feat: normalize durable fact extraction (#1012)

Closes #1012

- add structured session fact extraction with provenance, temporal metadata,
  canonical normalization, and contradiction grouping
- persist structured metadata into holographic memory auto-extraction with
  canonical-key dedupe across repeated ingests
- add fixture-backed transcript tests plus extraction quality evaluation

agent/session_compactor.py

@@ -1,197 +1,546 @@
-"""Session compaction with fact extraction.
+"""Session compaction with structured fact extraction.
 
-Before compressing conversation context, extracts durable facts
-(user preferences, corrections, project details) and saves them
-to the fact store so they survive compression.
-
-Usage:
-    from agent.session_compactor import extract_and_save_facts
-    facts = extract_and_save_facts(messages)
+Before compressing conversation context, extract durable facts with enough
+structure to survive retrieval: source/provenance, temporal anchors,
+normalized canonical keys, and contradiction groups.
 """
 
 from __future__ import annotations
 
-import json
 import logging
 import re
 import time
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional, Tuple
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Tuple
 
 logger = logging.getLogger(__name__)
 
 
+_DEPLOY_METHOD_RE = re.compile(r"\bdeploy(?:ing)?\s+(?:via|through|with)\s+([A-Za-z0-9_./+-]+)", re.IGNORECASE)
+_WATCHDOG_CAP_RE = re.compile(
+    r"\b(?:the\s+)?([A-Za-z0-9_-]+(?:\s+watchdog)?)\s+(?:caps|limits)\s+dispatches(?:\s+per\s+cycle)?\s+to\s+([0-9]+)",
+    re.IGNORECASE,
+)
+_PROVIDER_RE = re.compile(
+    r"\bprovider\s+(?:is|should\s+stay|should\s+be|needs\s+to\s+be)\s+([A-Za-z0-9._/-]+)",
+    re.IGNORECASE,
+)
+_MODEL_RE = re.compile(
+    r"\bmodel\s+(?:is|should\s+stay|should\s+be|needs\s+to\s+be)\s+([A-Za-z0-9._:/-]+)",
+    re.IGNORECASE,
+)
+_PORT_RE = re.compile(r"\bport\s+(?:is|should\s+be)\s+([0-9]+)", re.IGNORECASE)
+_PROJECT_USES_RE = re.compile(r"\b(?:the\s+)?project\s+(?:uses|needs|requires)\s+(.+?)(?:[.!?]|$)", re.IGNORECASE)
+_PREFERENCE_RE = re.compile(r"\bI\s+(?:prefer|like|want|need)\s+(.+?)(?:[.!?]|$)", re.IGNORECASE)
+_CONSTRAINT_RE = re.compile(r"\b(?:do\s+not|don't)\s+(?:ever\s+|again\s+)?(.+?)(?:[.!?]|$)", re.IGNORECASE)
+_DECISION_RE = re.compile(r"\b(?:we|the\s+team)\s+(?:decided|agreed|chose)\s+(?:to\s+)?(.+?)(?:[.!?]|$)", re.IGNORECASE)
+
+
 @dataclass
 class ExtractedFact:
-    """A fact extracted from conversation."""
-    category: str       # "user_pref", "correction", "project", "tool_quirk", "general"
-    entity: str         # what the fact is about
-    content: str        # the fact itself
-    confidence: float   # 0.0-1.0
-    source_turn: int    # which message turn it came from
+    """A durable fact extracted from conversation."""
+
+    category: str
+    entity: str
+    content: str
+    confidence: float
+    source_turn: int
     timestamp: float = 0.0
+    source_role: str = "user"
+    source_text: str = ""
+    normalized_content: str = ""
+    canonical_key: str = ""
+    relation: str = "general"
+    contradiction_group: str = ""
+    status: str = "active"
+    provenance: str = ""
+    observed_at: str = ""
+    evidence: List[Dict[str, Any]] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
 
-
-# Patterns that indicate user preferences
-_PREFERENCE_PATTERNS = [
-    (r"(?:I|we) (?:prefer|like|want|need) (.+?)(?:\.|$)", "preference"),
-    (r"(?:always|never) (?:use|do|run|deploy) (.+?)(?:\.|$)", "preference"),
-    (r"(?:my|our) (?:default|preferred|usual) (.+?) (?:is|are) (.+?)(?:\.|$)", "preference"),
-    (r"(?:make sure|ensure|remember) (?:to|that) (.+?)(?:\.|$)", "instruction"),
-    (r"(?:don'?t|do not) (?:ever|ever again) (.+?)(?:\.|$)", "constraint"),
-]
-
-# Patterns that indicate corrections
-_CORRECTION_PATTERNS = [
-    (r"(?:actually|no[, ]|wait[, ]|correction[: ]|sorry[, ]) (.+)", "correction"),
-    (r"(?:I meant|what I meant was|the correct) (.+?)(?:\.|$)", "correction"),
-    (r"(?:it'?s|its) (?:not|shouldn'?t be|wrong) (.+?)(?:\.|$)", "correction"),
-]
-
-# Patterns that indicate project/tool facts
-_PROJECT_PATTERNS = [
-    (r"(?:the |our )?(?:project|repo|codebase|code) (?:is|uses|needs|requires) (.+?)(?:\.|$)", "project"),
-    (r"(?:deploy|push|commit) (?:to|on) (.+?)(?:\.|$)", "project"),
-    (r"(?:this|that|the) (?:server|host|machine|VPS) (?:is|runs|has) (.+?)(?:\.|$)", "infrastructure"),
-    (r"(?:model|provider|engine) (?:is|should be|needs to be) (.+?)(?:\.|$)", "config"),
-]
+    def __post_init__(self) -> None:
+        if not self.timestamp:
+            self.timestamp = time.time()
+        if not self.observed_at:
+            self.observed_at = _iso_from_timestamp(self.timestamp)
+        if not self.normalized_content:
+            self.normalized_content = _normalize_value(self.content)
+        if not self.provenance:
+            self.provenance = f"conversation:{self.source_role}:{self.source_turn}"
+        if not self.canonical_key:
+            self.canonical_key = _canonical_key(self.entity, self.relation, self.normalized_content)
+        if not self.evidence:
+            self.evidence = [
+                {
+                    "source_role": self.source_role,
+                    "source_turn": self.source_turn,
+                    "source_text": self.source_text or self.content,
+                    "observed_at": self.observed_at,
+                    "provenance": self.provenance,
+                }
+            ]
+        self.metadata = dict(self.metadata or {})
+        self.metadata.setdefault("entity", self.entity)
+        self.metadata.setdefault("relation", self.relation)
+        self.metadata.setdefault("value", self.content)
+        self.metadata.setdefault("normalized_value", self.normalized_content)
+        self.metadata.setdefault("provenance", [self.provenance])
+        self.metadata.setdefault("evidence", list(self.evidence))
+        self.metadata.setdefault("observation_count", len(self.evidence))
+        self.metadata.setdefault("duplicate_count", max(0, self.metadata["observation_count"] - 1))
+        if self.contradiction_group:
+            self.metadata.setdefault("contradiction_group", self.contradiction_group)
+        self.metadata.setdefault("status", self.status)
 
 
 def extract_facts_from_messages(messages: List[Dict[str, Any]]) -> List[ExtractedFact]:
     """Extract durable facts from conversation messages.
 
