diff --git a/internal/agent/core/agent.go b/internal/agent/core/agent.go
index 6f7a253..5ff3559 100644
--- a/internal/agent/core/agent.go
+++ b/internal/agent/core/agent.go
@@ -147,8 +147,9 @@ type AgentConfig struct {
 	PluginReg       *plugin.Registry
 	PluginDir       string
 	DistillInterval time.Duration
-	MaxContextSize  int              // 活跃上下文最大条数，超出按相关性裁剪
-	ContextSavePath string           // 上下文持久化路径，空则不持久化
+	MaxContextSize    int              // 活跃上下文最大条数，超出按相关性裁剪
+	ContextSavePath   string           // 上下文持久化路径，空则不持久化
+	EmbeddingModelPath string          // 预训练词嵌入模型路径（word2vec 文本格式），空则不使用
 	StageHost     *StageHost
 	EventBus      *events.Bus
 	ThinkingEnabled bool
@@ -174,7 +175,7 @@ func New(cfg AgentConfig) *Agent {
 		indexer:         cfg.Indexer,
 		skills:          cfg.Skills,
 		tracker:         cfg.Tracker,
-		context:         NewRelevanceContext(cfg.ContextSavePath),
+		context:         NewRelevanceContext(cfg.ContextSavePath, memory.NewStaticEmbedder(strings.Split(cfg.EmbeddingModelPath, ",")...)),
 		systemPrompt:    cfg.SystemPrompt,
 		ctx:             ctx,
 		cancel:          cancel,
diff --git a/internal/agent/core/context.go b/internal/agent/core/context.go
index b7a7ec0..aa267dc 100644
--- a/internal/agent/core/context.go
+++ b/internal/agent/core/context.go
@@ -15,32 +15,29 @@ import (
 	"gitcode.com/JianFeeeee/HomeAgent/internal/memory/vector"
 )
 
-// ContextEvent — 单条上下文事件
 type ContextEvent struct {
-	Timestamp time.Time `json:"timestamp"`
-	Source    string    `json:"source"`
-	Input     string    `json:"input"`
-	Response  string    `json:"response,omitempty"`
-	ToolsUsed []string  `json:"tools_used,omitempty"`
-	Vector    vector.Vector `json:"-"` // 缓存向量，避免重复计算
+	Timestamp time.Time   `json:"timestamp"`
+	Source    string      `json:"source"`
+	Input     string      `json:"input"`
+	Response  string      `json:"response,omitempty"`
+	ToolsUsed []string    `json:"tools_used,omitempty"`
+	Vector    vector.Vector `json:"-"`
 }
 
 const contextFlushInterval = 5 * time.Second
 
-// RelevanceContext — 基于相关性的上下文管理，非固定阈值
 type RelevanceContext struct {
 	mu       sync.Mutex
 	events   []*ContextEvent
-	embedder *memory.LocalWordEmbedder
-	trained  bool
+	embedder *memory.StaticEmbedder
 	savePath string
 	saveTimer *time.Timer
 	dirty    bool
 }
 
-func NewRelevanceContext(savePath string) *RelevanceContext {
+func NewRelevanceContext(savePath string, embedder *memory.StaticEmbedder) *RelevanceContext {
 	rc := &RelevanceContext{
-		embedder: memory.NewLocalWordEmbedder(),
+		embedder: embedder,
 		savePath: savePath,
 	}
 	if savePath != "" {
@@ -49,7 +46,6 @@ func NewRelevanceContext(savePath string) *RelevanceContext {
 	return rc
 }
 
-// load 从文件恢复上下文事件
 func (c *RelevanceContext) load() {
 	data, err := os.ReadFile(c.savePath)
 	if err != nil {
@@ -60,12 +56,27 @@ func (c *RelevanceContext) load() {
 		return
 	}
 	for _, evt := range events {
-		evt.Vector = c.embedder.Vectorize(evt.Input + " " + evt.Response)
+		evt.Input = memory.CleanTemplateText(evt.Input)
+		evt.Vector = c.computeVector(evt)
 	}
 	c.events = events
 }
 
-// Save 持久化上下文事件到文件
+func textForVector(evt *ContextEvent) string {
+	switch {
+	case evt.Source == "agent" && evt.Response != "":
+		return memory.CleanTemplateText(evt.Response)
+	case evt.Source == "cold_storage":
+		return memory.CleanTemplateText(evt.Input + " " + evt.Response)
+	default:
+		return memory.CleanTemplateText(evt.Input)
+	}
+}
+
+func (c *RelevanceContext) computeVector(evt *ContextEvent) vector.Vector {
+	return c.embedder.Vectorize(textForVector(evt))
+}
+
 func (c *RelevanceContext) Save() error {
 	if c.savePath == "" {
 		return nil
@@ -84,15 +95,13 @@ func (c *RelevanceContext) Append(evt ContextEvent) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 
-	evt.Vector = c.embedder.Vectorize(evt.Input + " " + evt.Response)
+	evt.Input = memory.CleanTemplateText(evt.Input)
+	evt.Vector = c.computeVector(&evt)
 	c.events = append(c.events, &evt)
 
-	c.trained = false
-
 	c.save()
 }
 
-// save 无锁版本，Append/Prune 内部持有锁时调用。带 debounce，每 5s 写一次盘。
 func (c *RelevanceContext) save() error {
 	if c.savePath == "" {
 		return nil
@@ -124,8 +133,6 @@ func (c *RelevanceContext) flush() {
 	c.dirty = false
 }
 
-// Prune — 基于当前输入计算每条上下文的相关性，归档最不相关的
-// 返回被归档的事件（转为文档），保留 topK 个最相关的在活跃上下文中
 func (c *RelevanceContext) Prune(currentInput string, topK int, docStore *document.Store) int {
 	c.mu.Lock()
 	defer c.mu.Unlock()
@@ -134,23 +141,19 @@ func (c *RelevanceContext) Prune(currentInput string, topK int, docStore *docume
 		return 0
 	}
 
