foxclaw 0.1.0 → 0.1.2

package/python/vendor/hmr/engines/evolution.py

@@ -0,0 +1,710 @@
+"""
+Self-Evolution Engine - 自我演化引擎
+HMR v2.0 新增
+
+解决的问题：
+  记忆库随时间膨胀，重复信息累积，结构越来越混乱
+  Self-Evolution Engine 让 HMR 主动优化自身：
+    1. PatternDetector    → 发现重复/矛盾/知识空白
+    2. KnowledgeAbstractor → 把相似记忆提炼为更高阶的抽象概念
+    3. ContradictionResolver → 处理互相矛盾的记忆
+    4. StrategyOptimizer  → 分析使用模式，推荐 LifecycleConfig 优化参数
+
+演化是渐进的、保守的：
+  - 不主动删除记忆（只降低权重）
+  - 每次只做小步改变
+  - 每次演化都记录操作日志（可追溯）
+"""
+
+import re
+import math
+import json
+import threading
+from pathlib import Path
+from typing import List, Dict, Any, Optional, Tuple, Set
+from dataclasses import dataclass, field
+from datetime import datetime
+from collections import Counter, defaultdict
+
+
+# ============================================================================
+# 演化日志
+# ============================================================================
+
+@dataclass
+class EvolutionLog:
+    """一次演化操作的记录"""
+    operation: str                   # abstract / resolve / optimize / cluster
+    affected_ids: List[str]          # 受影响的记忆 ID
+    result_id: Optional[str]         # 产生的新记忆 ID（如果有）
+    description: str
+    timestamp: datetime = field(default_factory=datetime.utcnow)
+
+    def to_dict(self) -> dict:
+        return {
+            "operation": self.operation,
+            "affected_ids": self.affected_ids,
+            "result_id": self.result_id,
+            "description": self.description,
+            "timestamp": self.timestamp.isoformat(),
+        }
+
+
+# ============================================================================
+# Pattern Detector（模式识别）
+# ============================================================================
+
+class PatternDetector:
+    """
+    在记忆库中发现模式：
+      - 高重复率的记忆簇（适合抽象）
+      - 互相矛盾的记忆（需要解决）
+      - 知识空白（高频查询但无对应记忆）
+    """
+
+    def find_duplicate_clusters(
+        self,
+        memories,
+        vector_store=None,
+        similarity_threshold: float = 0.75,
+        min_cluster_size: int = 3
+    ) -> List[List[Any]]:
+        """
+        找出相似度 > threshold 的记忆簇（适合压缩抽象）
+
+        使用向量余弦相似度（有 VectorStore 时），
+        降级到标题词重叠（无向量时）。
+        """
+        assigned: Set[str] = set()
+
+        if vector_store and len(vector_store.vectors) > 0:
+            return self._cluster_by_vector(
+                memories, vector_store, similarity_threshold, min_cluster_size, assigned
+            )
+        else:
+            return self._cluster_by_text(memories, min_cluster_size)
+
+    def _cluster_by_vector(
+        self, memories, vector_store, threshold, min_size, assigned
+    ) -> List[List[Any]]:
+        clusters = []
+        mem_list = [m for m in memories if m.id in vector_store.vectors]
+
+        for i, mem_a in enumerate(mem_list):
+            if mem_a.id in assigned:
+                continue
+            cluster = [mem_a]
+            vec_a = vector_store.vectors[mem_a.id]
+
+            for mem_b in mem_list[i+1:]:
+                if mem_b.id in assigned or mem_b.type != mem_a.type:
+                    continue
+                vec_b = vector_store.vectors.get(mem_b.id)
+                if vec_b is None:
+                    continue
+                sim = self._cosine(vec_a, vec_b)
+                if sim >= threshold:
+                    cluster.append(mem_b)
+
+            if len(cluster) >= min_size:
+                for m in cluster:
+                    assigned.add(m.id)
+                clusters.append(cluster)
+
+        return clusters
+
+    def _cluster_by_text(self, memories, min_size: int) -> List[List[Any]]:
+        """标题词重叠聚类（降级方案）"""
+        by_type: Dict[str, List[Any]] = defaultdict(list)
+        for m in memories:
+            by_type[m.type].append(m)
+
+        clusters = []
+        for mems in by_type.values():
+            if len(mems) >= min_size:
+                # 按标题相似度进一步分组
+                groups: Dict[str, List[Any]] = defaultdict(list)
+                for m in mems:
+                    key_words = frozenset(
