped.
+
+    Returns:
+        A node function ``macro_summary_node(state)`` compatible with LangGraph.
+    """
+
+    def macro_summary_node(state: dict) -> dict:
+        scan_summary = state.get("scan_summary") or {}
+
+        # Guard: abort early if scan data is absent or *only* contains an error
+        # (partial failures with real data + an "error" key are still usable)
+        if not scan_summary or (isinstance(scan_summary, dict) and scan_summary.keys() == {"error"}):
+            return {
+                "messages": [],
+                "macro_brief": "NO DATA AVAILABLE - ABORT MACRO",
+                "macro_memory_context": "",
+                "sender": "macro_summary_agent",
+            }
+
+        # ------------------------------------------------------------------
+        # Compress scan data to save tokens
+        # ------------------------------------------------------------------
+        executive_summary: str = scan_summary.get("executive_summary", "Not available")
+
+        macro_context: dict = scan_summary.get("macro_context", {})
+        macro_context_str = (
+            f"Economic cycle: {macro_context.get('economic_cycle', 'N/A')}\n"
+            f"Central bank stance: {macro_context.get('central_bank_stance', 'N/A')}\n"
+            f"Geopolitical risks: {macro_context.get('geopolitical_risks', 'N/A')}"
+        )
+
+        key_themes: list = scan_summary.get("key_themes", [])
+        key_themes_str = "\n".join(
+            f"- {t.get('theme', '?')} [{t.get('conviction', '?')}] "
+            f"({t.get('timeframe', '?')}): {t.get('description', '')}"
+            for t in key_themes
+        ) or "None"
+
+        # Strip verbose rationale — retain only what the brief needs
+        ticker_conviction = [
+            {
+                "ticker": t.get("ticker", "?"),
+                "conviction": t.get("conviction", "?"),
+                "thesis_angle": t.get("thesis_angle", "?"),
+            }
+            for t in scan_summary.get("stocks_to_investigate", [])
+        ]
+        ticker_conviction_str = json.dumps(ticker_conviction, indent=2) or "[]"
+
+        risk_factors: list = scan_summary.get("risk_factors", [])
+        risk_factors_str = "\n".join(f"- {r}" for r in risk_factors) or "None"
+
+        # ------------------------------------------------------------------
+        # Past macro regime history
+        # ------------------------------------------------------------------
+        if macro_memory is not None:
+            past_context = macro_memory.build_macro_context(limit=3)
+        else:
+            past_context = "No prior macro regime history available."
+
+        # ------------------------------------------------------------------
+        # Build system message
+        # ------------------------------------------------------------------
+        system_message = (
+            "You are a macro strategist compressing a scanner report into a concise regime brief.\n\n"
+            "## Past Macro Regime History\n"
+            f"{past_context}\n\n"
+            "## Current Scan Data\n"
+            "### Executive Summary\n"
+            f"{executive_summary}\n\n"
+            "### Macro Context\n"
+            f"{macro_context_str}\n\n"
+            "### Key Themes\n"
+            f"{key_themes_str}\n\n"
+            "### Candidate Tickers (conviction only)\n"
+            f"{ticker_conviction_str}\n\n"
+            "### Risk Factors\n"
+            f"{risk_factors_str}\n\n"
+            "Produce a structured macro brief in this exact format:\n\n"
+            "MACRO REGIME: [risk-on|risk-off|neutral|transition]\n\n"
+            "KEY NUMBERS: [retain ALL exact numeric values — VIX levels, %, yield values, "
+            "sector weightings — do not round or omit]\n\n"
+            "TOP 3 THEMES:\n"
+            "1. [theme]: [description — retain all numbers]\n"
+            "2. [theme]: [description — retain all numbers]\n"
+            "3. [theme]: [description — retain all numbers]\n\n"
+            "MACRO-ALIGNED TICKERS: [list tickers with high conviction and why they fit the regime]\n\n"
+            "REGIME MEMORY NOTE: [any relevant lesson from past macro history that applies now]\n\n"
+            "IMPORTANT: Do NOT restrict yourself to a word count. Retain every numeric value from the "
+            "scan data. If the scan data is incomplete, note it explicitly — do not guess or extrapolate."
+        )
+
+        prompt = ChatPromptTemplate.from_messages(
+            [
+                (
+                    "system",
+                    "You are a helpful AI assistant, collaborating with other assistants."
+                    " You have access to the following tools: {tool_names}.\n{system_message}"
+                    " For your reference, the current date is {current_date}.",
+                ),
+                MessagesPlaceholder(variable_name="messages"),
+            ]
+        )
+
+        prompt = prompt.partial(system_message=system_message)
+        prompt = prompt.partial(tool_names="none")
+        prompt = prompt.partial(current_date=state.get("analysis_date", ""))
+
+        chain = prompt | llm
+        result = chain.invoke(state["messages"])
+
+        # ------------------------------------------------------------------
+        # Persist macro regime call to memory
+        # ------------------------------------------------------------------
+        if macro_memory is not None:
+            _persist_regime(result.content, scan_summary, macro_memory, state)
+
+        return {
+            "messages": [result],
+            "macro_brief": result.content,
+            "macro_memory_context": past_context,
+            "sender": "macro_summary_agent",
+        }
+
+    return macro_summary_node
+
+
+def _persist_regime(
+    brief: str,
+    scan_summary: dict,
+    macro_memory: MacroMemory,
+    state: dict,
+) -> None:
+    """Extract MACRO REGIME line and persist to MacroMemory.
+
+    Fails silently — memory persistence must never break the pipeline.
+    """
+    try:
+        macro_call = "neutral"
+        match = re.search(r"MACRO REGIME:\s*([^\n]+)", brief, re.IGNORECASE)
+        if match:
+            raw_call = match.group(1).strip().lower()
+            # Normalise to one of the four valid values
+            for valid in ("risk-on", "risk-off", "transition", "neutral"):
+                if valid in raw_call:
+                    macro_call = valid
+                    break
+
+        # Best-effort VIX extraction — scan data rarely includes a bare float
+        vix_level = 0.0
+        vix_match = re.search(r"VIX[:\s]+([0-9]+(?:\.[0-9]+)?)", brief, re.IGNORECASE)
+        if vix_match:
+            try:
+                vix_level = float(vix_match.group(1))
+            except ValueError:
+                pass
+
+        key_themes = [
+            t.get("theme", "") for t in scan_summary.get("key_themes", []) if t.get("theme")
+        ]
+        sector_thesis = scan_summary.get("executive_summary", "")[:500]
+        analysis_date = state.get("analysis_date", "")
+
+        macro_memory.record_macro_state(
+            date=analysis_date,
+            vix_level=vix_level,
+            macro_call=macro_call,
+            sector_thesis=sector_thesis,
+            key_themes=key_themes,
+        )
+    except Exception:
+        logger.warning("macro_summary_agent: failed to persist regime to memory", exc_info=True)