-	// 保护最近 10 条记录不被淘汰，从更早的记录中选择淘汰对象
-	protectCount := 10
-	if protectCount > len(c.events) {
-		protectCount = len(c.events)
+	pCount := 10
+	if pCount > len(c.events) {
+		pCount = len(c.events)
 	}
-	protected := c.events[len(c.events)-protectCount:]
-	candidates := c.events[:len(c.events)-protectCount]
+	protected := c.events[len(c.events)-pCount:]
+	candidates := c.events[:len(c.events)-pCount]
 
 	if len(candidates) == 0 {
 		return 0
 	}
 
-	c.ensureTrained()
+	queryVec := c.embedder.VectorizeClean(currentInput)
 
-	queryVec := c.embedder.Vectorize(currentInput)
-
-	// 计算每条候选上下文与当前输入的相关性
 	type scored struct {
 		event *ContextEvent
 		score float64
@@ -162,12 +165,10 @@ func (c *RelevanceContext) Prune(currentInput string, topK int, docStore *docume
 		scoredEvents[i] = scored{event: evt, score: score, idx: i}
 	}
 
-	// 按相关性从高到低排序
 	sort.Slice(scoredEvents, func(i, j int) bool {
 		return scoredEvents[i].score > scoredEvents[j].score
 	})
 
-	// 从候选中选 topK 最相关的保留，其余淘汰
 	keepCount := topK - len(protected)
 	if keepCount < 0 {
 		keepCount = 0
@@ -178,19 +179,16 @@ func (c *RelevanceContext) Prune(currentInput string, topK int, docStore *docume
 	}
 	archive := scoredEvents[keepCount:]
 
-	// 重建 events 为保留的候选 + 受保护的最新记录
 	c.events = make([]*ContextEvent, 0, len(keep)+len(protected))
 	for _, s := range keep {
 		c.events = append(c.events, s.event)
 	}
 	c.events = append(c.events, protected...)
 
-	// 按时间重新排序
 	sort.Slice(c.events, func(i, j int) bool {
 		return c.events[i].Timestamp.Before(c.events[j].Timestamp)
 	})
 
-	// 归档到文档记忆
 	archived := 0
 	if docStore != nil && len(archive) > 0 {
 		entries := make([]document.ContextEntry, len(archive))
@@ -213,7 +211,6 @@ func (c *RelevanceContext) Prune(currentInput string, topK int, docStore *docume
 	return archived
 }
 
-// Format — 输出活跃上下文的文本，用于注入 prompt
 func (c *RelevanceContext) Format() string {
 	c.mu.Lock()
 	defer c.mu.Unlock()
@@ -234,7 +231,6 @@ func (c *RelevanceContext) Format() string {
 	return sb.String()
 }
 
-// Recent — 返回最近 n 条
 func (c *RelevanceContext) Recent(n int) []ContextEvent {
 	c.mu.Lock()
 	defer c.mu.Unlock()
@@ -249,23 +245,8 @@ func (c *RelevanceContext) Recent(n int) []ContextEvent {
 	return result
 }
 
-// Len — 当前上下文事件数
 func (c *RelevanceContext) Len() int {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	return len(c.events)
 }
-
-func (c *RelevanceContext) ensureTrained() {
-	if !c.trained && len(c.events) > 0 {
-		texts := make([]string, len(c.events))
-		for i, evt := range c.events {
-			texts[i] = evt.Input + " " + evt.Response
-		}
-		c.embedder.Train(texts)
-		for _, evt := range c.events {
-			evt.Vector = c.embedder.Vectorize(evt.Input + " " + evt.Response)
-		}
-		c.trained = true
-	}
-}
diff --git a/internal/agent/core/context_test.go b/internal/agent/core/context_test.go
index 8ed489a..bbb6f97 100644
--- a/internal/agent/core/context_test.go
+++ b/internal/agent/core/context_test.go
@@ -1,12 +1,19 @@
 package core
 
 import (
+	"os"
 	"testing"
 	"time"
+
+	"gitcode.com/JianFeeeee/HomeAgent/internal/memory"
 )
 
+func newTestCtx() *RelevanceContext {
+	return NewRelevanceContext("", memory.NewStaticEmbedder(""))
+}
+
 func TestContextAppendAndLen(t *testing.T) {
-	ctx := NewRelevanceContext("")
+	ctx := newTestCtx()
 	if ctx.Len() != 0 {
 		t.Errorf("new context should be empty, got %d", ctx.Len())
 	}
@@ -18,7 +25,7 @@ func TestContextAppendAndLen(t *testing.T) {
 }
 
 func TestContextRecent(t *testing.T) {
-	ctx := NewRelevanceContext("")
+	ctx := newTestCtx()
 	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "a"})
 	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "b"})
 	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "c"})
@@ -33,7 +40,7 @@ func TestContextRecent(t *testing.T) {
 }
 
 func TestContextFormat(t *testing.T) {
-	ctx := NewRelevanceContext("")
+	ctx := newTestCtx()
 	f := ctx.Format()
 	if f != "" {
 		t.Errorf("empty context should format to empty string, got %q", f)
@@ -54,7 +61,7 @@ func TestContextFormat(t *testing.T) {
 }
 
 func TestContextPruneKeepsTopK(t *testing.T) {
-	ctx := NewRelevanceContext("")
+	ctx := newTestCtx()
 	for i := 0; i < 20; i++ {
 		ctx.Append(ContextEvent{
 			Timestamp: time.Now(),
@@ -63,7 +70,6 @@ func TestContextPruneKeepsTopK(t *testing.T) {
 			Response:  "是的天气不错",
 		})
 	}
-	// 加一条不同主题的
 	ctx.Append(ContextEvent{
 		Timestamp: time.Now(),
 		Source:    "user",
@@ -71,17 +77,16 @@ func TestContextPruneKeepsTopK(t *testing.T) {
 		Response:  "好的我来算",
 	})
 