+                        re.findall(r'\b\w{3,}\b', m.title.lower())[:3]
+                    )
+                    groups[str(sorted(key_words))].append(m)
+                for g in groups.values():
+                    if len(g) >= min_size:
+                        clusters.append(g)
+        return clusters
+
+    def find_contradictions(self, memories) -> List[Tuple[Any, Any, str]]:
+        """
+        找出互相矛盾的记忆对。
+
+        简单规则：
+        - 同一主题（标题词重叠 > 60%），但 confidence 差异 > 0.4
+        - 包含明确的否定词对：
+            ("不应该", "应该"), ("避免", "推荐"), ("失败", "成功")
+        """
+        contradictions = []
+        positive = {"应该", "推荐", "最佳", "成功", "有效", "recommend", "best", "success"}
+        negative = {"不应该", "避免", "失败", "禁止", "avoid", "fail", "never", "wrong"}
+
+        mem_list = list(memories)
+        for i, ma in enumerate(mem_list):
+            for mb in mem_list[i+1:]:
+                if ma.type != mb.type:
+                    continue
+                # 标题重叠
+                words_a = set(re.findall(r'\b\w{3,}\b', ma.title.lower()))
+                words_b = set(re.findall(r'\b\w{3,}\b', mb.title.lower()))
+                if not words_a or not words_b:
+                    continue
+                overlap = len(words_a & words_b) / min(len(words_a), len(words_b))
+                if overlap < 0.4:
+                    continue
+
+                # 语义方向相反
+                text_a = (ma.content + ma.title).lower()
+                text_b = (mb.content + mb.title).lower()
+                a_pos = any(w in text_a for w in positive)
+                a_neg = any(w in text_a for w in negative)
+                b_pos = any(w in text_b for w in positive)
+                b_neg = any(w in text_b for w in negative)
+
+                if (a_pos and b_neg) or (a_neg and b_pos):
+                    contradictions.append((
+                        ma, mb,
+                        f"方向相反：A={'肯定' if a_pos else '否定'}，"
+                        f"B={'肯定' if b_pos else '否定'}"
+                    ))
+
+                # 置信度差异大（同主题不同结论）
+                elif abs(ma.confidence - mb.confidence) > 0.4 and overlap > 0.6:
+                    contradictions.append((
+                        ma, mb,
+                        f"置信度矛盾：A={ma.confidence:.2f}，B={mb.confidence:.2f}"
+                    ))
+
+        return contradictions[:20]  # 最多返回 20 对
+
+    def find_knowledge_gaps(
+        self,
+        memories,
+        recall_history: List[Dict[str, Any]],
+        min_query_count: int = 3
+    ) -> List[str]:
+        """
+        找出知识空白：频繁被查询但没有对应记忆的主题。
+
+        recall_history 格式：[{"query": str, "hit_count": int, "miss": bool}]
+        """
+        # 统计未命中的查询
+        miss_queries: Counter = Counter()
+        for event in recall_history:
+            if event.get("miss"):
+                miss_queries[event["query"]] += 1
+
+        # 找出频繁失败的查询
+        gaps = []
+        for query, count in miss_queries.most_common(10):
+            if count >= min_query_count:
+                # 检查是否真的没有对应记忆
+                has_memory = any(
+                    query.lower() in (m.title + m.content).lower()
+                    for m in memories
+                )
+                if not has_memory:
+                    gaps.append(f"知识空白：'{query}'（查询{count}次，无对应记忆）")
+
+        return gaps
+
+    @staticmethod
+    def _cosine(a: List[float], b: List[float]) -> float:
+        dot = sum(x * y for x, y in zip(a, b))
+        na = math.sqrt(sum(x * x for x in a))
+        nb = math.sqrt(sum(x * x for x in b))
+        return dot / (na * nb) if na and nb else 0.0
+
+
+# ============================================================================
+# Knowledge Abstractor（知识抽象）
+# ============================================================================
+
+class KnowledgeAbstractor:
+    """
+    把相似/重复的记忆提炼为更高阶的抽象概念。
+
+    两种抽象模式：
+    1. 归纳抽象：多个 execution 记录 → 一条 insight（规律）
+    2. 概念提升：多个 concept → 一条更高阶 concept
+    """
+
+    def abstract_cluster(
+        self,
+        memories,
+        llm_fn=None
+    ) -> Optional[Dict[str, Any]]:
+        """