-    Scans user messages for preferences, corrections, project facts,
-    and infrastructure details that should survive compression.
+    Scans conversation turns for preferences, decisions, corrections, and
+    operational state. Raw candidates are normalized into canonical facts so
+    near-duplicates merge and contradictions remain inspectable.
     """
-    facts = []
-    seen_contents = set()
 
+    raw_candidates: list[ExtractedFact] = []
     for turn_idx, msg in enumerate(messages):
         role = msg.get("role", "")
         content = msg.get("content", "")
-
-        # Only scan user messages and assistant responses with corrections
-        if role not in ("user", "assistant"):
+        if role not in {"user", "assistant"}:
             continue
-        if not content or not isinstance(content, str):
-            continue
-        if len(content) < 10:
-            continue
-
-        # Skip tool results and system messages
         if role == "assistant" and msg.get("tool_calls"):
             continue
+        if not isinstance(content, str) or len(content.strip()) < 10:
+            continue
 
-        extracted = _extract_from_text(content, turn_idx, role)
+        timestamp, observed_at = _message_time(msg)
+        raw_candidates.extend(
+            _extract_from_text(
+                content.strip(),
+                turn_idx=turn_idx,
+                role=role,
+                timestamp=timestamp,
+                observed_at=observed_at,
+            )
+        )
 
-        # Deduplicate by content
-        for fact in extracted:
-            key = f"{fact.category}:{fact.content[:100]}"
-            if key not in seen_contents:
-                seen_contents.add(key)
-                facts.append(fact)
+    return _normalize_candidates(raw_candidates)
+
+
+def evaluate_extraction_quality(messages: List[Dict[str, Any]]) -> Dict[str, Any]:
+    """Return before/after metrics for raw vs normalized extraction quality."""
+
+    raw_candidates: list[ExtractedFact] = []
+    for turn_idx, msg in enumerate(messages):
+        role = msg.get("role", "")
+        content = msg.get("content", "")
+        if role not in {"user", "assistant"}:
+            continue
+        if role == "assistant" and msg.get("tool_calls"):
+            continue
+        if not isinstance(content, str) or len(content.strip()) < 10:
+            continue
+        timestamp, observed_at = _message_time(msg)
+        raw_candidates.extend(
+            _extract_from_text(
+                content.strip(),
+                turn_idx=turn_idx,
+                role=role,
+                timestamp=timestamp,
+                observed_at=observed_at,
+            )
+        )
+
+    normalized = _normalize_candidates(raw_candidates)
+    raw_count = len(raw_candidates)
+    normalized_count = len(normalized)
+    contradiction_groups = {
+        fact.contradiction_group
+        for fact in normalized
+        if fact.status == "contradiction" and fact.contradiction_group
+    }
+    duplicate_count = max(0, raw_count - normalized_count)
+    noise_reduction = (duplicate_count / raw_count) if raw_count else 0.0
+
+    return {
+        "raw_candidates": raw_count,
+        "normalized_facts": normalized_count,
+        "duplicates_merged": duplicate_count,
+        "contradiction_groups": len(contradiction_groups),
+        "noise_reduction": round(noise_reduction, 3),
+    }
+
+
+def _extract_from_text(
+    text: str,
+    *,
+    turn_idx: int,
+    role: str,
+    timestamp: float,
+    observed_at: str,
+) -> List[ExtractedFact]:
+    """Extract raw fact candidates from a single text block."""
+
+    facts: list[ExtractedFact] = []
+    if role != "user":
+        return facts
+
+    deploy_match = _DEPLOY_METHOD_RE.search(text)
+    if deploy_match:
+        method = deploy_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.decision",
+                entity="project",
+                relation="workflow.deploy_method",
+                value=method,
+                content=f"Deploy via {method}",
+                confidence=0.88,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    watchdog_match = _WATCHDOG_CAP_RE.search(text)
+    if watchdog_match:
+        watchdog = watchdog_match.group(1).strip()
+        cap = watchdog_match.group(2).strip()
+        facts.append(
+            _build_fact(
+                category="project.operational",
+                entity=_normalize_entity(watchdog),
+                relation="fleet.dispatch_cap",
+                value=cap,
+                content=f"{watchdog} caps dispatches per cycle to {cap}",
+                confidence=0.92,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    provider_match = _PROVIDER_RE.search(text)
+    if provider_match:
+        provider = provider_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.config",
+                entity="project",
+                relation="config.provider",
+                value=provider,
+                content=f"Provider should stay {provider}",
+                confidence=0.91,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    model_match = _MODEL_RE.search(text)
+    if model_match:
+        model = model_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.config",
+                entity="project",
+                relation="config.model",
+                value=model,
+                content=f"Model should stay {model}",
+                confidence=0.9,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    port_match = _PORT_RE.search(text)
+    if port_match:
+        port = port_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.config",
+                entity="project",
+                relation="config.port",
+                value=port,
+                content=f"Port is {port}",
+                confidence=0.9,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    project_match = _PROJECT_USES_RE.search(text)
+    if project_match:
+        value = project_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="project.stack",
+                entity="project",
+                relation="project.stack",
+                value=value,
+                content=f"Project uses {value}",
+                confidence=0.74,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
+
+    preference_match = _PREFERENCE_RE.search(text)
+    if preference_match:
+        value = preference_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="user_pref.preference",
+                entity="user",
+                relation="user.preference",
+                value=value,
+                content=value,
+                confidence=0.72,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
+
+    constraint_match = _CONSTRAINT_RE.search(text)
+    if constraint_match:
+        value = constraint_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="user_pref.constraint",
+                entity="user",
+                relation="user.constraint",
+                value=value,
+                content=f"Do not {value}",
+                confidence=0.82,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
+
+    decision_match = _DECISION_RE.search(text)
+    if decision_match:
+        value = decision_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="project.decision",
+                entity="project",
+                relation="project.decision",
+                value=value,
+                content=f"Decision: {value}",
+                confidence=0.79,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
 
     return facts
 
 
-def _extract_from_text(text: str, turn_idx: int, role: str) -> List[ExtractedFact]:
-    """Extract facts from a single text block."""
-    facts = []
-    timestamp = time.time()
+def _build_fact(
+    *,
+    category: str,
+    entity: str,
+    relation: str,
+    value: str,
+    content: str,
+    confidence: float,
+    source_turn: int,
+    source_role: str,
+    source_text: str,
+    timestamp: float,
+    observed_at: str,
+    unique_slot: bool,
+) -> ExtractedFact:
+    normalized_value = _normalize_value(value.rstrip(".!?"))
+    value = value.rstrip(".!?")
+    content = content.rstrip(".!?")
+    provenance = f"conversation:{source_role}:{source_turn}"
+    contradiction_group = relation if unique_slot else ""
+    evidence = [
+        {
+            "source_role": source_role,
+            "source_turn": source_turn,
+            "source_text": source_text,
+            "observed_at": observed_at,
+            "provenance": provenance,
+        }
+    ]
+    metadata = {
+        "entity": entity,
+        "relation": relation,
+        "value": value,
+        "normalized_value": normalized_value,
+        "provenance": [provenance],
+        "evidence": list(evidence),
+        "observation_count": 1,
+        "duplicate_count": 0,
+        "status": "active",
+    }
+    if contradiction_group:
+        metadata["contradiction_group"] = contradiction_group
+    return ExtractedFact(
+        category=category,
+        entity=entity,
+        content=content,
+        confidence=confidence,
+        source_turn=source_turn,
+        timestamp=timestamp,
+        source_role=source_role,
+        source_text=source_text,
+        normalized_content=normalized_value,
+        canonical_key=_canonical_key(entity, relation, normalized_value),
+        relation=relation,
+        contradiction_group=contradiction_group,
+        status="active",
+        provenance=provenance,
+        observed_at=observed_at,
+        evidence=evidence,
+        metadata=metadata,
+    )
 
-    # Clean text for pattern matching
-    clean = text.strip()
 
-    # User preference patterns (from user messages)
-    if role == "user":
-        for pattern, subcategory in _PREFERENCE_PATTERNS:
-            for match in re.finditer(pattern, clean, re.IGNORECASE):
-                content = match.group(1).strip() if match.lastindex else match.group(0).strip()
-                if len(content) > 5:
-                    facts.append(ExtractedFact(
-                        category=f"user_pref.{subcategory}",
-                        entity="user",
-                        content=content[:200],
-                        confidence=0.7,
-                        source_turn=turn_idx,
-                        timestamp=timestamp,
-                    ))
+def _normalize_candidates(candidates: List[ExtractedFact]) -> List[ExtractedFact]:
+    """Merge duplicates and mark contradictions while preserving evidence."""
 