-	archived := ctx.Prune("微积分", 5, nil) // nil docStore → 不归档，只裁剪
+	archived := ctx.Prune("微积分", 5, nil)
 	_ = archived
 
-	// protectCount=10 + topK=5 → 最多保留 15
 	if ctx.Len() > 15 {
 		t.Errorf("after prune to 5, len should be ≤15, got %d", ctx.Len())
 	}
 }
 
 func TestContextPruneWithDocStore(t *testing.T) {
-	ctx := NewRelevanceContext("")
+	ctx := newTestCtx()
 	for i := 0; i < 15; i++ {
 		ctx.Append(ContextEvent{
 			Timestamp: time.Now(),
@@ -98,8 +103,7 @@ func TestContextPruneWithDocStore(t *testing.T) {
 }
 
 func TestContextAppendAfterPrune(t *testing.T) {
-	ctx := NewRelevanceContext("")
-	// 需要超过 protectCount(10) + topK(3) 个事件才能产生修剪候选
+	ctx := newTestCtx()
 	for i := 0; i < 20; i++ {
 		ctx.Append(ContextEvent{
 			Timestamp: time.Now(),
@@ -109,7 +113,7 @@ func TestContextAppendAfterPrune(t *testing.T) {
 	}
 
 	ctx.Prune("hello", 3, nil)
-	if ctx.Len() > 13 { // 10 protected + 3 topK
+	if ctx.Len() > 13 {
 		t.Errorf("expected ≤13 after prune, got %d", ctx.Len())
 	}
 