+        将一组相似记忆抽象为一条新记忆。
+
+        返回：适合传给 hmr.ingest 的参数 dict，或 None（如果不值得抽象）
+        """
+        if len(memories) < 2:
+            return None
+
+        mtypes = [m.type for m in memories]
+        dominant_type = Counter(mtypes).most_common(1)[0][0]
+
+        # 根据类型选择目标类型
+        target_type = {
+            "execution": "reflection",   # 执行记录 → 反思洞察
+            "concept":   "concept",      # 概念 → 更高阶概念
+            "decision":  "concept",      # 决策 → 决策模式
+            "task":      "workflow",     # 任务 → 工作流
+        }.get(dominant_type, "concept")
+
+        # 生成内容
+        content = self._generate_abstract(memories, dominant_type, llm_fn)
+        if not content:
+            return None
+
+        # 合并所有标签和依赖
+        all_tags = list(set(t for m in memories for t in m.tags))[:8]
+        all_deps = list(set(d for m in memories for d in m.runtime_dependencies))[:5]
+        avg_confidence = sum(m.confidence for m in memories) / len(memories)
+
+        return {
+            "content": content,
+            "memory_type": target_type,
+            "title": f"[抽象] {self._extract_topic(memories)}",
+            "metadata": {
+                "tags": all_tags + ["abstracted", f"from_{len(memories)}_memories"],
+                "runtime_dependencies": all_deps,
+                "confidence": round(min(0.95, avg_confidence + 0.1), 2),
+            }
+        }
+
+    def _generate_abstract(
+        self, memories, dominant_type: str, llm_fn=None
+    ) -> Optional[str]:
+        """生成抽象内容（LLM 或 TF-IDF）"""
+        if llm_fn:
+            try:
+                combined = "\n---\n".join(
+                    f"[{m.type}] {m.title}: {m.content[:200]}"
+                    for m in memories
+                )
+                return llm_fn(combined, len(memories))
+            except Exception:
+                pass
+
+        # TF-IDF 降级
+        return self._tfidf_abstract(memories, dominant_type)
+
+    def _tfidf_abstract(self, memories, dominant_type: str) -> str:
+        stop = {"的","了","是","在","a","an","the","is","to","of","and","in","for"}
+        all_text = " ".join(m.content for m in memories)
+        words = re.findall(r'\b\w{2,}\b', all_text.lower())
+        kws = [w for w in words if w not in stop]
+        top_kws = [w for w, _ in Counter(kws).most_common(10)]
+
+        type_prefix = {
+            "execution": "综合执行经验规律",
+            "concept":   "抽象知识提炼",
+            "decision":  "决策模式总结",
+        }.get(dominant_type, "知识抽象")
+
+        return (
+            f"{type_prefix}（来自 {len(memories)} 条记忆）：\n"
+            f"核心主题：{', '.join(top_kws[:6])}。\n"
+            f"时间跨度：{memories[0].created_at.date()} ~ {memories[-1].created_at.date()}。"
+        )
+
+    def _extract_topic(self, memories) -> str:
+        words = []
+        for m in memories:
+            words.extend(re.findall(r'\b\w{3,}\b', m.title))
+        if not words:
+            return "综合记忆"
+        top = Counter(words).most_common(2)
+        return " + ".join(w for w, _ in top)
+
+
+# ============================================================================
+# Contradiction Resolver（矛盾解决）
+# ============================================================================
+
+class ContradictionResolver:
+    """
+    处理互相矛盾的记忆对。
+
+    策略：
+    - 保留置信度更高的那条（主要版本）
+    - 降低另一条的置信度和 temporal_weight
+    - 创建一条新的 reflection 记忆记录这个矛盾和解决过程
+    """
+
+    def resolve(
+        self,
+        mem_a,
+        mem_b,
+        reason: str,
+        memory_fs=None
+    ) -> Dict[str, Any]:
+        """
+        解决两条记忆之间的矛盾。
+
+        返回：
+        {
+            "winner": mem,        # 保留的记忆
+            "loser": mem,         # 降权的记忆
+            "reflection_content": str  # 建议存储的矛盾说明
+        }
+        """
+        # 根据置信度 + 访问频率决定哪条是"正确的"
+        score_a = mem_a.confidence * 0.6 + min(1.0, mem_a.access_count / 10) * 0.4
+        score_b = mem_b.confidence * 0.6 + min(1.0, mem_b.access_count / 10) * 0.4
+
+        winner, loser = (mem_a, mem_b) if score_a >= score_b else (mem_b, mem_a)
+
+        # 降低 loser 的权重