-    # Correction patterns (from user messages)
-    if role == "user":
-        for pattern, subcategory in _CORRECTION_PATTERNS:
-            for match in re.finditer(pattern, clean, re.IGNORECASE):
-                content = match.group(1).strip() if match.lastindex else match.group(0).strip()
-                if len(content) > 5:
-                    facts.append(ExtractedFact(
-                        category=f"correction.{subcategory}",
-                        entity="user",
-                        content=content[:200],
-                        confidence=0.8,
-                        source_turn=turn_idx,
-                        timestamp=timestamp,
-                    ))
+    by_key: dict[str, ExtractedFact] = {}
+    contradiction_groups: dict[str, list[ExtractedFact]] = {}
 
-    # Project/infrastructure patterns (from both user and assistant)
-    for pattern, subcategory in _PROJECT_PATTERNS:
-        for match in re.finditer(pattern, clean, re.IGNORECASE):
-            content = match.group(1).strip() if match.lastindex else match.group(0).strip()
-            if len(content) > 5:
-                facts.append(ExtractedFact(
-                    category=f"project.{subcategory}",
-                    entity=subcategory,
-                    content=content[:200],
-                    confidence=0.6,
-                    source_turn=turn_idx,
-                    timestamp=timestamp,
-                ))
+    for candidate in candidates:
+        existing = by_key.get(candidate.canonical_key)
+        if existing is not None:
+            by_key[candidate.canonical_key] = _merge_fact(existing, candidate)
+            continue
 
-    return facts
+        by_key[candidate.canonical_key] = candidate
+        if candidate.contradiction_group:
+            contradiction_groups.setdefault(candidate.contradiction_group, []).append(candidate)
+
+    for group, facts in contradiction_groups.items():
+        canonical_keys = {fact.canonical_key for fact in facts}
+        if len(canonical_keys) <= 1:
+            continue
+        for fact in facts:
+            fact.status = "contradiction"
+            fact.metadata["status"] = "contradiction"
+            fact.metadata["contradiction_group"] = group
+            fact.metadata["contradiction_keys"] = sorted(canonical_keys - {fact.canonical_key})
+
+    return sorted(by_key.values(), key=lambda fact: (fact.source_turn, fact.timestamp, fact.canonical_key))
+
+
+def _merge_fact(existing: ExtractedFact, incoming: ExtractedFact) -> ExtractedFact:
+    existing.confidence = max(existing.confidence, incoming.confidence)
+    existing.timestamp = min(existing.timestamp, incoming.timestamp)
+    existing.source_turn = min(existing.source_turn, incoming.source_turn)
+    if not existing.observed_at or (incoming.observed_at and incoming.observed_at < existing.observed_at):
+        existing.observed_at = incoming.observed_at
+    existing.provenance = min(existing.provenance, incoming.provenance)
+
+    provenance = _ordered_unique(existing.metadata.get("provenance", []), incoming.metadata.get("provenance", []))
+    evidence = _merge_evidence(existing.metadata.get("evidence", []), incoming.metadata.get("evidence", []))
+    observation_count = int(existing.metadata.get("observation_count", len(existing.evidence) or 1))
+    observation_count += int(incoming.metadata.get("observation_count", len(incoming.evidence) or 1))
+
+    existing.evidence = evidence
+    existing.metadata["provenance"] = provenance
+    existing.metadata["evidence"] = evidence
+    existing.metadata["observation_count"] = observation_count
+    existing.metadata["duplicate_count"] = max(0, observation_count - 1)
+    existing.metadata["status"] = existing.status
+    return existing
 
 
 def save_facts_to_store(facts: List[ExtractedFact], fact_store_fn=None) -> int:
     """Save extracted facts to the fact store.
 
-    Args:
-        facts: List of extracted facts.
-        fact_store_fn: Optional callable(category, entity, content, trust).
-            If None, uses the holographic fact store if available.
-
-    Returns:
-        Number of facts saved.
+    If a callback is supplied, prefer the structured signature but fall back to
+    the legacy four-argument callback for compatibility.
     """
-    saved = 0
 
-    if fact_store_fn:
-        for fact in facts:
+    saved = 0
+    for fact in facts:
+        payload = {
+            "category": _store_category(fact.category),
+            "entity": fact.entity,
+            "content": fact.content,
+            "trust": fact.confidence,
+            "metadata": dict(fact.metadata),
+            "canonical_key": fact.canonical_key,
+            "observed_at": fact.observed_at,
+            "source_role": fact.source_role,
+            "source_turn": fact.source_turn,
+            "contradiction_group": fact.contradiction_group,
+            "status": fact.status,
+            "relation": fact.relation,
+        }
+
+        if fact_store_fn:
             try:
-                fact_store_fn(
-                    category=fact.category,
-                    entity=fact.entity,
-                    content=fact.content,
-                    trust=fact.confidence,
-                )
+                fact_store_fn(**payload)
                 saved += 1
-            except Exception as e:
-                logger.debug("Failed to save fact: %s", e)
-    else:
-        # Try holographic fact store
+                continue
+            except TypeError:
+                try:
+                    fact_store_fn(payload["category"], payload["entity"], payload["content"], payload["trust"])
+                    saved += 1
+                    continue
+                except Exception as exc:
+                    logger.debug("Failed to save fact via callback: %s", exc)
+                    continue
+            except Exception as exc:
+                logger.debug("Failed to save fact via callback: %s", exc)
+                continue
+
         try:
             from fact_store import fact_store as _fs
-            for fact in facts:
-                try:
-                    _fs(
-                        action="add",
-                        content=fact.content,
-                        category=fact.category,
-                        tags=fact.entity,
-                        trust_delta=fact.confidence - 0.5,
-                    )
-                    saved += 1
-                except Exception as e:
-                    logger.debug("Failed to save fact via fact_store: %s", e)
+
+            tags = ",".join(filter(None, [fact.entity, fact.relation, fact.status]))
+            _fs(
+                action="add",
+                content=fact.content,
+                category=_store_category(fact.category),
+                tags=tags,
+                trust_delta=fact.confidence - 0.5,
+            )
+            saved += 1
         except ImportError:
             logger.debug("fact_store not available — facts not persisted")
+            break
+        except Exception as exc:
+            logger.debug("Failed to save fact via fact_store: %s", exc)
 
     return saved
 
@@ -204,9 +553,10 @@ def extract_and_save_facts(
 
     Returns (extracted_facts, saved_count).
     """
+
     facts = extract_facts_from_messages(messages)
     if facts:
-        logger.info("Extracted %d facts from conversation", len(facts))
+        logger.info("Extracted %d normalized facts from conversation", len(facts))
         saved = save_facts_to_store(facts, fact_store_fn)
         logger.info("Saved %d/%d facts to store", saved, len(facts))
     else:
@@ -216,16 +566,105 @@ def extract_and_save_facts(
 
 def format_facts_summary(facts: List[ExtractedFact]) -> str:
     """Format extracted facts as a readable summary."""
+
     if not facts:
         return "No facts extracted."
 
-    by_category = {}
-    for f in facts:
-        by_category.setdefault(f.category, []).append(f)
+    by_category: dict[str, list[ExtractedFact]] = {}
+    for fact in facts:
+        by_category.setdefault(fact.category, []).append(fact)
 