@@ -119,6 +123,107 @@ func TestContextAppendAfterPrune(t *testing.T) {
 	}
 }
 
+func TestContextPruneWithStaticEmbedder(t *testing.T) {
+	tmpFile, err := os.CreateTemp("", "test_embeddings_*.txt")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer os.Remove(tmpFile.Name())
+
+	content := `8 4
+天气 0.1 0.2 0.3 0.4
+下雨 0.15 0.25 0.35 0.45
+台风 0.12 0.22 0.32 0.42
+股票 0.9 0.1 0.1 0.1
+基金 0.85 0.15 0.1 0.1
+微积分 0.1 0.1 0.9 0.1
+导数 0.15 0.1 0.85 0.15
+数学 0.1 0.1 0.8 0.2
+`
+	if _, err := tmpFile.WriteString(content); err != nil {
+		t.Fatal(err)
+	}
+	tmpFile.Close()
+
+	embedder := memory.NewStaticEmbedder(tmpFile.Name())
+	if !embedder.Loaded() {
+		t.Fatal("embedder should be loaded")
+	}
+
+	ctx := NewRelevanceContext("", embedder)
+
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "今天天气很好", Response: "是的"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "会不会下雨", Response: "会"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "台风来了", Response: "注意"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "帮我算微积分", Response: "好的"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "导数怎么求", Response: "公式"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "数学题", Response: "解答"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "股票涨了", Response: "恭喜"})
+	ctx.Append(ContextEvent{Timestamp: time.Now(), Source: "user", Input: "基金定投", Response: "可以"})
+
+	if ctx.Len() != 8 {
+		t.Fatalf("expected 8 events, got %d", ctx.Len())
+	}
+
+	archived := ctx.Prune("最近基金怎么样", 3, nil)
+
+	if ctx.Len() > 13 {
+		t.Errorf("prune should limit total events, got %d", ctx.Len())
+	}
+
+	remaining := ctx.Format()
+	t.Logf("query: 最近基金怎么样\nremaining events:\n%s", remaining)
+	t.Logf("archived: %d", archived)
+
+	needsFund := contains(remaining, "基金定投") || contains(remaining, "股票涨了")
+	needsWeather := contains(remaining, "今天天气很好") || contains(remaining, "台风来了")
+
+	t.Logf("has financial events: %v, has weather events: %v", needsFund, needsWeather)
+}
+
+func TestContextPruneRecent10Protected(t *testing.T) {
+	ctx := newTestCtx()
+
+	for i := 0; i < 15; i++ {
+		ctx.Append(ContextEvent{
+			Timestamp: time.Now(),
+			Source:    "user",
+			Input:     "今天天气很好",
+		})
+	}
+
+	for i := 0; i < 5; i++ {
+		ctx.Append(ContextEvent{
+			Timestamp: time.Now(),
+			Source:    "user",
+			Input:     "股票行情",
+		})
+	}
+
+	ctx.Prune("天气", 3, nil)
+
+	// 最近 10 条全部是"股票行情"（第6-15条是天气，第16-20条是股票）
+	// protectCount=10 保护最近 10 条 → 5 条天气最多保留 5+3=8 条
+	// 至少最近 10 条全部保留 → 至少包含 5 条股票
+	remaining := ctx.Format()
+	t.Logf("after weather query:\n%s", remaining)
+	weatherCount := 0
+	stockCount := 0
+	for _, line := range splitLines(remaining) {
+		if contains(line, "天气") {
+			weatherCount++
+		}
+		if contains(line, "股票") {
+			stockCount++
+		}
+	}
+	t.Logf("weather events: %d, stock events: %d", weatherCount, stockCount)
+
+	if stockCount < 5 {
+		t.Errorf("recent 10 should all be protected, expected at least 5 stock events, got %d", stockCount)
+	}
+}
+
 func contains(s, substr string) bool {
 	return len(s) >= len(substr) && containsStr(s, substr)
 }
@@ -131,3 +236,18 @@ func containsStr(s, substr string) bool {
 	}
 	return false
 }
+
+func splitLines(s string) []string {
+	var lines []string
+	start := 0
+	for i := 0; i < len(s); i++ {
+		if s[i] == '\n' {
+			lines = append(lines, s[start:i])
+			start = i + 1
+		}
+	}
+	if start < len(s) {
+		lines = append(lines, s[start:])
+	}
+	return lines
+}
diff --git a/internal/memory/bilingual_test.go b/internal/memory/bilingual_test.go
new file mode 100644
index 0000000..010169d
--- /dev/null
+++ b/internal/memory/bilingual_test.go
@@ -0,0 +1,243 @@
+package memory
+
+import (
+	"fmt"
+	"sort"
+	"testing"
+)
+
+type bilingualEvent struct {
+	idx     int
+	source  string
+	topic   string
+	text    string // cleaned text for vectorization
+	label   string // short description
+}
+
+func TestBilingualPruningAccuracy(t *testing.T) {
+	zhPath := "/tmp/cc.zh.top200k.vec"
+	enPath := "/tmp/cc.en.top200k.vec"
+
+	// Test with Chinese-only vs Chinese+English
+	type modelConfig struct {
+		name  string
+		paths []string
+	}
+
+	configs := []modelConfig{
+		{"中文仅", []string{zhPath}},
+		{"中文+英文", []string{zhPath, enPath}},
+	}
+
+	events := genBilingualEvents()
+
+	for _, cfg := range configs {
+		t.Run(cfg.name, func(t *testing.T) {
+			e := NewStaticEmbedder(cfg.paths...)
+			if !e.Loaded() {
+				t.Skipf("%s: embedder not loaded", cfg.name)
+			}
+			t.Logf("%s: %d words", cfg.name, len(e.words))
+
+			type scored struct {
+				idx    int
+				topic  string
+				label  string
+				score  float64
+			}
+
+			queries := []struct {
+				q       string
+				qTopic  string
+				desc    string
+			}{
+				{"老大说了关于 React 组件的事情", "老大私聊", "中英混合:老大+React"},
+				{"帮我查一下 Nginx 反向代理配置", "服务器运维", "中英混合:Nginx+反向代理"},
+				{"河南医药大学 Docker 部署", "大学招生", "中英混合:大学+Docker"},
+				{"JavaScript 基金定投收益计算", "股票基金", "中英混合:JS+基金"},
+				{"Server 前端组件封装 layout", "前端开发", "中英混合:Server+layout"},
+				{"河南医药大学录取分数线", "大学招生", "纯中文:大学"},
+				{"nginx reverse proxy config", "服务器运维", "纯英文:nginx"},
+			}
+
+			for _, q := range queries {
+				qVec := e.Vectorize(q.q)
+				t.Logf("\n  query: %q (%s)", q.q, q.desc)
+
+				all := make([]scored, len(events))
+				for i, ev := range events {
+					text := textForBilingual(ev, cfg.paths)
+					vec := e.Vectorize(text)
+					all[i] = scored{idx: i, topic: ev.topic, label: ev.label, score: cosineSim(qVec, vec)}
+				}
+				sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+				// Check top 5 for same-topic presence