tradingagents/agents/portfolio/micro_summary_agent.py

@@ -0,0 +1,185 @@
+"""Micro Summary Agent.
+
+Pure-reasoning LLM node (no tools). Compresses holding reviews and ranked
+candidates into a 1-page micro brief, injecting per-ticker reflexion memory.
+
+Pattern: ``create_micro_summary_agent(llm, micro_memory)`` → closure
+(mirrors macro_synthesis pattern).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+
+from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+
+from tradingagents.memory.reflexion import ReflexionMemory
+
+logger = logging.getLogger(__name__)
+
+
+def create_micro_summary_agent(llm, micro_memory: ReflexionMemory | None = None):
+    """Create a micro summary agent node.
+
+    Args:
+        llm: A LangChain chat model instance (mid_think or deep_think recommended).
+        micro_memory: Optional ReflexionMemory instance for per-ticker history
+            injection. When None, memory features are skipped.
+
+    Returns:
+        A node function ``micro_summary_node(state)`` compatible with LangGraph.
+    """
+
+    def micro_summary_node(state: dict) -> dict:
+        analysis_date = state.get("analysis_date") or ""
+
+        # ------------------------------------------------------------------
+        # Parse inputs — handle missing / malformed gracefully
+        # ------------------------------------------------------------------
+        holding_reviews_raw = state.get("holding_reviews") or "{}"
+        candidates_raw = state.get("prioritized_candidates") or "[]"
+
+        holding_reviews: dict = _parse_json_safely(holding_reviews_raw, default={})
+        candidates: list = _parse_json_safely(candidates_raw, default=[])
+
+        # Optional: per-ticker trading graph analyses (fundamentals, technicals, etc.)
+        ticker_analyses: dict = state.get("ticker_analyses") or {}
+
+        # ------------------------------------------------------------------
+        # Collect all tickers and retrieve per-ticker memory
+        # ------------------------------------------------------------------
+        holding_tickers = list(holding_reviews.keys()) if isinstance(holding_reviews, dict) else []
+        candidate_tickers = [
+            c.get("ticker", "") for c in candidates if isinstance(c, dict) and c.get("ticker")
+        ]
+        all_tickers = list(dict.fromkeys(holding_tickers + candidate_tickers))  # preserve order, dedupe
+
+        ticker_memory_dict: dict[str, str] = {}
+        if micro_memory is not None:
+            for ticker in all_tickers:
+                ticker_memory_dict[ticker] = micro_memory.build_context(ticker, limit=2)
+
+        ticker_memory_str = json.dumps(ticker_memory_dict)
+
+        # ------------------------------------------------------------------
+        # Build concise per-ticker input table
+        # ------------------------------------------------------------------
+        table_rows: list[str] = []
+
+        for ticker in holding_tickers:
+            review = holding_reviews.get(ticker, {}) if isinstance(holding_reviews, dict) else {}
+            if not isinstance(review, dict):
+                review = {}
+            rec = review.get("recommendation", "?")
+            confidence = review.get("confidence", "")
+            label = f"HOLDING | {rec} | conf:{confidence}" if confidence else f"HOLDING | {rec}"
+            # Enrich with trading graph analysis if available
+            analysis = ticker_analyses.get(ticker, {}) if isinstance(ticker_analyses, dict) else {}
+            key_number = analysis.get("final_trade_decision", "")[:80] if isinstance(analysis, dict) else ""
+            key_number = key_number or "-"
+            memory_snippet = (ticker_memory_dict.get(ticker, "")[:100] or "no memory")
+            table_rows.append(f"{ticker} | {label} | {key_number} | {memory_snippet}")
+
+        for c in candidates:
+            if not isinstance(c, dict):
+                continue
+            ticker = c.get("ticker", "?")
+            conviction = c.get("conviction", "?")
+            thesis = c.get("thesis_angle", "?")
+            score = c.get("score", "")
+            key_number = f"score:{score}" if score != "" else "-"
+            label = f"CANDIDATE | {conviction} | {thesis}"
+            memory_snippet = (ticker_memory_dict.get(ticker, "")[:100] or "no memory")
+            table_rows.append(f"{ticker} | {label} | {key_number} | {memory_snippet}")
+
+        ticker_table = "\n".join(table_rows) or "No tickers available."
+
+        # Serialise full detail for LLM context
+        holding_reviews_str = (
+            json.dumps(holding_reviews, indent=2)
+            if holding_reviews
+            else "No holding reviews available."
+        )
+        candidates_str = (
+            json.dumps(candidates, indent=2)
+            if candidates
+            else "No candidates available."
+        )
+
+        # ------------------------------------------------------------------
+        # Build system message
+        # ------------------------------------------------------------------
+        system_message = (
+            "You are a micro analyst compressing position-level data into a concise brief "
+            "for a portfolio manager.\n\n"
+            "## Per-Ticker Data\n"
+            f"{ticker_table}\n\n"
+            "## Holding Reviews (full detail)\n"
+            f"{holding_reviews_str}\n\n"
+            "## Prioritized Candidates (full detail)\n"
+            f"{candidates_str}\n\n"
+            "Produce a structured micro brief in this exact format:\n\n"
+            "HOLDINGS TABLE:\n"
+            "| TICKER | ACTION | KEY NUMBER | FLAG | MEMORY |\n"
+            "|--------|--------|------------|------|--------|\n"
+            "[one row per holding — if data is missing, write \"NO DATA\" in KEY NUMBER and FLAG columns]\n\n"
+            "CANDIDATES TABLE:\n"
+            "| TICKER | CONVICTION | THESIS ANGLE | KEY NUMBER | FLAG | MEMORY |\n"
+            "|--------|------------|--------------|------------|------|--------|\n"
+            "[one row per candidate — if data is missing, write \"NO DATA\"]\n\n"
+            "RED FLAGS: [list any tickers with accounting anomalies, high debt, or historical losses "
+            "— cite exact numbers]\n"
+            "GREEN FLAGS: [list tickers with strong momentum, insider buying, or positive memory "
+            "— cite exact numbers]\n\n"
+            "IMPORTANT: Retain exact debt ratios, P/E multiples, EPS values, and unrealized P&L "
+            "percentages. Never round or omit a numeric value. If a ticker has no data, write "
+            "\"NO DATA\" — do not guess."
+        )
+
+        prompt = ChatPromptTemplate.from_messages(
+            [
+                (
+                    "system",
+                    "You are a helpful AI assistant, collaborating with other assistants."
+                    " You have access to the following tools: {tool_names}.\n{system_message}"
+                    " For your reference, the current date is {current_date}.",
+                ),
+                MessagesPlaceholder(variable_name="messages"),
+            ]
+        )
+
+        prompt = prompt.partial(system_message=system_message)
+        prompt = prompt.partial(tool_names="none")
+        prompt = prompt.partial(current_date=analysis_date)
+
+        chain = prompt | llm
+        result = chain.invoke(state["messages"])
+
+        return {
+            "messages": [result],
+            "micro_brief": result.content,
+            "micro_memory_context": ticker_memory_str,
+            "sender": "micro_summary_agent",
+        }
+
+    return micro_summary_node
+
+
+def _parse_json_safely(raw: str, *, default):
+    """Parse a JSON string, returning *default* on any parse error.
+
+    Args:
+        raw:     Raw string (may be JSON or empty/malformed).
+        default: Value to return when parsing fails.
+    """
+    if not raw or not raw.strip():
+        return default
+    try:
+        return json.loads(raw)
+    except (json.JSONDecodeError, TypeError):
+        logger.warning(
+            "micro_summary_agent: could not parse JSON input (first 100): %s",
+            raw[:100],
+        )
+        return default