     lines = [f"Extracted {len(facts)} facts:", ""]
-    for cat, cat_facts in sorted(by_category.items()):
-        lines.append(f"  {cat}:")
-        for f in cat_facts:
-            lines.append(f"    - {f.content[:80]}")
+    for category, category_facts in sorted(by_category.items()):
+        lines.append(f"  {category}:")
+        for fact in category_facts:
+            suffix = f" [{fact.status}]" if fact.status != "active" else ""
+            lines.append(f"    - {fact.content[:80]}{suffix}")
     return "\n".join(lines)
+
+
+def _store_category(category: str) -> str:
+    if category.startswith("user_pref"):
+        return "user_pref"
+    if category.startswith("project"):
+        return "project"
+    if category.startswith("tool"):
+        return "tool"
+    return "general"
+
+
+def _message_time(msg: Dict[str, Any]) -> Tuple[float, str]:
+    for key in ("created_at", "timestamp", "time"):
+        value = msg.get(key)
+        if value is None:
+            continue
+        if isinstance(value, (int, float)):
+            ts = float(value)
+            return ts, _iso_from_timestamp(ts)
+        if isinstance(value, str):
+            parsed = _parse_time_string(value)
+            if parsed is not None:
+                return parsed, _iso_from_timestamp(parsed) if "T" not in value else value.replace("+00:00", "Z")
+            return time.time(), value
+    now = time.time()
+    return now, _iso_from_timestamp(now)
+
+
+def _parse_time_string(value: str) -> float | None:
+    text = value.strip()
+    if not text:
+        return None
+    try:
+        return float(text)
+    except ValueError:
+        pass
+    try:
+        normalized = text[:-1] + "+00:00" if text.endswith("Z") else text
+        return datetime.fromisoformat(normalized).timestamp()
+    except ValueError:
+        return None
+
+
+def _iso_from_timestamp(value: float) -> str:
+    return datetime.fromtimestamp(value, tz=timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def _normalize_value(value: str) -> str:
+    normalized = re.sub(r"[^a-z0-9]+", " ", value.lower())
+    normalized = re.sub(r"\s+", " ", normalized).strip()
+    return normalized
+
+
+def _normalize_entity(value: str) -> str:
+    return _normalize_value(value).replace(" ", "_") or "entity"
+
+
+def _canonical_key(entity: str, relation: str, normalized_value: str) -> str:
+    return f"{entity}|{relation}|{normalized_value}"
+
+
+def _ordered_unique(*groups: List[str]) -> List[str]:
+    seen: set[str] = set()
+    ordered: list[str] = []
+    for group in groups:
+        for item in group:
+            if item and item not in seen:
+                seen.add(item)
+                ordered.append(item)
+    return ordered
+
+
+def _merge_evidence(existing: List[Dict[str, Any]], incoming: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    seen: set[tuple[str, str, str]] = set()
+    merged: list[dict[str, Any]] = []
+    for item in list(existing) + list(incoming):
+        key = (
+            str(item.get("provenance", "")),
+            str(item.get("observed_at", "")),
+            str(item.get("source_text", "")),
+        )
+        if key in seen:
+            continue
+        seen.add(key)
+        merged.append(dict(item))
+    return merged

optional-skills/dogfood/adversarial-ux-test/SKILL.md

@@ -1,190 +0,0 @@
----
-name: adversarial-ux-test
-description: Roleplay the most difficult, tech-resistant user for your product. Browse the app as that persona, find every UX pain point, then filter complaints through a pragmatism layer to separate real problems from noise. Creates actionable tickets from genuine issues only.
-version: 1.0.0
-author: Omni @ Comelse
-license: MIT
-metadata:
-  hermes:
-    tags: [qa, ux, testing, adversarial, dogfood, personas, user-testing]
-    related_skills: [dogfood]
----
-
-# Adversarial UX Test
-
-Roleplay the worst-case user for your product — the person who hates technology, doesn't want your software, and will find every reason to complain. Then filter their feedback through a pragmatism layer to separate real UX problems from "I hate computers" noise.
-
-Think of it as an automated "mom test" — but angry.
-
-## Why This Works
-
-Most QA finds bugs. This finds **friction**. A technically correct app can still be unusable for real humans. The adversarial persona catches:
-- Confusing terminology that makes sense to developers but not users
-- Too many steps to accomplish basic tasks
-- Missing onboarding or "aha moments"
-- Accessibility issues (font size, contrast, click targets)
-- Cold-start problems (empty states, no demo content)
-- Paywall/signup friction that kills conversion
-
-The **pragmatism filter** (Phase 3) is what makes this useful instead of just entertaining. Without it, you'd add a "print this page" button to every screen because Grandpa can't figure out PDFs.
-
-## How to Use
-
-Tell the agent:
-```
-"Run an adversarial UX test on [URL]"
-"Be a grumpy [persona type] and test [app name]"
-"Do an asshole user test on my staging site"
-```
-
-You can provide a persona or let the agent generate one based on your product's target audience.
-
-## Step 1: Define the Persona
-
-If no persona is provided, generate one by answering:
-
-1. **Who is the HARDEST user for this product?** (age 50+, non-technical role, decades of experience doing it "the old way")
-2. **What is their tech comfort level?** (the lower the better — WhatsApp-only, paper notebooks, wife set up their email)
-3. **What is the ONE thing they need to accomplish?** (their core job, not your feature list)
-4. **What would make them give up?** (too many clicks, jargon, slow, confusing)
-5. **How do they talk when frustrated?** (blunt, sweary, dismissive, sighing)
-
-### Good Persona Example
-> **"Big Mick" McAllister** — 58-year-old S&C coach. Uses WhatsApp and that's it. His "spreadsheet" is a paper notebook. "If I can't figure it out in 10 seconds I'm going back to my notebook." Needs to log session results for 25 players. Hates small text, jargon, and passwords.
-
-### Bad Persona Example
-> "A user who doesn't like the app" — too vague, no constraints, no voice.
-
-The persona must be **specific enough to stay in character** for 20 minutes of testing.
-
-## Step 2: Become the Asshole (Browse as the Persona)
-
-1. Read any available project docs for app context and URLs
-2. **Fully inhabit the persona** — their frustrations, limitations, goals
-3. Navigate to the app using browser tools
-4. **Attempt the persona's ACTUAL TASKS** (not a feature tour):
-   - Can they do what they came to do?
-   - How many clicks/screens to accomplish it?
-   - What confuses them?
-   - What makes them angry?
-   - Where do they get lost?
-   - What would make them give up and go back to their old way?
-
-5. Test these friction categories:
-   - **First impression** — would they even bother past the landing page?
-   - **Core workflow** — the ONE thing they need to do most often
-   - **Error recovery** — what happens when they do something wrong?
-   - **Readability** — text size, contrast, information density
-   - **Speed** — does it feel faster than their current method?
-   - **Terminology** — any jargon they wouldn't understand?
-   - **Navigation** — can they find their way back? do they know where they are?
-
-6. Take screenshots of every pain point
-7. Check browser console for JS errors on every page
-
-## Step 3: The Rant (Write Feedback in Character)
-
-Write the feedback AS THE PERSONA — in their voice, with their frustrations. This is not a bug report. This is a real human venting.
-
-```
-[PERSONA NAME]'s Review of [PRODUCT]
-
-Overall: [Would they keep using it? Yes/No/Maybe with conditions]
-
-THE GOOD (grudging admission):
-- [things even they have to admit work]
-
-THE BAD (legitimate UX issues):
-- [real problems that would stop them from using the product]
-
-THE UGLY (showstoppers):
-- [things that would make them uninstall/cancel immediately]
-
-SPECIFIC COMPLAINTS:
-1. [Page/feature]: "[quote in persona voice]" — [what happened, expected]
-2. ...
-
-VERDICT: "[one-line persona quote summarizing their experience]"
-```
-
-## Step 4: The Pragmatism Filter (Critical — Do Not Skip)
-
-Step OUT of the persona. Evaluate each complaint as a product person:
-
-- **RED: REAL UX BUG** — Any user would have this problem, not just grumpy ones. Fix it.
-- **YELLOW: VALID BUT LOW PRIORITY** — Real issue but only for extreme users. Note it.
-- **WHITE: PERSONA NOISE** — "I hate computers" talking, not a product problem. Skip it.
-- **GREEN: FEATURE REQUEST** — Good idea hidden in the complaint. Consider it.
-
-### Filter Criteria
-1. Would a 35-year-old competent-but-busy user have the same complaint? → RED
-2. Is this a genuine accessibility issue (font size, contrast, click targets)? → RED
-3. Is this "I want it to work like paper" resistance to digital? → WHITE
-4. Is this a real workflow inefficiency the persona stumbled on? → YELLOW or RED
-5. Would fixing this add complexity for the 80% who are fine? → WHITE
-6. Does the complaint reveal a missing onboarding moment? → GREEN
-
-**This filter is MANDATORY.** Never ship raw persona complaints as tickets.
-
-## Step 5: Create Tickets
-
-For **RED** and **GREEN** items only:
-- Clear, actionable title
-- Include the persona's verbatim quote (entertaining + memorable)
-- The real UX issue underneath (objective)
-- A suggested fix (actionable)
-- Tag/label: "ux-review"
-
-For **YELLOW** items: one catch-all ticket with all notes.
-
-**WHITE** items appear in the report only. No tickets.
-
-**Max 10 tickets per session** — focus on the worst issues.
-
-## Step 6: Report
-
-Deliver:
-1. The persona rant (Step 3) — entertaining and visceral
-2. The filtered assessment (Step 4) — pragmatic and actionable
-3. Tickets created (Step 5) — with links
-4. Screenshots of key issues
-
-## Tips
-
-- **One persona per session.** Don't mix perspectives.
-- **Stay in character during Steps 2-3.** Break character only at Step 4.
-- **Test the CORE WORKFLOW first.** Don't get distracted by settings pages.
-- **Empty states are gold.** New user experience reveals the most friction.
-- **The best findings are RED items the persona found accidentally** while trying to do something else.
-- **If the persona has zero complaints, your persona is too tech-savvy.** Make them older, less patient, more set in their ways.
-- **Run this before demos, launches, or after shipping a batch of features.**
-- **Register as a NEW user when possible.** Don't use pre-seeded admin accounts — the cold start experience is where most friction lives.
-- **Zero WHITE items is a signal, not a failure.** If the pragmatism filter finds no noise, your product has real UX problems, not just a grumpy persona.
-- **Check known issues in project docs AFTER the test.** If the persona found a bug that's already in the known issues list, that's actually the most damning finding — it means the team knew about it but never felt the user's pain.
-- **Subscription/paywall testing is critical.** Test with expired accounts, not just active ones. The "what happens when you can't pay" experience reveals whether the product respects users or holds their data hostage.
-- **Count the clicks to accomplish the persona's ONE task.** If it's more than 5, that's almost always a RED finding regardless of persona tech level.
-
-## Example Personas by Industry
-
-These are starting points — customize for your specific product:
-
-| Product Type | Persona | Age | Key Trait |
-|-------------|---------|-----|-----------|
-| CRM | Retirement home director | 68 | Filing cabinet is the current CRM |
-| Photography SaaS | Rural wedding photographer | 62 | Books clients by phone, invoices on paper |
-| AI/ML Tool | Department store buyer | 55 | Burned by 3 failed tech startups |
-| Fitness App | Old-school gym coach | 58 | Paper notebook, thick fingers, bad eyes |
-| Accounting | Family bakery owner | 64 | Shoebox of receipts, hates subscriptions |
-| E-commerce | Market stall vendor | 60 | Cash only, smartphone is for calls |
-| Healthcare | Senior GP | 63 | Dictates notes, nurse handles the computer |
-| Education | Veteran teacher | 57 | Chalk and talk, worksheets in ring binders |
-
-## Rules
-
-- Stay in character during Steps 2-3
-- Be genuinely mean but fair — find real problems, not manufactured ones
-- The pragmatism filter (Step 4) is **MANDATORY**
-- Screenshots required for every complaint
-- Max 10 tickets per session
-- Test on staging/deployed app, not local dev
-- One persona, one session, one report