+        loser.confidence = max(0.1, loser.confidence * 0.5)
+        loser.temporal_weight = max(0.1, loser.temporal_weight * 0.5)
+        if "contradicted" not in loser.tags:
+            loser.tags.append("contradicted")
+
+        if memory_fs:
+            memory_fs.write_memory(loser)
+
+        reflection_content = (
+            f"发现矛盾记忆对：\n"
+            f"  保留：[{winner.type}]《{winner.title}》（置信度={winner.confidence:.2f}）\n"
+            f"  降权：[{loser.type}]《{loser.title}》\n"
+            f"  原因：{reason}"
+        )
+
+        return {
+            "winner": winner,
+            "loser": loser,
+            "reflection_content": reflection_content
+        }
+
+
+# ============================================================================
+# Strategy Optimizer（策略优化器）
+# ============================================================================
+
+class StrategyOptimizer:
+    """
+    分析 HMR 的使用模式，推荐优化参数。
+    """
+
+    def suggest_lifecycle_config(
+        self,
+        memory_stats: Dict[str, Any],
+        scheduler_stats: Dict[str, Any],
+        lifecycle_stats: Dict[str, Any],
+    ) -> Dict[str, Any]:
+        """
+        分析当前系统状态，推荐 LifecycleConfig 参数。
+        """
+        suggestions = {}
+        reasons = []
+
+        total = memory_stats.get("total_memories", 0)
+        by_type = memory_stats.get("memory_types", {})
+        dominant_strategy = scheduler_stats.get("dominant_strategy", "semantic")
+        hit_rate = float(scheduler_stats.get("cache_hit_rate", "0%").rstrip("%")) / 100
+        dormant = lifecycle_stats.get("by_state", {}).get("dormant", 0)
+
+        # 1. max_memories_per_type
+        max_per_type = max(by_type.values(), default=0)
+        if max_per_type > 150:
+            suggestions["max_memories_per_type"] = 80
+            reasons.append(f"某类型记忆数 {max_per_type} 过多，建议降低阈值到 80")
+        elif max_per_type < 20:
+            suggestions["max_memories_per_type"] = 200
+            reasons.append("记忆数量较少，放宽压缩限制")
+
+        # 2. check_interval_ingests
+        if total > 500:
+            suggestions["check_interval_ingests"] = 5
+            reasons.append("记忆库较大，建议加密检查频率（每5次）")
+        elif total < 50:
+            suggestions["check_interval_ingests"] = 20
+            reasons.append("记忆库较小，降低检查频率（每20次）")
+
+        # 3. prune_retrievability
+        if dormant > total * 0.3:
+            suggestions["prune_retrievability"] = 0.08
+            reasons.append(f"休眠记忆占比 {dormant/max(total,1):.0%}，适当提高删除阈值")
+
+        # 4. JIT 相关
+        if dominant_strategy == "jit" and hit_rate < 0.4:
+            reasons.append("JIT 使用多但命中率低，建议检查查询质量或增加记忆数量")
+
+        return {
+            "suggested_config": suggestions,
+            "reasons": reasons,
+            "current_health": self._compute_health(total, hit_rate, dormant)
+        }
+
+    def analyze_recall_patterns(
+        self, scheduler_stats: Dict[str, Any]
+    ) -> List[str]:
+        """分析召回使用模式，给出优化建议"""
+        insights = []
+        counts = scheduler_stats.get("strategy_counts", {})
+        hit_rate = scheduler_stats.get("cache_hit_rate", "0%")
+
+        total_calls = sum(counts.values())
+        if total_calls == 0:
+            return ["尚无足够数据进行分析"]
+
+        for strat, cnt in counts.items():
+            ratio = cnt / total_calls
+            if ratio > 0.6:
+                insights.append(f"{strat} 策略占主导（{ratio:.0%}），系统查询风格一致")
+            elif ratio > 0.3:
+                insights.append(f"{strat} 策略频繁使用（{ratio:.0%}）")
+
+        hit = float(str(hit_rate).rstrip("%")) / 100
+        if hit > 0.5:
+            insights.append(f"热缓存命中率 {hit_rate}，查询重复度高，可考虑扩大缓存容量")
+        elif hit < 0.2:
+            insights.append(f"热缓存命中率 {hit_rate}，查询多样性高，缓存效益有限")
+
+        return insights
+
+    def _compute_health(self, total: int, hit_rate: float, dormant: int) -> str:
+        if total == 0:
+            return "空库"
+        dormant_ratio = dormant / total
+        if dormant_ratio > 0.5:
+            return "⚠️ 记忆老化严重，建议运行压缩"
+        elif dormant_ratio > 0.3:
+            return "⚡ 一般，有一定老化"