tradingagents/agents/portfolio/pm_decision_agent.py

@@ -1,7 +1,7 @@
 """Portfolio Manager Decision Agent.
 
-Pure reasoning LLM agent (no tools).  Synthesizes risk metrics, holding
+Pure reasoning LLM agent (no tools).  Synthesizes macro and micro briefs into a
-reviews, and prioritized candidates into a structured investment decision.
+fully auditable, structured investment decision via Pydantic-schema-driven output.
 
 Pattern: ``create_pm_decision_agent(llm)`` → closure (macro_synthesis pattern).
 """
@@ -10,20 +10,95 @@ from __future__ import annotations
 
 import json
 import logging
+from typing import Literal
 
+from langchain_core.messages import AIMessage
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+from pydantic import BaseModel
 
 from tradingagents.agents.utils.json_utils import extract_json
 
 logger = logging.getLogger(__name__)
 
 
-def create_pm_decision_agent(llm, config: dict | None = None):
+# ---------------------------------------------------------------------------
+# Pydantic output schema
+# ---------------------------------------------------------------------------
+
+
+class ForensicReport(BaseModel):
+    """Audit trail for the PM's decision confidence and risk posture."""
+
+    regime_alignment: str
+    key_risks: list[str]
+    decision_confidence: Literal["high", "medium", "low"]
+    position_sizing_rationale: str
+
+
+class BuyOrder(BaseModel):
+    """A fully justified buy order with risk parameters."""
+
+    ticker: str
+    shares: float
+    price_target: float
+    stop_loss: float
+    take_profit: float
+    sector: str
+    rationale: str
+    thesis: str
+    macro_alignment: str
+    memory_note: str
+    position_sizing_logic: str
+
+
+class SellOrder(BaseModel):
+    """A sell order with macro-driven flag."""
+
+    ticker: str
+    shares: float
+    rationale: str
+    macro_driven: bool
+
+
+class HoldOrder(BaseModel):
+    """A hold decision with rationale."""
+
+    ticker: str
+    rationale: str
+
+
+class PMDecisionSchema(BaseModel):
+    """Full PM decision output — structured and auditable."""
+
+    macro_regime: Literal["risk-on", "risk-off", "neutral", "transition"]
+    regime_alignment_note: str
+    sells: list[SellOrder]
+    buys: list[BuyOrder]
+    holds: list[HoldOrder]
+    cash_reserve_pct: float
+    portfolio_thesis: str
+    risk_summary: str
+    forensic_report: ForensicReport
+
+
+# ---------------------------------------------------------------------------
+# Factory
+# ---------------------------------------------------------------------------
+
+
+def create_pm_decision_agent(
+    llm,
+    config: dict | None = None,
+    macro_memory=None,
+    micro_memory=None,
+):
     """Create a PM decision agent node.
 