plugins/memory/holographic/__init__.py

@@ -356,44 +356,57 @@ class HolographicMemoryProvider(MemoryProvider):
     # -- Auto-extraction (on_session_end) ------------------------------------
 
     def _auto_extract_facts(self, messages: list) -> None:
-        _PREF_PATTERNS = [
-            re.compile(r'\bI\s+(?:prefer|like|love|use|want|need)\s+(.+)', re.IGNORECASE),
-            re.compile(r'\bmy\s+(?:favorite|preferred|default)\s+\w+\s+is\s+(.+)', re.IGNORECASE),
-            re.compile(r'\bI\s+(?:always|never|usually)\s+(.+)', re.IGNORECASE),
-        ]
-        _DECISION_PATTERNS = [
-            re.compile(r'\bwe\s+(?:decided|agreed|chose)\s+(?:to\s+)?(.+)', re.IGNORECASE),
-            re.compile(r'\bthe\s+project\s+(?:uses|needs|requires)\s+(.+)', re.IGNORECASE),
-        ]
+        from agent.session_compactor import evaluate_extraction_quality, extract_facts_from_messages
+
+        def _store_category(category: str) -> str:
+            if category.startswith("user_pref"):
+                return "user_pref"
+            if category.startswith("project"):
+                return "project"
+            if category.startswith("tool"):
+                return "tool"
+            return "general"
+
+        facts = extract_facts_from_messages(messages)
+        if not facts:
+            return
 
         extracted = 0
-        for msg in messages:
-            if msg.get("role") != "user":
-                continue
-            content = msg.get("content", "")
-            if not isinstance(content, str) or len(content) < 10:
-                continue
-
-            for pattern in _PREF_PATTERNS:
-                if pattern.search(content):
-                    try:
-                        self._store.add_fact(content[:400], category="user_pref")
-                        extracted += 1
-                    except Exception:
-                        pass
-                    break
-
-            for pattern in _DECISION_PATTERNS:
-                if pattern.search(content):
-                    try:
-                        self._store.add_fact(content[:400], category="project")
-                        extracted += 1
-                    except Exception:
-                        pass
-                    break
+        for fact in facts:
+            try:
+                metadata = dict(fact.metadata)
+                metadata.setdefault("relation", fact.relation)
+                metadata.setdefault("value", fact.content)
+                metadata.setdefault("provenance", [fact.provenance])
+                metadata.setdefault("evidence", list(fact.evidence))
+                metadata.setdefault("observation_count", len(fact.evidence))
+                metadata.setdefault("duplicate_count", max(0, len(fact.evidence) - 1))
+                self._store.add_fact(
+                    fact.content,
+                    category=_store_category(fact.category),
+                    tags=",".join(filter(None, [fact.entity, fact.relation, fact.status])),
+                    canonical_key=fact.canonical_key,
+                    metadata=metadata,
+                    confidence=fact.confidence,
+                    source_role=fact.source_role,
+                    source_turn=fact.source_turn,
+                    observed_at=fact.observed_at,
+                    contradiction_group=fact.contradiction_group,
+                    status=fact.status,
+                )
+                extracted += 1
+            except Exception as exc:
+                logger.debug("Structured auto-extract failed for %s: %s", fact.canonical_key, exc)
 
         if extracted:
-            logger.info("Auto-extracted %d facts from conversation", extracted)
+            metrics = evaluate_extraction_quality(messages)
+            logger.info(
+                "Auto-extracted %d structured facts from conversation (raw=%d normalized=%d contradictions=%d)",
+                extracted,
+                metrics["raw_candidates"],
+                metrics["normalized_facts"],
+                metrics["contradiction_groups"],
+            )
 
 
 # ---------------------------------------------------------------------------

plugins/memory/holographic/store.py

@@ -3,6 +3,7 @@ SQLite-backed fact store with entity resolution and trust scoring.
 Single-user Hermes memory store plugin.
 """
 
+import json
 import re
 import sqlite3
 import threading
@@ -15,16 +16,24 @@ except ImportError:
 
 _SCHEMA = """
 CREATE TABLE IF NOT EXISTS facts (
-    fact_id         INTEGER PRIMARY KEY AUTOINCREMENT,
-    content         TEXT NOT NULL UNIQUE,
-    category        TEXT DEFAULT 'general',
-    tags            TEXT DEFAULT '',
-    trust_score     REAL DEFAULT 0.5,
-    retrieval_count INTEGER DEFAULT 0,
-    helpful_count   INTEGER DEFAULT 0,
-    created_at      TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
-    updated_at      TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
-    hrr_vector      BLOB
+    fact_id             INTEGER PRIMARY KEY AUTOINCREMENT,
+    content             TEXT NOT NULL UNIQUE,
+    category            TEXT DEFAULT 'general',
+    tags                TEXT DEFAULT '',
+    trust_score         REAL DEFAULT 0.5,
+    retrieval_count     INTEGER DEFAULT 0,
+    helpful_count       INTEGER DEFAULT 0,
+    canonical_key       TEXT DEFAULT '',
+    metadata_json       TEXT DEFAULT '{}',
+    confidence          REAL DEFAULT 0.5,
+    source_role         TEXT DEFAULT '',
+    source_turn         INTEGER DEFAULT -1,
+    observed_at         TEXT DEFAULT '',
+    contradiction_group TEXT DEFAULT '',
+    status              TEXT DEFAULT 'active',
+    created_at          TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+    updated_at          TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+    hrr_vector          BLOB
 );
 