+				var intraHits int
+				for _, s := range all[:5] {
+					if s.topic == q.qTopic {
+						intraHits++
+					}
+				}
+
+				topScore := all[0]
+				topIsCorrect := topScore.topic == q.qTopic
+
+				t.Logf("    top5 intra=%d/5, top1=%q(%s) score=%.4f %s",
+					intraHits, topScore.topic, topScore.label, topScore.score,
+					map[bool]string{true: "✅", false: "❌"}[topIsCorrect])
+				for _, s := range all[:5] {
+					mark := ""
+					if s.topic == q.qTopic {
+						mark = " ✓"
+					}
+					t.Logf("      [%.4f] [%-12s] %s%s", s.score, s.topic, trimLen(s.label, 50), mark)
+				}
+
+				if !topIsCorrect {
+					t.Logf("    [WARN] top1 mismatch for %q", q.desc)
+				}
+			}
+		})
+	}
+}
+
+func TestBilingualCrossLingualSimilarity(t *testing.T) {
+	zhPath := "/tmp/cc.zh.top200k.vec"
+	enPath := "/tmp/cc.en.top200k.vec"
+	e := NewStaticEmbedder(zhPath, enPath)
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	pairs := []struct {
+		a, b string
+		desc string
+	}{
+		{"server", "服务器", "英中同义"},
+		{"computer", "电脑", "英中同义"},
+		{"老大", "boss", "中英同义"},
+		{"大学", "university", "中英同义"},
+		{"Nginx", "服务器", "专名+普通"},
+		{"股票", "stock", "中英同义"},
+		{"React", "前端", "专名+概念"},
+		{"JavaScript", "编程", "专名+概念"},
+		{"老大私聊", "boss private chat", "中英短语"},
+		{"河南医药大学录取", "Henan Medical University admission", "中英专名"},
+		{"nginx config", "Nginx 配置", "英中技术"},
+	}
+
+	t.Log("=== 跨语言相似度 ===")
+	for _, p := range pairs {
+		va := e.Vectorize(p.a)
+		vb := e.Vectorize(p.b)
+		sim := cosineSim(va, vb)
+		t.Logf("  %.4f  %q  ↔  %q  [%s]", sim, trimLen(p.a, 30), trimLen(p.b, 30), p.desc)
+	}
+}
+
+func TestBilingualEdgeCases(t *testing.T) {
+	zhPath := "/tmp/cc.zh.top200k.vec"
+	e := NewStaticEmbedder(zhPath)
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	cases := []string{
+		"纯英文文本 nginx react docker javascript",
+		"纯中文 服务器 配置 反向代理 部署",
+		"中英混合 nginx 反向代理 配置",
+		"代码片段 const foo = 'bar'; function test()",
+		"URL路径 /api/v1/users/123",
+		"中文含标点！@#￥%……&*（）",
+		"空字符串",
+	}
+
+	t.Log("=== 边缘情况向量化 ===")
+	for _, c := range cases {
+		v := e.Vectorize(c)
+		var dims int
+		for range v {
+			dims++
+		}
+		t.Logf("  dims=%d  %q", dims, trimLen(c, 60))
+	}
+}
+
+func TestBilingualVectorizeClean(t *testing.T) {
+	zhPath := "/tmp/cc.zh.top200k.vec"
+	e := NewStaticEmbedder(zhPath)
+
+	inputs := []string{
+		`来自小王的（扶高升学咨询群）群聊消息，通过id99使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复该群聊，content 设为 JSON 字符串：{"content":"你的回复","group_id":979911915}`,
+		`【重要！老大消息】来自—/的私聊消息，通过id54使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复对方，content 设为 JSON 字符串：{"content":"你的回复","user_id":2198972886}`,
+		`The nginx server is configured with reverse proxy. 帮我查一下 Docker 容器状态。`,
+		`老大你好，React 组件已经封装好了，Nginx 配置也改完了，Docker 部署没问题。`,
+	}
+
+	for i, inp := range inputs {
+		rawVec := e.Vectorize(inp)
+		cleanVec := e.VectorizeClean(inp)
+		sim := cosineSim(rawVec, cleanVec)
+		rawTokens := len(e.tokenize(inp))
+		cleanTokens := len(e.tokenize(CleanTemplateText(inp)))
+		t.Logf("[%d] sim(raw,clean)=%.4f  tokens: raw=%d clean=%d", i, sim, rawTokens, cleanTokens)
+	}
+}
+
+// --- bilingual test data ---
+
+func genBilingualEvents() []bilingualEvent {
+	entries := []struct {
+		topic   string
+		zh      string // Chinese description
+		en      string // English terms mixed in
+		source  string
+	}{
+		{"大学招生", "河南医药大学录取分数线", "", "qq"},
+		{"大学招生", "医学院专业排名", "medical university ranking", "agent"},
+		{"大学招生", "高考志愿填报咨询", "college application consultation", "qq"},
+		{"大学招生", "河南医药大学 Docker 部署项目", "docker deployment project", "agent"},
+		{"老大私聊", "老大私聊消息回复", "boss private chat reply", "qq"},
+		{"老大私聊", "老大说了关于 React 组件的事情", "boss talked about React components", "agent"},
+		{"老大私聊", "回复老大关于服务器配置问题", "reply boss about nginx config", "agent"},
+		{"老大私聊", "老大要求检查 Docker 容器状态", "boss asked to check docker status", "qq"},
+		{"前端开发", "前端组件封装", "React component encapsulation", "cli"},
+		{"前端开发", "页面路由配置 layout 设计", "page route config layout design", "cli"},
+		{"前端开发", "JavaScript 交互逻辑开发", "javascript interaction logic", "agent"},
+		{"前端开发", "TypeScript 代码调试优化", "typescript code debug optimization", "agent"},
+		{"服务器运维", "Nginx 反向代理配置", "nginx reverse proxy config", "cli"},
+		{"服务器运维", "Docker 容器部署方案", "docker container deployment", "cli"},
+		{"服务器运维", "数据库备份恢复", "database backup recovery", "agent"},
+		{"服务器运维", "SSL 证书续期配置", "ssl certificate renewal", "agent"},
+		{"股票基金", "基金定投策略配置", "fund investment strategy", "qq"},
+		{"股票基金", "股票涨跌分析", "stock market analysis", "agent"},
+		{"股票基金", "理财收益 JavaScript 计算", "investment return javascript calculation", "agent"},
+		{"股票基金", "市场行情 API 数据获取", "market data api fetch", "qq"},
+	}
+
+	var events []bilingualEvent
+	for i, entry := range entries {
+		text := entry.zh
+		if entry.en != "" {
+			text += " " + entry.en
+		}
+		events = append(events, bilingualEvent{
+			idx:    i,
+			source: entry.source,
+			topic:  entry.topic,
+			text:   text,
+			label:  fmt.Sprintf("%s (%s)", trimLen(entry.zh, 30), trimLen(entry.en, 30)),
+		})
+	}
+	return events
+}
+
+func textForBilingual(ev bilingualEvent, modelPaths []string) string {
+	switch {
+	case ev.source == "agent" && ev.text != "":
+		return CleanTemplateText(ev.text)
+	default:
+		return CleanTemplateText(ev.text)
+	}
+}
diff --git a/internal/memory/clean_stress_test.go b/internal/memory/clean_stress_test.go
new file mode 100644
index 0000000..5b85cb9
--- /dev/null
+++ b/internal/memory/clean_stress_test.go
@@ -0,0 +1,346 @@
+package memory