+        elif hit_rate > 0.4:
+            return "✅ 健康，缓存效率良好"
+        else:
+            return "✅ 健康"
+
+
+# ============================================================================
+# Self-Evolution Engine（主类）
+# ============================================================================
+
+class SelfEvolutionEngine:
+    """
+    自我演化引擎
+
+    用法：
+        evo = SelfEvolutionEngine(storage_path="./hmr_data/evolution")
+        evo.register_ingest_fn(hmr.ingest)
+
+        # 运行完整演化周期
+        report = evo.evolve(
+            memories=hmr.memory_fs.list_memories(),
+            vector_store=hmr.vector_store,
+            memory_fs=hmr.memory_fs,
+        )
+
+        print(report["summary"])
+        for action in report["actions"]:
+            print(f"  {action}")
+    """
+
+    def __init__(self, storage_path: Optional[str] = None):
+        self.storage_path = Path(storage_path) if storage_path else None
+        self._lock = threading.Lock()
+
+        self.detector   = PatternDetector()
+        self.abstractor = KnowledgeAbstractor()
+        self.resolver   = ContradictionResolver()
+        self.optimizer  = StrategyOptimizer()
+
+        self.logs: List[EvolutionLog] = []
+        self._ingest_fn = None
+        self._llm_fn = None
+
+        if self.storage_path:
+            self.storage_path.mkdir(parents=True, exist_ok=True)
+            self._load_logs()
+
+    def register_ingest_fn(self, fn):
+        self._ingest_fn = fn
+
+    def register_llm_fn(self, fn):
+        """注入 LLM 摘要函数（可选）"""
+        self._llm_fn = fn
+
+    def evolve(
+        self,
+        memories,
+        vector_store=None,
+        memory_fs=None,
+        system_stats: Optional[Dict[str, Any]] = None,
+        dry_run: bool = False,
+    ) -> Dict[str, Any]:
+        """
+        运行一次完整的演化周期。
+
+        dry_run=True：只分析，不实际修改（用于预览）
+
+        返回演化报告：
+        {
+            "summary": str,
+            "actions": List[str],
+            "abstractions": int,
+            "contradictions_resolved": int,
+            "suggestions": dict,
+            "logs": List[EvolutionLog]
+        }
+        """
+        mem_list = list(memories)
+        actions = []
+        new_logs = []
+        abstractions = 0
+        contradictions_resolved = 0
+
+        print(f"[Evolution] 开始演化，记忆库 {len(mem_list)} 条...")
+
+        # ── Phase 1：发现重复簇并抽象 ─────────────────────────────────────
+        clusters = self.detector.find_duplicate_clusters(
+            mem_list, vector_store, similarity_threshold=0.75, min_cluster_size=3
+        )
+        print(f"[Evolution] 发现 {len(clusters)} 个重复簇")
+
+        for cluster in clusters[:5]:  # 每次最多处理 5 个簇
+            abstract_kwargs = self.abstractor.abstract_cluster(
+                cluster, llm_fn=self._llm_abstract_fn()
+            )
+            if abstract_kwargs and not dry_run and self._ingest_fn:
+                new_mem = self._ingest_fn(**abstract_kwargs)
+                log = EvolutionLog(
+                    operation="abstract",
+                    affected_ids=[m.id for m in cluster],
+                    result_id=new_mem.id,
+                    description=f"抽象 {len(cluster)} 条 → 《{abstract_kwargs['title']}》"
+                )
+                new_logs.append(log)
+                actions.append(f"✅ 抽象 {len(cluster)} 条 [{cluster[0].type}] → 《{abstract_kwargs['title']}》")
+                abstractions += 1
+            elif dry_run:
+                actions.append(f"[预览] 可抽象 {len(cluster)} 条 [{cluster[0].type}] 记忆")
+
+        # ── Phase 2：检测并处理矛盾 ────────────────────────────────────────
+        contradictions = self.detector.find_contradictions(mem_list)
+        print(f"[Evolution] 发现 {len(contradictions)} 对矛盾记忆")
+
+        for mem_a, mem_b, reason in contradictions[:5]:
+            if not dry_run:
+                result = self.resolver.resolve(mem_a, mem_b, reason, memory_fs)
+                # 存储矛盾说明到记忆
+                if self._ingest_fn:
+                    self._ingest_fn(
+                        content=result["reflection_content"],