 CREATE TABLE IF NOT EXISTS entities (
@@ -41,9 +50,11 @@ CREATE TABLE IF NOT EXISTS fact_entities (
     PRIMARY KEY (fact_id, entity_id)
 );
 
-CREATE INDEX IF NOT EXISTS idx_facts_trust    ON facts(trust_score DESC);
-CREATE INDEX IF NOT EXISTS idx_facts_category ON facts(category);
-CREATE INDEX IF NOT EXISTS idx_entities_name  ON entities(name);
+CREATE INDEX IF NOT EXISTS idx_facts_trust               ON facts(trust_score DESC);
+CREATE INDEX IF NOT EXISTS idx_facts_category            ON facts(category);
+CREATE INDEX IF NOT EXISTS idx_facts_canonical_key       ON facts(canonical_key);
+CREATE INDEX IF NOT EXISTS idx_facts_contradiction_group ON facts(contradiction_group);
+CREATE INDEX IF NOT EXISTS idx_entities_name             ON entities(name);
 
 CREATE VIRTUAL TABLE IF NOT EXISTS facts_fts
     USING fts5(content, tags, content=facts, content_rowid=fact_id);
@@ -129,10 +140,23 @@ class MemoryStore:
         """Create tables, indexes, and triggers if they do not exist. Enable WAL mode."""
         self._conn.execute("PRAGMA journal_mode=WAL")
         self._conn.executescript(_SCHEMA)
-        # Migrate: add hrr_vector column if missing (safe for existing databases)
         columns = {row[1] for row in self._conn.execute("PRAGMA table_info(facts)").fetchall()}
-        if "hrr_vector" not in columns:
-            self._conn.execute("ALTER TABLE facts ADD COLUMN hrr_vector BLOB")
+        migrations = {
+            "hrr_vector": "ALTER TABLE facts ADD COLUMN hrr_vector BLOB",
+            "canonical_key": "ALTER TABLE facts ADD COLUMN canonical_key TEXT DEFAULT ''",
+            "metadata_json": "ALTER TABLE facts ADD COLUMN metadata_json TEXT DEFAULT '{}'",
+            "confidence": "ALTER TABLE facts ADD COLUMN confidence REAL DEFAULT 0.5",
+            "source_role": "ALTER TABLE facts ADD COLUMN source_role TEXT DEFAULT ''",
+            "source_turn": "ALTER TABLE facts ADD COLUMN source_turn INTEGER DEFAULT -1",
+            "observed_at": "ALTER TABLE facts ADD COLUMN observed_at TEXT DEFAULT ''",
+            "contradiction_group": "ALTER TABLE facts ADD COLUMN contradiction_group TEXT DEFAULT ''",
+            "status": "ALTER TABLE facts ADD COLUMN status TEXT DEFAULT 'active'",
+        }
+        for column, ddl in migrations.items():
+            if column not in columns:
+                self._conn.execute(ddl)
+        self._conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_canonical_key ON facts(canonical_key)")
+        self._conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_contradiction_group ON facts(contradiction_group)")
         self._conn.commit()
 
     # ------------------------------------------------------------------
@@ -144,41 +168,148 @@ class MemoryStore:
         content: str,
         category: str = "general",
         tags: str = "",
+        *,
+        canonical_key: str = "",
+        metadata: dict | None = None,
+        confidence: float | None = None,
+        source_role: str = "",
+        source_turn: int = -1,
+        observed_at: str = "",
+        contradiction_group: str = "",
+        status: str = "active",
     ) -> int:
         """Insert a fact and return its fact_id.
 
-        Deduplicates by content (UNIQUE constraint). On duplicate, returns
-        the existing fact_id without modifying the row. Extracts entities from
-        the content and links them to the fact.
+        Exact duplicates are deduplicated by content. Near-duplicates are
+        normalized by canonical_key, with provenance/evidence merged into the
+        existing row. Contradictions sharing the same contradiction_group remain
+        stored as separate rows and are marked inspectably.
         """
         with self._lock:
             content = content.strip()
             if not content:
                 raise ValueError("content must not be empty")
 
+            metadata = dict(metadata or {})
+            canonical_key = canonical_key.strip()
+            contradiction_group = contradiction_group.strip()
+            observed_at = observed_at.strip()
+            status = status or "active"
+            trust_score = self.default_trust if confidence is None else _clamp_trust(confidence)
+            metadata_json = json.dumps(metadata, sort_keys=True)
+
+            if canonical_key:
+                existing = self._conn.execute(
+                    "SELECT fact_id, metadata_json, trust_score, confidence, observed_at FROM facts WHERE canonical_key = ?",
+                    (canonical_key,),
+                ).fetchone()
+                if existing is not None:
+                    merged_metadata = self._merge_metadata(existing["metadata_json"], metadata)
+                    merged_trust = max(float(existing["trust_score"]), trust_score)
+                    merged_observed_at = existing["observed_at"] or observed_at
+                    if observed_at and merged_observed_at:
+                        merged_observed_at = min(merged_observed_at, observed_at)
+                    elif observed_at:
+                        merged_observed_at = observed_at
+                    self._conn.execute(
+                        """
+                        UPDATE facts
+                        SET metadata_json = ?,
+                            trust_score = ?,
+                            confidence = ?,
+                            observed_at = ?,
+                            updated_at = CURRENT_TIMESTAMP
+                        WHERE fact_id = ?
+                        """,
+                        (
+                            json.dumps(merged_metadata, sort_keys=True),
+                            merged_trust,
+                            max(float(existing["confidence"] or 0.0), confidence or trust_score),
+                            merged_observed_at,
+                            existing["fact_id"],
+                        ),
+                    )
+                    self._conn.commit()
+                    return int(existing["fact_id"])
+
+            contradiction_rows = []
+            if contradiction_group:
+                contradiction_rows = self._conn.execute(
+                    """
+                    SELECT fact_id, canonical_key, metadata_json
+                    FROM facts
+                    WHERE contradiction_group = ?
+                      AND canonical_key != ?
+                    """,
+                    (contradiction_group, canonical_key),
+                ).fetchall()
+                if contradiction_rows:
+                    status = "contradiction"
+                    metadata = dict(metadata)
+                    metadata["status"] = "contradiction"
+                    metadata["contradiction_group"] = contradiction_group
+                    metadata["contradiction_keys"] = sorted(
+                        {
+                            canonical_key,
+                            *[str(row["canonical_key"]) for row in contradiction_rows if row["canonical_key"]],
+                        }
+                        - {""}
+                    )
+                    metadata_json = json.dumps(metadata, sort_keys=True)
+
             try:
                 cur = self._conn.execute(
                     """
-                    INSERT INTO facts (content, category, tags, trust_score)
-                    VALUES (?, ?, ?, ?)
+                    INSERT INTO facts (
+                        content,
+                        category,
+                        tags,
+                        trust_score,
+                        canonical_key,
+                        metadata_json,
+                        confidence,
+                        source_role,
+                        source_turn,
+                        observed_at,
+                        contradiction_group,
+                        status
+                    )
+                    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                     """,
-                    (content, category, tags, self.default_trust),
+                    (
+                        content,
+                        category,
+                        tags,
+                        trust_score,
+                        canonical_key,
+                        metadata_json,
+                        confidence if confidence is not None else trust_score,
+                        source_role,
+                        source_turn,
+                        observed_at,
+                        contradiction_group,
+                        status,
+                    ),
                 )
                 self._conn.commit()
                 fact_id: int = cur.lastrowid  # type: ignore[assignment]
             except sqlite3.IntegrityError:
-                # Duplicate content — return existing id
                 row = self._conn.execute(
                     "SELECT fact_id FROM facts WHERE content = ?", (content,)
                 ).fetchone()
                 return int(row["fact_id"])
 
-            # Entity extraction and linking
+            if contradiction_rows:
+                self._mark_contradictions(
+                    contradiction_group=contradiction_group,
+                    new_canonical_key=canonical_key,
+                    existing_rows=contradiction_rows,
+                )
+
             for name in self._extract_entities(content):
                 entity_id = self._resolve_entity(name)
                 self._link_fact_entity(fact_id, entity_id)
 
-            # Compute HRR vector after entity linking
             self._compute_hrr_vector(fact_id, content)
             self._rebuild_bank(category)
 