     Args:
         llm: A LangChain chat model instance (deep_think recommended).
         config: Portfolio configuration dictionary containing constraints.
+        macro_memory: Reserved for future direct retrieval; briefs come via state.
+        micro_memory: Reserved for future direct retrieval; briefs come via state.
 
     Returns:
         A node function ``pm_decision_node(state)`` compatible with LangGraph.
@@ -38,54 +113,56 @@ def create_pm_decision_agent(llm, config: dict | None = None):
 
     def pm_decision_node(state):
         analysis_date = state.get("analysis_date") or ""
+
+        # Read brief fields written by upstream summary agents
+        _macro_brief_raw = state.get("macro_brief") or ""
+        if not _macro_brief_raw or "NO DATA AVAILABLE" in _macro_brief_raw:
+            # Macro scanner failed — give PM explicit guidance rather than passing sentinel
+            macro_brief = (
+                "MACRO DATA UNAVAILABLE: No scanner output was produced. "
+                "Proceed with micro brief only. Adopt a conservative posture: "
+                "hold existing positions and avoid new buys unless micro thesis is very strong."
+            )
+        else:
+            macro_brief = _macro_brief_raw
+        micro_brief = state.get("micro_brief") or "No micro brief available."
+
+        # Build compressed portfolio summary — avoid passing the full blob
         portfolio_data_str = state.get("portfolio_data") or "{}"
-        risk_metrics_str = state.get("risk_metrics") or "{}"
+        try:
-        holding_reviews_str = state.get("holding_reviews") or "{}"
+            pd_raw = json.loads(portfolio_data_str)
-        prioritized_candidates_str = state.get("prioritized_candidates") or "[]"
+            portfolio = pd_raw.get("portfolio") or {}
+            holdings = pd_raw.get("holdings") or []
+            compressed = {
+                "cash": portfolio.get("cash", 0.0),
+                "n_positions": len(holdings),
+                "total_value": portfolio.get("total_value"),
+            }
+            compressed_str = json.dumps(compressed)
+        except Exception:
+            # Fallback: truncated raw string keeps token count bounded
+            compressed_str = portfolio_data_str[:200]
 
-        context = f"""## Portfolio Constraints
+        context = (
-{constraints_str}
+            f"## Portfolio Constraints\n{constraints_str}\n\n"
-
+            f"## Portfolio Summary\n{compressed_str}\n\n"
-## Portfolio Data
+            f"## Input A — Macro Context & Memory\n{macro_brief}\n\n"
-{portfolio_data_str}
+            f"## Input B — Micro Context & Memory\n{micro_brief}\n"
-
+        )
-## Risk Metrics
-{risk_metrics_str}
-
-## Holding Reviews
-{holding_reviews_str}
-
-## Prioritized Candidates
-{prioritized_candidates_str}
-"""
 
         system_message = (
-            "You are a portfolio manager making final investment decisions. "
+            "You are a portfolio manager making final, risk-adjusted investment decisions. "
-            "Given the constraints, risk metrics, holding reviews, and prioritized investment candidates, "
+            "You receive two inputs: (A) a macro regime brief with memory, and (B) a micro brief "
-            "produce a structured JSON investment decision. "
+            "with per-ticker signals and memory. Synthesize A and B into a Forensic Execution "
+            "Dashboard — a fully auditable decision plan where every trade is justified by both "
+            "macro alignment and micro thesis.\n\n"
             "## CONSTRAINTS COMPLIANCE:\n"
-            "You MUST ensure your suggested buys and position sizes adhere to the portfolio constraints. "
+            "You MUST ensure all buys adhere to the portfolio constraints. "
-            "If a high-conviction candidate would exceed the max position size or sector limit, "
+            "If a high-conviction candidate exceeds max position size or sector limit, "
-            "YOU MUST adjust the suggested 'shares' downward to fit within the limit. "
+            "adjust shares downward to fit. For every BUY: set stop_loss (5-15% below entry) "
-            "Do not suggest buys that you know will be rejected by the risk engine.\n\n"
+            "and take_profit (10-30% above entry). "
-            "Consider: reducing risk where metrics are poor, acting on SELL recommendations, "
+            "Every buy must have macro_alignment (how it fits the regime), "
-            "and adding positions in high-conviction candidates that pass constraints. "
+            "memory_note (any relevant historical lesson), and position_sizing_logic.\n\n"
-            "For every BUY you MUST set a stop_loss price (maximum acceptable loss level, "
-            "typically 5-15% below entry) and a take_profit price (expected sell target, "
-            "typically 10-30% above entry based on your thesis). "
-            "Output ONLY valid JSON matching this exact schema:\n"
-            "{\n"
-            '  "sells": [{"ticker": "...", "shares": 0.0, "rationale": "..."}],\n'
-            '  "buys": [{"ticker": "...", "shares": 0.0, "price_target": 0.0, '
-            '"stop_loss": 0.0, "take_profit": 0.0, '
-            '"sector": "...", "rationale": "...", "thesis": "..."}],\n'
-            '  "holds": [{"ticker": "...", "rationale": "..."}],\n'
-            '  "cash_reserve_pct": 0.10,\n'
-            '  "portfolio_thesis": "...",\n'
-            '  "risk_summary": "..."\n'
-            "}\n\n"
-            "IMPORTANT: Output ONLY valid JSON. Start your response with '{' and end with '}'. "
-            "Do NOT use markdown code fences. Do NOT include any explanation or preamble before or after the JSON.\n\n"
             f"{context}"
         )
 