+                        memory_type="reflection",
+                        title=f"[矛盾解决] {mem_a.title[:20]}",
+                        metadata={"tags": ["contradiction", "resolved"], "confidence": 0.8}
+                    )
+                log = EvolutionLog(
+                    operation="resolve",
+                    affected_ids=[mem_a.id, mem_b.id],
+                    result_id=None,
+                    description=f"解决矛盾：{reason}"
+                )
+                new_logs.append(log)
+                actions.append(f"🔧 解决矛盾：《{mem_a.title[:25]}》vs《{mem_b.title[:25]}》（{reason}）")
+                contradictions_resolved += 1
+            else:
+                actions.append(f"[预览] 矛盾：《{mem_a.title[:25]}》vs《{mem_b.title[:25]}》")
+
+        # ── Phase 3：策略优化建议 ──────────────────────────────────────────
+        suggestions = {}
+        if system_stats:
+            suggestions = self.optimizer.suggest_lifecycle_config(
+                memory_stats=system_stats.get("memory_fs", {}),
+                scheduler_stats=system_stats.get("scheduler", {}),
+                lifecycle_stats=system_stats.get("lifecycle", {}),
+            )
+            if suggestions.get("reasons"):
+                for r in suggestions["reasons"]:
+                    actions.append(f"💡 建议：{r}")
+
+        # ── 保存日志 ───────────────────────────────────────────────────────
+        with self._lock:
+            self.logs.extend(new_logs)
+
+        if not dry_run and self.storage_path:
+            self._save_logs()
+
+        summary = (
+            f"演化完成：抽象 {abstractions} 个簇，"
+            f"解决 {contradictions_resolved} 对矛盾，"
+            f"{'健康度: ' + suggestions.get('current_health', '') if suggestions else ''}"
+        )
+        print(f"[Evolution] {summary}")
+
+        return {
+            "summary": summary,
+            "actions": actions,
+            "abstractions": abstractions,
+            "contradictions_resolved": contradictions_resolved,
+            "suggestions": suggestions,
+            "logs": new_logs,
+            "dry_run": dry_run,
+        }
+
+    def get_stats(self) -> Dict[str, Any]:
+        by_op: Dict[str, int] = Counter(log.operation for log in self.logs)
+        return {
+            "total_operations": len(self.logs),
+            "by_operation": dict(by_op),
+            "recent_operations": [
+                {"op": l.operation, "desc": l.description, "time": l.timestamp.isoformat()}
+                for l in self.logs[-5:]
+            ]
+        }
+
+    def _llm_abstract_fn(self):
+        """返回 LLM 抽象函数（如果有注入）"""
+        if not self._llm_fn:
+            return None
+        llm = self._llm_fn
+
+        def fn(combined: str, count: int) -> Optional[str]:
+            try:
+                return llm(
+                    f"从以下 {count} 条记忆中提炼核心规律（150字以内）：\n{combined[:2000]}"
+                )
+            except Exception:
+                return None
+        return fn
+
+    def _save_logs(self):
+        if not self.storage_path:
+            return
+        path = self.storage_path / "evolution_logs.json"
+        tmp = path.with_suffix(".tmp")
+        data = [l.to_dict() for l in self.logs[-500:]]
+        with open(tmp, "w", encoding="utf-8") as f:
+            json.dump(data, f, ensure_ascii=False, indent=2, default=str)
+        tmp.replace(path)
+
+    def _load_logs(self):
+        path = self.storage_path / "evolution_logs.json"
+        if not path.exists():
+            return
+        try:
+            with open(path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            for d in data:
+                self.logs.append(EvolutionLog(
+                    operation=d["operation"],
+                    affected_ids=d["affected_ids"],
+                    result_id=d.get("result_id"),
+                    description=d["description"],
+                    timestamp=datetime.fromisoformat(d["timestamp"])
+                ))
+            print(f"[Evolution] 加载 {len(self.logs)} 条演化日志")
+        except Exception as e:
+            print(f"[Evolution] 日志加载失败: {e}")