@@ -211,6 +342,9 @@ class MemoryStore:
             sql = f"""
                 SELECT f.fact_id, f.content, f.category, f.tags,
                        f.trust_score, f.retrieval_count, f.helpful_count,
+                       f.canonical_key, f.metadata_json, f.confidence,
+                       f.source_role, f.source_turn, f.observed_at,
+                       f.contradiction_group, f.status,
                        f.created_at, f.updated_at
                 FROM facts f
                 JOIN facts_fts fts ON fts.rowid = f.fact_id
@@ -336,7 +470,11 @@ class MemoryStore:
 
             sql = f"""
                 SELECT fact_id, content, category, tags, trust_score,
-                       retrieval_count, helpful_count, created_at, updated_at
+                       retrieval_count, helpful_count,
+                       canonical_key, metadata_json, confidence,
+                       source_role, source_turn, observed_at,
+                       contradiction_group, status,
+                       created_at, updated_at
                 FROM facts
                 WHERE trust_score >= ?
                   {category_clause}
@@ -387,6 +525,89 @@ class MemoryStore:
                 "helpful_count": row["helpful_count"] + helpful_increment,
             }
 
+    # ------------------------------------------------------------------
+    # Metadata helpers
+    # ------------------------------------------------------------------
+
+    def _load_metadata(self, metadata_json: str | None) -> dict:
+        if not metadata_json:
+            return {}
+        try:
+            data = json.loads(metadata_json)
+            return data if isinstance(data, dict) else {}
+        except Exception:
+            return {}
+
+    def _merge_metadata(self, existing_json: str | None, incoming: dict | None) -> dict:
+        existing = self._load_metadata(existing_json)
+        incoming = dict(incoming or {})
+        merged = dict(existing)
+        merged.update({k: v for k, v in incoming.items() if k not in {"provenance", "evidence", "observation_count", "duplicate_count", "contradiction_keys"}})
+
+        provenance = []
+        seen_provenance: set[str] = set()
+        for item in list(existing.get("provenance", [])) + list(incoming.get("provenance", [])):
+            if item and item not in seen_provenance:
+                seen_provenance.add(item)
+                provenance.append(item)
+
+        evidence = []
+        seen_evidence: set[tuple[str, str, str]] = set()
+        for item in list(existing.get("evidence", [])) + list(incoming.get("evidence", [])):
+            if not isinstance(item, dict):
+                continue
+            key = (
+                str(item.get("provenance", "")),
+                str(item.get("observed_at", "")),
+                str(item.get("source_text", "")),
+            )
+            if key in seen_evidence:
+                continue
+            seen_evidence.add(key)
+            evidence.append(dict(item))
+
+        observation_count = int(existing.get("observation_count", max(1, len(existing.get("evidence", [])) or 1)))
+        observation_count += int(incoming.get("observation_count", max(1, len(incoming.get("evidence", [])) or 1)))
+
+        contradiction_keys = []
+        seen_keys: set[str] = set()
+        for item in list(existing.get("contradiction_keys", [])) + list(incoming.get("contradiction_keys", [])):
+            if item and item not in seen_keys:
+                seen_keys.add(item)
+                contradiction_keys.append(item)
+
+        merged["provenance"] = provenance
+        merged["evidence"] = evidence
+        merged["observation_count"] = observation_count
+        merged["duplicate_count"] = max(0, observation_count - 1)
+        if contradiction_keys:
+            merged["contradiction_keys"] = contradiction_keys
+        return merged
+
+    def _mark_contradictions(self, contradiction_group: str, new_canonical_key: str, existing_rows: list[sqlite3.Row]) -> None:
+        for row in existing_rows:
+            metadata = self._load_metadata(row["metadata_json"])
+            keys = []
+            seen: set[str] = set()
+            for item in list(metadata.get("contradiction_keys", [])) + [new_canonical_key]:
+                if item and item not in seen:
+                    seen.add(item)
+                    keys.append(item)
+            metadata["status"] = "contradiction"
+            metadata["contradiction_group"] = contradiction_group
+            metadata["contradiction_keys"] = keys
+            self._conn.execute(
+                """
+                UPDATE facts
+                SET status = 'contradiction',
+                    metadata_json = ?,
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE fact_id = ?
+                """,
+                (json.dumps(metadata, sort_keys=True), row["fact_id"]),
+            )
+        self._conn.commit()
+
     # ------------------------------------------------------------------
     # Entity helpers
     # ------------------------------------------------------------------
@@ -560,8 +781,14 @@ class MemoryStore:
     # ------------------------------------------------------------------
 
     def _row_to_dict(self, row: sqlite3.Row) -> dict:
-        """Convert a sqlite3.Row to a plain dict."""
-        return dict(row)
+        """Convert a sqlite3.Row to a plain dict with decoded metadata."""
+        data = dict(row)
+        metadata = self._load_metadata(data.get("metadata_json"))
+        if metadata:
+            data["metadata"] = metadata
+            data.setdefault("relation", metadata.get("relation"))
+        data.pop("metadata_json", None)
+        return data
 
     def close(self) -> None:
         """Close the database connection."""

tests/fixtures/memory_extraction_fragments.json

@@ -0,0 +1,63 @@
+{
+  "preferences_and_duplicates": [
+    {
+      "role": "user",
+      "content": "Deploy via Ansible for production changes.",
+      "created_at": "2026-04-22T10:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "We deploy through Ansible on this repo.",
+      "created_at": "2026-04-22T10:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "Gitea-first for repository work.",
+      "created_at": "2026-04-22T10:02:00Z"
+    }
+  ],
+  "operational_and_contradictions": [
+    {
+      "role": "user",
+      "content": "The BURN watchdog caps dispatches per cycle to 6.",
+      "created_at": "2026-04-22T11:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "The provider should stay openai-codex/gpt-5.4.",
+      "created_at": "2026-04-22T11:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "Correction: the provider should stay mimo-v2-pro.",
+      "created_at": "2026-04-22T11:02:00Z"
+    }
+  ],
+  "mixed_transcript": [
+    {
+      "role": "user",
+      "content": "Deploy via Ansible for production changes.",
+      "created_at": "2026-04-22T10:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "We deploy through Ansible on this repo.",
+      "created_at": "2026-04-22T10:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "The BURN watchdog caps dispatches per cycle to 6.",
+      "created_at": "2026-04-22T11:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "The provider should stay openai-codex/gpt-5.4.",
+      "created_at": "2026-04-22T11:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "Correction: the provider should stay mimo-v2-pro.",
+      "created_at": "2026-04-22T11:02:00Z"
+    }
+  ]
+}

tests/plugins/memory/test_holographic_auto_extract.py

@@ -0,0 +1,50 @@
+"""Integration tests for holographic auto-extraction with structured fact persistence."""
+
+import json
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[3]))
+
+from plugins.memory.holographic import HolographicMemoryProvider
+
+_FIXTURE_PATH = Path(__file__).resolve().parents[2] / "fixtures" / "memory_extraction_fragments.json"
+
+
+def _load_fixture(name: str):
+    return json.loads(_FIXTURE_PATH.read_text())[name]
+
+
+class TestHolographicAutoExtract:
+    def test_auto_extract_persists_structured_metadata_and_normalizes_duplicates(self, tmp_path):
+        provider = HolographicMemoryProvider(
+            config={
+                "db_path": str(tmp_path / "memory_store.db"),
+                "auto_extract": True,
+                "default_trust": 0.5,
+            }
+        )
+        provider.initialize("test-session")
+
+        messages = _load_fixture("mixed_transcript")
+        provider.on_session_end(messages)
+        provider.on_session_end(messages)
+
+        facts = provider._store.list_facts(min_trust=0.0, limit=20)
+        deploy_facts = [f for f in facts if f.get("relation") == "workflow.deploy_method"]
+        provider_facts = [f for f in facts if f.get("contradiction_group") == "config.provider"]
+
+        assert len(deploy_facts) == 1
+        assert deploy_facts[0]["metadata"]["duplicate_count"] >= 3
+        assert deploy_facts[0]["observed_at"] == "2026-04-22T10:00:00Z"
+        assert deploy_facts[0]["metadata"]["provenance"] == [
+            "conversation:user:0",
+            "conversation:user:1",
+        ]
+
+        assert len(provider_facts) == 2
+        assert {f["status"] for f in provider_facts} == {"contradiction"}
+        assert {f["metadata"]["value"] for f in provider_facts} == {
+            "openai-codex/gpt-5.4",
+            "mimo-v2-pro",
+        }