@@ -100,27 +177,41 @@ def create_pm_decision_agent(llm, config: dict | None = None):
                 MessagesPlaceholder(variable_name="messages"),
             ]
         )
-
         prompt = prompt.partial(system_message=system_message)
         prompt = prompt.partial(tool_names="none")
         prompt = prompt.partial(current_date=analysis_date)
 
-        chain = prompt | llm
+        # Primary path: structured output via Pydantic schema
-        result = chain.invoke(state["messages"])
+        structured_llm = llm.with_structured_output(PMDecisionSchema)
+        chain = prompt | structured_llm
 
-        raw = result.content or "{}"
         try:
-            parsed = extract_json(raw)
+            result = chain.invoke(state["messages"])
-            decision_str = json.dumps(parsed)
+            decision_str = result.model_dump_json()
-        except (ValueError, json.JSONDecodeError):
+        except Exception as exc:
             logger.warning(
-                "pm_decision_agent: could not extract JSON; storing raw (first 200): %s",
+                "pm_decision_agent: structured output failed (%s); falling back to raw", exc
-                raw[:200],
             )
-            decision_str = raw
+            # Fallback: plain LLM + extract_json
+            chain_raw = prompt | llm
+            raw_result = chain_raw.invoke(state["messages"])
+            raw = raw_result.content or "{}"
+            try:
+                parsed = extract_json(raw)
+                decision_str = json.dumps(parsed)
+            except (ValueError, json.JSONDecodeError):
+                decision_str = raw
+            return {
+                "messages": [raw_result],
+                "pm_decision": decision_str,
+                "sender": "pm_decision_agent",
+            }
 
+        # with_structured_output returns the Pydantic model directly, not an AIMessage.
+        # Wrap in a synthetic AIMessage so downstream message-history nodes stay consistent.
+        synthetic_msg = AIMessage(content=decision_str)
         return {
-            "messages": [result],
+            "messages": [synthetic_msg],
             "pm_decision": decision_str,
             "sender": "pm_decision_agent",
         }

tradingagents/graph/portfolio_graph.py

@@ -2,13 +2,18 @@
 
 from __future__ import annotations
 
+import os
 from typing import Any, List, Optional
 
 from tradingagents.default_config import DEFAULT_CONFIG
 from tradingagents.llm_clients import create_llm_client
+from tradingagents.memory.macro_memory import MacroMemory
+from tradingagents.memory.reflexion import ReflexionMemory
 from tradingagents.agents.portfolio import (
     create_holding_reviewer,
     create_pm_decision_agent,
+    create_macro_summary_agent,
+    create_micro_summary_agent,
 )
 from .portfolio_setup import PortfolioGraphSetup
 
@@ -50,12 +55,31 @@ class PortfolioGraph:
 
         portfolio_config = self._get_portfolio_config()
 
+        mongo_uri = self.config.get("mongo_uri") or os.environ.get("TRADINGAGENTS_MONGO_URI")
+        macro_mem = MacroMemory(mongo_uri=mongo_uri)
+        micro_mem = ReflexionMemory(
+            mongo_uri=mongo_uri,
+            collection_name="micro_reflexion",
+            fallback_path="reports/micro_reflexion.json",  # distinct from pipeline reflexion.json
+        )
+
         agents = {
             "review_holdings": create_holding_reviewer(mid_llm),
-            "pm_decision": create_pm_decision_agent(deep_llm, config=portfolio_config),
+            "macro_summary":   create_macro_summary_agent(mid_llm, macro_mem),
+            "micro_summary":   create_micro_summary_agent(mid_llm, micro_mem),
+            "pm_decision":     create_pm_decision_agent(
+                deep_llm, config=portfolio_config,
+                macro_memory=macro_mem, micro_memory=micro_mem,
+            ),
         }
 
-        setup = PortfolioGraphSetup(agents, repo=self._repo, config=portfolio_config)
+        setup = PortfolioGraphSetup(
+            agents,
+            repo=self._repo,
+            config=portfolio_config,
+            macro_memory=macro_mem,
+            micro_memory=micro_mem,
+        )
         self.graph = setup.setup_graph()
 
     def _get_portfolio_config(self) -> dict[str, Any]:
@@ -163,6 +187,10 @@ class PortfolioGraph:
             "risk_metrics": "",
             "holding_reviews": "",
             "prioritized_candidates": "",
+            "macro_brief": "",
+            "micro_brief": "",
+            "macro_memory_context": "",
+            "micro_memory_context": "",
             "pm_decision": "",
             "execution_result": "",
             "sender": "",

tradingagents/graph/portfolio_setup.py

@@ -1,12 +1,15 @@
 """Portfolio Manager workflow graph setup.
 