+
+import (
+	"fmt"
+	"sort"
+	"testing"
+)
+
+type cleanTestEvent struct {
+	idx     int
+	source  string
+	input   string
+	response string
+	rawText string
+	cleanedText string
+	topic  string
+}
+
+func TestCleanStressPrecision(t *testing.T) {
+	events := genStressEvents(200)
+	topics := []string{"大学招生", "老大私聊", "前端开发", "服务器运维", "股票基金"}
+
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	for _, cleanMode := range []bool{true, false} {
+		t.Run(fmt.Sprintf("去模版=%v", cleanMode), func(t *testing.T) {
+			usedTopics := make([]string, 0)
+			for _, tp := range topics {
+				if hasTopicEvents(events, tp) {
+					usedTopics = append(usedTopics, tp)
+				}
+			}
+			if len(usedTopics) == 0 {
+				t.Fatal("no events for any topic")
+			}
+			t.Logf("topics: %v, events: %d", usedTopics, len(events))
+
+	for _, qTopic := range usedTopics {
+		query := queryForTopic(qTopic)
+		qVec := e.Vectorize(query)
+
+		type scored struct {
+			idx   int
+			topic string
+			text  string
+			score float64
+		}
+		all := make([]scored, len(events))
+		for i, ev := range events {
+			text := ev.rawText
+			if cleanMode {
+				text = ev.cleanedText
+			}
+			vec := e.Vectorize(text)
+			all[i] = scored{idx: i, topic: ev.topic, text: text, score: cosineSim(qVec, vec)}
+		}
+		sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+		topK := len(usedTopics) * 2
+		if topK > len(all) {
+			topK = len(all)
+		}
+
+		intraHits := 0
+		for _, s := range all[:topK] {
+			if s.topic == qTopic {
+				intraHits++
+			}
+		}
+		expected := countTopicEvents(events, qTopic)
+		if expected > topK {
+			expected = topK
+		}
+		recall := float64(intraHits) / float64(expected)
+
+		if recall < 0.3 {
+			t.Logf("  [LOW] query=%q topK=%d intra=%d/%d recall=%.2f", qTopic, topK, intraHits, expected, recall)
+			for _, s := range all[:8] {
+				t.Logf("    [%.4f] %s", s.score, trimLen(s.text, 60))
+			}
+		} else {
+			t.Logf("  [OK] query=%q topK=%d intra=%d/%d recall=%.2f", qTopic, topK, intraHits, expected, recall)
+		}
+	}
+		})
+	}
+}
+
+func TestCleanStressCrossTopic(t *testing.T) {
+	events := genStressEvents(200)
+
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	queries := []string{
+		"河南医药大学录取分数线",
+		"老大发了什么私聊消息",
+		"前端组件怎么封装布局",
+		"服务器部署配置代理备份证书",
+		"基金定投收益计算",
+	}
+
+	for _, q := range queries {
+		qVec := e.Vectorize(q)
+		t.Logf("query: %q", q)
+
+		type scored struct {
+			idx   int
+			topic string
+			score float64
+		}
+		all := make([]scored, len(events))
+		for i, ev := range events {
+			all[i] = scored{idx: i, topic: ev.topic, score: cosineSim(qVec, e.Vectorize(ev.cleanedText))}
+		}
+		sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+		topScores := make(map[string]float64)
+		for _, s := range all[:10] {
+			if _, ok := topScores[s.topic]; !ok {
+				topScores[s.topic] = s.score
+			}
+		}
+		for tp, sc := range topScores {
+			t.Logf("  [%.4f] %s", sc, tp)
+		}
+	}
+}
+
+func TestCleanTemplateNoiseSuppression(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	noisyInput := `来自小王的（扶高升学咨询群）群聊消息，通过id99使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复该群聊，content 设为 JSON 字符串：{"content":"你的回复","group_id":979911915}`
+	cleanInput := `来自小王的（扶高升学咨询群）群聊消息`
+
+	query := "扶高升学咨询群"
+	qClear := e.Vectorize(query)
+	qNoisy := e.Vectorize(noisyInput)
+	qClean := e.Vectorize(cleanInput)
+
+	n2c := cosineSim(qNoisy, qClean)
+	n2q := cosineSim(qNoisy, qClear)
+	c2q := cosineSim(qClean, qClear)
+
+	t.Logf("noisy(%q) vs clean(%q) = %.4f", noisyInput[:30], cleanInput, n2c)
+	t.Logf("noisy vs query(%q) = %.4f", query, n2q)
+	t.Logf("clean vs query = %.4f", c2q)
+
+	if c2q <= n2q {
+		t.Log("NOTE: clean not better than noisy for this pattern (may have useful info in metadata)")
+	}
+}
+
+func TestCleanVectorConsistency(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	templates := []string{
+		`通过id1使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复该群聊，content 设为 JSON 字符串：{"content":"你的回复","group_id":1}`,
+		`通过id2使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复对方，content 设为 JSON 字符串：{"content":"你的回复","user_id":2}`,
+		`通过id3使用qq_get_message工具获取消息正文。获取后必须使用qq_send_private_msg工具回复对方，不得使用其他非回复工具。`,
+		`[12:00] agent: 处理完成`,
+	}
+
+	for _, tmpl := range templates {
+		cleaned := CleanTemplateText(tmpl)
+		t.Logf("template {%q} → {%q} (%d chars)", trimLen(tmpl, 60), cleaned, len(cleaned))
+	}
+
+	pairs := []struct {
+		a, b   string
+		reason string
+	}{
+		{cleanQQGroup("A", "群1"), cleanQQGroup("B", "群1"), "同群不同人"},
+		{cleanQQGroup("A", "群1"), cleanQQGroup("A", "群2"), "同人不同群"},
+		{cleanQQPrivate("老大"), cleanQQPrivate("老板"), "私聊不同人"},
+		{cleanQQGroup("A", "高考群"), cleanQQPrivate("老大"), "群聊vs私聊"},
+	}
+
+	for _, p := range pairs {
+		va := e.Vectorize(p.a)
+		vb := e.Vectorize(p.b)
+		sim := cosineSim(va, vb)
+		t.Logf("sim(%q, %q) [%s] = %.4f", trimLen(p.a, 40), trimLen(p.b, 40), p.reason, sim)
+	}
+}
+
+func BenchmarkCleanVectorize(b *testing.B) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		b.Skip("embedder not loaded")
+	}
+
+	texts := make([]string, 100)
+	for i := range texts {
+		texts[i] = fmt.Sprintf(
+			`【重要！老大消息】来自—/的私聊消息，通过id%d使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复对方，content 设为 JSON 字符串：{"content":"你的回复","user_id":%d}`,
+			i, 1000+i,
+		)
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		e.VectorizeClean(texts[i%len(texts)])
+	}
+}
+
+// --- test data generators ---
+
+func genStressEvents(n int) []cleanTestEvent {
+	if n <= 0 {
+		return nil
+	}