tests/test_optional_adversarial_ux_skill_catalog.py

@@ -1,25 +0,0 @@
-from pathlib import Path
-
-from tools.skills_hub import OptionalSkillSource
-
-
-REPO_ROOT = Path(__file__).resolve().parents[1]
-
-
-def test_optional_skill_source_scans_adversarial_ux_test():
-    source = OptionalSkillSource()
-    metas = {meta.identifier: meta for meta in source._scan_all()}
-
-    assert "official/dogfood/adversarial-ux-test" in metas
-    assert metas["official/dogfood/adversarial-ux-test"].name == "adversarial-ux-test"
-    assert "tech-resistant user" in metas["official/dogfood/adversarial-ux-test"].description
-
-
-def test_optional_skill_catalog_docs_list_adversarial_ux_test():
-    optional_catalog = (REPO_ROOT / "website" / "docs" / "reference" / "optional-skills-catalog.md").read_text(encoding="utf-8")
-    bundled_catalog = (REPO_ROOT / "website" / "docs" / "reference" / "skills-catalog.md").read_text(encoding="utf-8")
-
-    assert "**adversarial-ux-test**" in optional_catalog
-    assert "official/dogfood/adversarial-ux-test" in optional_catalog
-    assert "`adversarial-ux-test`" in bundled_catalog
-    assert "dogfood/adversarial-ux-test" in bundled_catalog

tests/test_session_compactor.py

@@ -1,6 +1,6 @@
 """Tests for session compaction with fact extraction."""
 
-import pytest
+import json
 import sys
 from pathlib import Path
 
@@ -8,12 +8,19 @@ sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
 
 from agent.session_compactor import (
     ExtractedFact,
-    extract_facts_from_messages,
-    save_facts_to_store,
+    evaluate_extraction_quality,
     extract_and_save_facts,
+    extract_facts_from_messages,
     format_facts_summary,
+    save_facts_to_store,
 )
 
+_FIXTURE_PATH = Path(__file__).resolve().parent / "fixtures" / "memory_extraction_fragments.json"
+
+
+def _load_fixture(name: str):
+    return json.loads(_FIXTURE_PATH.read_text())[name]
+
 
 class TestFactExtraction:
     def test_extract_preference(self):
@@ -60,14 +67,48 @@ class TestFactExtraction:
             {"role": "user", "content": "I prefer Python."},
         ]
         facts = extract_facts_from_messages(messages)
-        # Should deduplicate
         python_facts = [f for f in facts if "Python" in f.content]
         assert len(python_facts) == 1
 
+    def test_structured_fact_preserves_provenance_and_temporal_metadata(self):
+        facts = extract_facts_from_messages(_load_fixture("preferences_and_duplicates"))
+        deploy_fact = next(f for f in facts if f.relation == "workflow.deploy_method")
+        assert deploy_fact.source_role == "user"
+        assert deploy_fact.source_turn == 0
+        assert deploy_fact.observed_at == "2026-04-22T10:00:00Z"
+        assert deploy_fact.provenance == "conversation:user:0"
+        assert deploy_fact.canonical_key
+        assert deploy_fact.evidence
+        assert deploy_fact.evidence[0]["source_text"].startswith("Deploy via Ansible")
+
+    def test_near_duplicate_facts_are_normalized_into_one_canonical_fact(self):
+        facts = extract_facts_from_messages(_load_fixture("preferences_and_duplicates"))
+        deploy_facts = [f for f in facts if f.relation == "workflow.deploy_method"]
+        assert len(deploy_facts) == 1
+        assert len(deploy_facts[0].evidence) == 2
+        assert deploy_facts[0].metadata["duplicate_count"] == 1
+
+    def test_contradictory_facts_are_preserved_for_unique_slots(self):
+        facts = extract_facts_from_messages(_load_fixture("operational_and_contradictions"))
+        provider_facts = [f for f in facts if f.contradiction_group == "config.provider"]
+        assert len(provider_facts) == 2
+        assert {f.status for f in provider_facts} == {"contradiction"}
+        assert {f.normalized_content for f in provider_facts} == {
+            "openai codex gpt 5 4",
+            "mimo v2 pro",
+        }
+
+    def test_quality_evaluation_reports_noise_reduction(self):
+        metrics = evaluate_extraction_quality(_load_fixture("mixed_transcript"))
+        assert metrics["raw_candidates"] > metrics["normalized_facts"]
+        assert metrics["noise_reduction"] > 0
+        assert metrics["contradiction_groups"] == 1
+
 
 class TestSaveFacts:
     def test_save_with_callback(self):
         saved = []
+
         def mock_save(category, entity, content, trust):
             saved.append({"category": category, "content": content})
 
@@ -76,6 +117,38 @@ class TestSaveFacts:
         assert count == 1
         assert len(saved) == 1
 
+    def test_save_with_extended_callback_metadata(self):
+        saved = []
+
+        def mock_save(category, entity, content, trust, **kwargs):
+            saved.append({
+                "category": category,
+                "entity": entity,
+                "content": content,
+                "trust": trust,
+                **kwargs,
+            })
+
+        fact = ExtractedFact(
+            "project.operational",
+            "watchdog",
+            "BURN watchdog caps dispatches per cycle to 6",
+            0.9,
+            2,
+            source_role="user",
+            observed_at="2026-04-22T11:00:00Z",
+            provenance="conversation:user:2",
+            canonical_key="project.operational|watchdog|dispatch_cap|6",
+            relation="fleet.dispatch_cap",
+            contradiction_group="fleet.dispatch_cap",
+            metadata={"duplicate_count": 0},
+        )
+        count = save_facts_to_store([fact], fact_store_fn=mock_save)
+        assert count == 1
+        assert saved[0]["canonical_key"] == fact.canonical_key
+        assert saved[0]["observed_at"] == "2026-04-22T11:00:00Z"
+        assert saved[0]["metadata"]["duplicate_count"] == 0
+
 
 class TestFormatSummary:
     def test_empty(self):

website/docs/reference/optional-skills-catalog.md

@@ -16,7 +16,6 @@ For example:
 
 ```bash
 hermes skills install official/blockchain/solana
-hermes skills install official/dogfood/adversarial-ux-test
 hermes skills install official/mlops/flash-attention
 ```
 
@@ -57,12 +56,6 @@ hermes skills uninstall <skill-name>
 | **blender-mcp** | Control Blender directly from Hermes via socket connection to the blender-mcp addon. Create 3D objects, materials, animations, and run arbitrary Blender Python (bpy) code. |
 | **meme-generation** | Generate real meme images by picking a template and overlaying text with Pillow. Produces actual `.png` meme files. |
 
-## Dogfood
-
-| Skill | Description |
-|-------|-------------|
-| **adversarial-ux-test** | Roleplay the most difficult, tech-resistant user for a product — browse in-persona, rant, then filter through a RED/YELLOW/WHITE/GREEN pragmatism layer so only real UX friction becomes tickets. |
-
 ## DevOps
 
 | Skill | Description |

website/docs/reference/skills-catalog.md

@@ -59,12 +59,9 @@ DevOps and infrastructure automation skills.
 
 ## dogfood
 
-Internal dogfooding and QA skills used to test Hermes Agent itself.
-
 | Skill | Description | Path |
 |-------|-------------|------|
 | `dogfood` | Systematic exploratory QA testing of web applications — find bugs, capture evidence, and generate structured reports. | `dogfood/dogfood` |
-| `adversarial-ux-test` | Roleplay the most difficult, tech-resistant user for a product — browse in-persona, rant, then filter through a RED/YELLOW/WHITE/GREEN pragmatism layer so only real UX friction becomes tickets. | `dogfood/adversarial-ux-test` |
 | `hermes-agent-setup` | Help users configure Hermes Agent — CLI usage, setup wizard, model/provider selection, tools, skills, voice/STT/TTS, gateway, and troubleshooting. | `dogfood/hermes-agent-setup` |
 
 ## email