-Sequential workflow:
+Fan-out/fan-in workflow:
   START → load_portfolio → compute_risk → review_holdings
-        → prioritize_candidates → pm_decision → execute_trades → END
+        → prioritize_candidates → macro_summary (parallel)
+                                → micro_summary  (parallel)
+        → make_pm_decision → execute_trades → END
 
 Non-LLM nodes (load_portfolio, compute_risk, prioritize_candidates,
 execute_trades) receive ``repo`` and ``config`` via closure.
-LLM nodes (review_holdings, pm_decision) are created externally and passed in.
+LLM nodes (review_holdings, macro_summary, micro_summary, pm_decision)
+are created externally and passed in.
 """
 
 from __future__ import annotations
@@ -34,13 +37,16 @@ _EMPTY_PORTFOLIO_DICT = {
 
 
 class PortfolioGraphSetup:
-    """Builds the sequential Portfolio Manager workflow graph.
+    """Builds the Portfolio Manager workflow graph with parallel summary fan-out.
 
     Args:
-        agents: Dict with keys ``review_holdings`` and ``pm_decision``
+        agents: Dict with keys ``review_holdings``, ``macro_summary``,
-                mapping to LLM agent node functions.
+                ``micro_summary``, and ``pm_decision`` mapping to LLM agent
+                node functions.
         repo: PortfolioRepository instance (injected into closure nodes).
         config: Portfolio config dict.
+        macro_memory: MacroMemory instance forwarded to summary agents.
+        micro_memory: ReflexionMemory instance forwarded to summary agents.
     """
 
     def __init__(
@@ -48,10 +54,17 @@ class PortfolioGraphSetup:
         agents: dict[str, Any],
         repo=None,
         config: dict[str, Any] | None = None,
+        macro_memory=None,
+        micro_memory=None,
     ) -> None:
         self.agents = agents
         self._repo = repo
         self._config = config or {}
+        # Memory instances are already baked into the agent closures at the call site
+        # (portfolio_graph.py passes them to create_macro/micro_summary_agent).
+        # Stored here for future direct access by non-LLM closure nodes if needed.
+        self._macro_memory = macro_memory
+        self._micro_memory = micro_memory
 
     # ------------------------------------------------------------------
     # Node factories (non-LLM)
@@ -206,7 +219,13 @@ class PortfolioGraphSetup:
     # ------------------------------------------------------------------
 
     def setup_graph(self):
-        """Build and compile the sequential portfolio workflow graph.
+        """Build and compile the portfolio workflow graph with parallel summary fan-out.
+
+        Topology:
+            START → load_portfolio → compute_risk → review_holdings
+                  → prioritize_candidates → macro_summary (parallel)
+                                          → micro_summary  (parallel)
+                  → make_pm_decision → execute_trades → END
 
         Returns:
             A compiled LangGraph graph ready to invoke.
@@ -221,14 +240,25 @@ class PortfolioGraphSetup:
 
         # Register LLM nodes
         workflow.add_node("review_holdings", self.agents["review_holdings"])
+        workflow.add_node("macro_summary", self.agents["macro_summary"])
+        workflow.add_node("micro_summary", self.agents["micro_summary"])
         workflow.add_node("make_pm_decision", self.agents["pm_decision"])
 
-        # Sequential edges
+        # Sequential backbone
         workflow.add_edge(START, "load_portfolio")
         workflow.add_edge("load_portfolio", "compute_risk")
         workflow.add_edge("compute_risk", "review_holdings")
         workflow.add_edge("review_holdings", "prioritize_candidates")
-        workflow.add_edge("prioritize_candidates", "make_pm_decision")
+
+        # Fan-out: prioritize_candidates → both summary nodes (parallel)
+        workflow.add_edge("prioritize_candidates", "macro_summary")
+        workflow.add_edge("prioritize_candidates", "micro_summary")
+
+        # Fan-in: both summary nodes → make_pm_decision
+        workflow.add_edge("macro_summary", "make_pm_decision")
+        workflow.add_edge("micro_summary", "make_pm_decision")
+
+        # Tail
         workflow.add_edge("make_pm_decision", "execute_trades")
         workflow.add_edge("execute_trades", END)
 

tradingagents/memory/__init__.py

@@ -1 +1,6 @@
 """Agent memory subsystem for TradingAgents."""
+
+from tradingagents.memory.reflexion import ReflexionMemory
+from tradingagents.memory.macro_memory import MacroMemory
+
+__all__ = ["ReflexionMemory", "MacroMemory"]