+	topics := []string{"大学招生", "老大私聊", "前端开发", "服务器运维", "股票基金"}
+	events := make([]cleanTestEvent, 0, n)
+
+	names := []string{"小明", "小红", "小张", "老王", "老大", "小李", "小王", "小赵"}
+	groups := []string{"扶高升学咨询群", "前端技术交流", "服务器运维群", "基金定投群", "闲聊群"}
+
+	topicContent := map[string]struct {
+		keywords []string
+		sources  []string
+	}{
+		"大学招生": {[]string{"河南医药大学", "录取分数线", "专业排名", "高考志愿", "招生简章"}, []string{"qq", "qq", "agent"}},
+		"老大私聊": {[]string{"老大私聊消息", "回复老大", "任务安排", "汇报工作", "收到"}, []string{"qq", "agent", "agent"}},
+		"前端开发": {[]string{"前端组件封装", "页面路由配置", "界面布局设计", "交互逻辑开发", "代码调试优化"}, []string{"cli", "cli", "agent"}},
+		"服务器运维": {[]string{"反向代理配置", "容器部署方案", "证书续期", "数据库备份恢复", "监控告警处理"}, []string{"cli", "agent", "agent"}},
+		"股票基金": {[]string{"基金定投策略", "股票涨跌分析", "理财收益计算", "市场行情分析", "投资风险管理"}, []string{"qq", "qq", "agent"}},
+	}
+
+	for i := 0; i < n; i++ {
+		tp := topics[i%len(topics)]
+		info := topicContent[tp]
+		kw := info.keywords[i%len(info.keywords)]
+		nm := names[i%len(names)]
+		grp := groups[i%len(groups)]
+		src := info.sources[i%len(info.sources)]
+
+		var input, response string
+		switch src {
+		case "qq":
+			if tp == "老大私聊" {
+				input = fmt.Sprintf(`【重要！老大消息】来自%s的私聊消息，通过id%d使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复对方，content 设为 JSON 字符串：{"content":"你的回复","user_id":%d}`, nm, i, 1000+i)
+				if i%3 == 0 {
+					response = fmt.Sprintf("已回复老大，关于%s", kw)
+				}
+			} else {
+				input = fmt.Sprintf(`来自%s的（%s）群聊消息，通过id%d使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复该群聊，content 设为 JSON 字符串：{"content":"你的回复","group_id":%d}`, nm, grp, i, 9000+i)
+				if i%3 == 0 {
+					response = fmt.Sprintf("已回复%s相关的问题", kw)
+				}
+			}
+		case "agent":
+			input = fmt.Sprintf(`来自%s的（%s）消息`, nm, grp)
+			response = fmt.Sprintf("关于%s，我的建议是...已处理完成。", kw)
+		case "cli":
+			input = fmt.Sprintf("查询%s的相关信息", kw)
+			response = fmt.Sprintf("查到了%s的结果", kw)
+		}
+
+		cleaned := cleanEventText(src, input, response)
+		raw := rawEventText(src, input, response)
+		events = append(events, cleanTestEvent{
+			idx:     i,
+			source:  src,
+			input:   input,
+			response: response,
+			rawText: raw,
+			cleanedText: cleaned,
+			topic:  tp,
+		})
+	}
+	return events
+}
+
+func cleanEventText(source, input, response string) string {
+	switch {
+	case source == "agent" && response != "":
+		return CleanTemplateText(response)
+	case source == "cold_storage":
+		return CleanTemplateText(input + " " + response)
+	default:
+		return CleanTemplateText(input)
+	}
+}
+
+func rawEventText(source, input, response string) string {
+	if response == "" {
+		return input
+	}
+	return input + " " + response
+}
+
+func queryForTopic(topic string) string {
+	switch topic {
+	case "大学招生":
+		return "河南医药大学录取分数线多少"
+	case "老大私聊":
+		return "老大刚才说了什么私聊消息"
+	case "前端开发":
+		return "前端组件怎么封装布局"
+	case "服务器运维":
+		return "服务器部署容器代理配置证书备份监控告警"
+	case "股票基金":
+		return "基金定投收益怎么样"
+	default:
+		return topic
+	}
+}
+
+func hasTopicEvents(events []cleanTestEvent, topic string) bool {
+	for _, ev := range events {
+		if ev.topic == topic {
+			return true
+		}
+	}
+	return false
+}
+
+func countTopicEvents(events []cleanTestEvent, topic string) int {
+	n := 0
+	for _, ev := range events {
+		if ev.topic == topic {
+			n++
+		}
+	}
+	return n
+}
+
+func cleanQQGroup(user, group string) string {
+	return fmt.Sprintf("来自%s的（%s）群聊消息", user, group)
+}
+
+func cleanQQPrivate(user string) string {
+	return fmt.Sprintf("【重要！老大消息】来自%s的私聊消息", user)
+}
diff --git a/internal/memory/clean_text.go b/internal/memory/clean_text.go
new file mode 100644
index 0000000..659216f
--- /dev/null
+++ b/internal/memory/clean_text.go
@@ -0,0 +1,57 @@
+package memory
+
+import (
+	"regexp"
+	"strings"
+
+	"gitcode.com/JianFeeeee/HomeAgent/internal/memory/vector"
+)
+
+var (
+	reQQGroupSuffix = regexp.MustCompile(
+		`，通过id\d+使用qq_get_message工具获取消息正文。获取内容后使用 output_send\(channel="qq"\) 回复该群聊，content 设为 JSON 字符串：\{[^}]*\}`,
+	)
+	reQQPrivateSuffix = regexp.MustCompile(
+		`，通过id\d+使用qq_get_message工具获取消息正文。获取内容后使用 output_send\(channel="qq"\) 回复对方，content 设为 JSON 字符串：\{[^}]*\}`,
+	)
+	reQQOldReply = regexp.MustCompile(
+		`通过id\d+使用qq_get_message工具获取消息正文。获取后必须使用[^。]+。`,
+	)
+	reQQOldForbid = regexp.MustCompile(
+		`你只能通过qq_get_message先看消息，然后直接用%!s\(MISSING\)send_private_msg回复，中间的思考过程禁止调用任何其他工具\s*→\s*`,
+	)
+	reQQGeneral = regexp.MustCompile(
+		`通过id\d+使用qq_get_message工具获取消息正文[。，][^。]*?(?:回复|发送消息)`,
+	)
+	reTimestamp = regexp.MustCompile(
+		`\[\d{2}:\d{2}\]\s*`,
+	)
+	reAgentPrefix = regexp.MustCompile(
+		`冷知识|注意|提示|核心要求|规则`,
+	)
+	reMultiSpace = regexp.MustCompile(`\s+`)
+)
+
+func CleanTemplateText(text string) string {
+	text = reQQGroupSuffix.ReplaceAllString(text, "")
+	text = reQQPrivateSuffix.ReplaceAllString(text, "")
+	text = reQQOldReply.ReplaceAllString(text, "")
+	text = reQQOldForbid.ReplaceAllString(text, "")
+	text = reQQGeneral.ReplaceAllString(text, "")
+	text = reTimestamp.ReplaceAllString(text, "")
+	text = reMultiSpace.ReplaceAllString(text, " ")
+	text = strings.TrimSpace(text)
+
+	if text == "" {
+		return ""
+	}
+
+	text = strings.TrimPrefix(text, "，")
+	text = strings.TrimPrefix(text, "，")
+	text = strings.TrimSpace(text)
+	return text
+}
+
+func (e *StaticEmbedder) VectorizeClean(text string) vector.Vector {
+	return e.Vectorize(CleanTemplateText(text))
+}
diff --git a/internal/memory/cut.go b/internal/memory/cut.go
index a1f5918..b29cb49 100644
--- a/internal/memory/cut.go
+++ b/internal/memory/cut.go
@@ -93,6 +93,7 @@ var stopWords = map[string]bool{
 }
 