tradingagents/memory/macro_memory.py

@@ -0,0 +1,281 @@
+"""Macro memory — learn from past regime-level market context.
+
+Stores macro regime states (VIX level, risk-on/off call, sector thesis, key
+themes) and later associates outcomes, enabling agents to *reflect* on
+regime accuracy and adjust forward-looking bias accordingly.
+
+Unlike ReflexionMemory (which is per-ticker), MacroMemory operates at the
+market-wide level. Each record captures the macro environment on a given date,
+independent of any single security.
+
+Backed by MongoDB when available; falls back to a local JSON file when not.
+
+Schema (``macro_memory`` collection)::
+
+    {
+        "regime_date":   str,          # ISO date "2026-03-26"
+        "vix_level":     float,        # e.g. 25.3
+        "macro_call":    str,          # "risk-on" | "risk-off" | "neutral" | "transition"
+        "sector_thesis": str,          # free-form regime summary
+        "key_themes":    list,         # list of top macro theme strings
+        "run_id":        str | None,
+        "outcome":       dict | None,  # filled later by record_outcome()
+        "created_at":    datetime,
+    }
+
+Usage::
+
+    from tradingagents.memory.macro_memory import MacroMemory
+
+    mem = MacroMemory("mongodb://localhost:27017")
+    mem.record_macro_state(
+        date="2026-03-26",
+        vix_level=25.3,
+        macro_call="risk-off",
+        sector_thesis="Energy under pressure, Fed hawkish",
+        key_themes=["rate hikes", "oil volatility"],
+    )
+    context = mem.build_macro_context(limit=3)
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+_COLLECTION = "macro_memory"
+
+_VALID_MACRO_CALLS = {"risk-on", "risk-off", "neutral", "transition"}
+
+
+class MacroMemory:
+    """MongoDB-backed macro regime memory.
+
+    Falls back to a local JSON file when MongoDB is unavailable, so the
+    feature always works (though with degraded query performance on the
+    local variant).
+    """
+
+    def __init__(
+        self,
+        mongo_uri: str | None = None,
+        db_name: str = "tradingagents",
+        fallback_path: str | Path = "reports/macro_memory.json",
+    ) -> None:
+        self._col = None
+        self._fallback_path = Path(fallback_path)
+
+        if mongo_uri:
+            try:
+                from pymongo import DESCENDING, MongoClient
+
+                client = MongoClient(mongo_uri)
+                db = client[db_name]
+                self._col = db[_COLLECTION]
+                self._col.create_index([("regime_date", DESCENDING)])
+                self._col.create_index("created_at")
+                logger.info("MacroMemory using MongoDB (db=%s)", db_name)
+            except Exception:
+                logger.warning(
+                    "MacroMemory: MongoDB unavailable — using local file",
+                    exc_info=True,
+                )
+
+    # ------------------------------------------------------------------
+    # Record macro state
+    # ------------------------------------------------------------------
+
+    def record_macro_state(
+        self,
+        date: str,
+        vix_level: float,
+        macro_call: str,
+        sector_thesis: str,
+        key_themes: list[str],
+        run_id: str | None = None,
+    ) -> None:
+        """Store a macro regime state for later reflection.
+
+        Args:
+            date:          ISO date string, e.g. "2026-03-26".
+            vix_level:     VIX index level at the time of the call.
+            macro_call:    Regime classification: "risk-on", "risk-off",
+                           "neutral", or "transition".
+            sector_thesis: Free-form summary of the prevailing sector view.
+            key_themes:    Top macro themes driving the regime call.
+            run_id:        Optional run identifier for traceability.
+        """
+        normalized_call = macro_call.lower()
+        if normalized_call not in _VALID_MACRO_CALLS:
+            logger.warning(
+                "MacroMemory: unexpected macro_call %r (expected one of %s)",
+                macro_call,
+                _VALID_MACRO_CALLS,
+            )
+
+        doc: dict[str, Any] = {
+            "regime_date": date,
+            "vix_level": float(vix_level),
+            "macro_call": normalized_call,
+            "sector_thesis": sector_thesis,
+            "key_themes": list(key_themes),
+            "run_id": run_id,
+            "outcome": None,
+            "created_at": datetime.now(timezone.utc),
+        }
+
+        if self._col is not None:
+            self._col.insert_one(doc)
+        else:
+            # Local JSON fallback uses ISO string (JSON has no datetime type)
+            doc["created_at"] = doc["created_at"].isoformat()
+            self._append_local(doc)
+
+    # ------------------------------------------------------------------
+    # Record outcome (feedback loop)
+    # ------------------------------------------------------------------
+
+    def record_outcome(self, date: str, outcome: dict[str, Any]) -> bool:
+        """Attach outcome to the most recent macro state for a given date.
+
+        Args:
+            date:    ISO date string matching the original ``regime_date``.
+            outcome: Dict with evaluation data, e.g.::
+
+                {
+                    "evaluation_date": "2026-04-26",
+                    "vix_at_evaluation": 18.2,
+                    "regime_confirmed": True,
+                    "notes": "Risk-off call was correct; market sold off",
+                }
+
+        Returns:
+            True if a matching state was found and updated.
+        """
+        if self._col is not None:
+            from pymongo import DESCENDING
+
+            doc = self._col.find_one_and_update(
+                {"regime_date": date, "outcome": None},
+                {"$set": {"outcome": outcome}},
+                sort=[("created_at", DESCENDING)],
+            )
+            return doc is not None
+        else:
+            return self._update_local_outcome(date, outcome)
+
+    # ------------------------------------------------------------------
+    # Query
+    # ------------------------------------------------------------------
+
+    def get_recent(self, limit: int = 3) -> list[dict[str, Any]]:
+        """Return most recent macro states, newest first.
+
+        Args:
+            limit: Maximum number of results.
+        """
+        if self._col is not None:
+            from pymongo import DESCENDING
+
+            cursor = self._col.find(
+                {},
+                {"_id": 0},
+            ).sort("regime_date", DESCENDING).limit(limit)
+            return list(cursor)
+        else:
+            return self._load_recent_local(limit)
+
+    def build_macro_context(self, limit: int = 3) -> str:
+        """Build a human-readable context string from recent macro states.
+
+        Suitable for injection into agent prompts. Returns a multi-line string
+        summarising recent regime calls and outcomes.
+
+        Format example::
+
+            - [2026-03-20] risk-off (VIX: 25.3)
+              Thesis: Energy sector under pressure, Fed hawkish
+              Themes: ['rate hikes', 'oil volatility']
+              Outcome: pending
+
+        Args:
+            limit: How many past states to include.
+
+        Returns:
+            Multi-line string summarising recent macro regime states.
+        """
+        recent = self.get_recent(limit=limit)
+        if not recent:
+            return "No prior macro regime states recorded."
+
+        lines: list[str] = []
+        for rec in recent:
+            dt = rec.get("regime_date", "?")
+            call = rec.get("macro_call", "?")
+            vix = rec.get("vix_level", "?")
+            thesis = rec.get("sector_thesis", "")[:300]
+            themes = rec.get("key_themes", [])
+
+            outcome = rec.get("outcome")
+            if outcome:
+                confirmed = outcome.get("regime_confirmed", "?")
+                notes = outcome.get("notes", "")
+                outcome_str = f"  Outcome: confirmed={confirmed} — {notes}" if notes else f"  Outcome: confirmed={confirmed}"
+            else:
+                outcome_str = "  Outcome: pending"
+
+            lines.append(
+                f"- [{dt}] {call} (VIX: {vix})\n"
+                f"  Thesis: {thesis}\n"
+                f"  Themes: {themes}\n"
+                f"{outcome_str}"
+            )
+        return "\n".join(lines)
+
+    # ------------------------------------------------------------------
+    # Local JSON fallback
+    # ------------------------------------------------------------------
+
+    def _load_all_local(self) -> list[dict[str, Any]]:
+        """Load all records from the local JSON file."""
+        if not self._fallback_path.exists():
+            return []
+        try:
+            return json.loads(self._fallback_path.read_text(encoding="utf-8"))
+        except (json.JSONDecodeError, OSError):
+            return []
+
+    def _save_all_local(self, records: list[dict[str, Any]]) -> None:
+        """Overwrite the local JSON file with all records."""
+        self._fallback_path.parent.mkdir(parents=True, exist_ok=True)
+        self._fallback_path.write_text(
+            json.dumps(records, indent=2), encoding="utf-8"
+        )
+
+    def _append_local(self, doc: dict[str, Any]) -> None:
+        """Append a single record to the local file."""
+        records = self._load_all_local()
+        records.append(doc)
+        self._save_all_local(records)
+
+    def _load_recent_local(self, limit: int) -> list[dict[str, Any]]:
+        """Load and sort all records by regime_date descending from the local file."""
+        records = self._load_all_local()
+        records.sort(key=lambda r: r.get("regime_date", ""), reverse=True)
+        return records[:limit]
+
+    def _update_local_outcome(self, date: str, outcome: dict[str, Any]) -> bool:
+        """Update the most recent matching macro state in the local file."""
+        records = self._load_all_local()
+        # Iterate newest first (reversed insertion order is a proxy)
+        for rec in reversed(records):
+            if rec.get("regime_date") == date and rec.get("outcome") is None:
+                rec["outcome"] = outcome
+                self._save_all_local(records)
+                return True
+        return False

tradingagents/memory/reflexion.py

@@ -56,6 +56,7 @@ class ReflexionMemory:
         mongo_uri: str | None = None,
         db_name: str = "tradingagents",
         fallback_path: str | Path = "reports/reflexion.json",
+        collection_name: str = "reflexion",
     ) -> None:
         self._col = None
         self._fallback_path = Path(fallback_path)
@@ -66,7 +67,7 @@ class ReflexionMemory:
 
                 client = MongoClient(mongo_uri)
                 db = client[db_name]
-                self._col = db[_COLLECTION]
+                self._col = db[collection_name]
                 self._col.create_index(
                     [("ticker", 1), ("decision_date", DESCENDING)]
                 )
@@ -184,7 +185,7 @@ class ReflexionMemory:
             from pymongo import DESCENDING
 
             cursor = self._col.find(
-                {"ticker": ticker.upper()},
+                {"ticker": ticker.upper()},  # Hard metadata filter — prevents cross-ticker contamination
                 {"_id": 0},
             ).sort("decision_date", DESCENDING).limit(limit)
             return list(cursor)
@@ -260,7 +261,7 @@ class ReflexionMemory:
     def _load_local(self, ticker: str, limit: int) -> list[dict[str, Any]]:
         """Load and filter records for a ticker from the local file."""
         records = self._load_all_local()
-        filtered = [r for r in records if r.get("ticker") == ticker]
+        filtered = [r for r in records if r.get("ticker") == ticker]  # Hard metadata filter — local fallback
         filtered.sort(key=lambda r: r.get("decision_date", ""), reverse=True)
         return filtered[:limit]
 

tradingagents/portfolio/dual_report_store.py

@@ -67,6 +67,11 @@ class DualReportStore:
         """The flow identifier set on this store, if any."""
         return self._local.flow_id
 
+    @property
+    def flow_id(self) -> str | None:
+        """The flow identifier set on this store, if any."""
+        return self._local.flow_id
+
     @property
     def run_id(self) -> str | None:
         """The run/flow identifier (flow_id takes precedence)."""

tradingagents/portfolio/mongo_report_store.py

@@ -96,6 +96,11 @@ class MongoReportStore:
         """The flow identifier set on this store, if any."""
         return self._flow_id
 
+    @property
+    def flow_id(self) -> str | None:
+        """The flow identifier set on this store, if any."""
+        return self._flow_id
+
     @property
     def run_id(self) -> str | None:
         """The run/flow identifier (flow_id takes precedence for backward compat)."""

tradingagents/portfolio/portfolio_states.py

@@ -35,6 +35,15 @@ class PortfolioManagerState(MessagesState):
     risk_metrics: Annotated[str, _last_value]
     holding_reviews: Annotated[str, _last_value]
     prioritized_candidates: Annotated[str, _last_value]
+
+    # Summary briefs (written by parallel summary agents)
+    macro_brief: Annotated[str, _last_value]
+    micro_brief: Annotated[str, _last_value]
+
+    # Pre-fetched memory context strings
+    macro_memory_context: Annotated[str, _last_value]
+    micro_memory_context: Annotated[str, _last_value]
+
     pm_decision: Annotated[str, _last_value]
     execution_result: Annotated[str, _last_value]