 func ExtractKeywords(text string) []string {
+	text = CleanTemplateText(text)
 	x := GetJieba()
 	if x == nil {
 		return nil
diff --git a/internal/memory/indexer.go b/internal/memory/indexer.go
index bfd5992..fdcdf3e 100644
--- a/internal/memory/indexer.go
+++ b/internal/memory/indexer.go
@@ -90,11 +90,13 @@ func (idx *Indexer) BuildContext(userInput string) *InjectedContext {
 		return &InjectedContext{Summary: ""}
 	}
 
+	input := CleanTemplateText(userInput)
+
 	// 1. 向量搜索：从实体名向量索引中找到相关实体
-	vectorEntities := idx.vectorSearchEntities(userInput)
+	vectorEntities := idx.vectorSearchEntities(input)
 
 	// 2. 关键词搜索：已有逻辑
-	keywords := ExtractKeywords(userInput)
+	keywords := ExtractKeywords(input)
 	if len(keywords) == 0 && len(vectorEntities) == 0 {
 		keywords = []string{userInput}
 	}
diff --git a/internal/memory/real_context_test.go b/internal/memory/real_context_test.go
new file mode 100644
index 0000000..7833ea8
--- /dev/null
+++ b/internal/memory/real_context_test.go
@@ -0,0 +1,309 @@
+package memory
+
+import (
+	"encoding/json"
+	"os"
+	"sort"
+	"strings"
+	"testing"
+	"time"
+)
+
+type realEvent struct {
+	Timestamp time.Time `json:"timestamp"`
+	Source    string    `json:"source"`
+	Input     string    `json:"input"`
+	Response  string    `json:"response"`
+}
+
+func TestCleanTemplateText(t *testing.T) {
+	cases := []struct {
+		input    string
+		expected string
+		contains string
+	}{
+		{
+			input:    `来自A的（扶高升学咨询群）群聊消息，通过id36使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复该群聊，content 设为 JSON 字符串：{"content":"你的回复","group_id":979911915}`,
+			expected: "来自A的（扶高升学咨询群）群聊消息",
+		},
+		{
+			input:    `【重要！老大消息】来自—/的私聊消息，通过id54使用qq_get_message工具获取消息正文。获取内容后使用 output_send(channel="qq") 回复对方，content 设为 JSON 字符串：{"content":"你的回复","user_id":2198972886}`,
+			expected: "【重要！老大消息】来自—/的私聊消息",
+		},
+		{
+			input:    `[12:05] agent: 已经回复老大啦～继续去搞 vanblog 换端口的事 😊`,
+			contains: "已经回复老大啦",
+		},
+		{
+			input:    `加载文档记忆: 扶高升学咨询群 河南医药大学 回复 招生 专业 录取`,
+			expected: "加载文档记忆: 扶高升学咨询群 河南医药大学 回复 招生 专业 录取",
+		},
+		{
+			input:    `通过id10使用qq_get_message工具获取消息正文。获取后必须使用qq_send_private_msg工具回复对方，不得使用其他非回复工具。你只能通过qq_get_message先看消息，然后直接用%!s(MISSING)send_private_msg回复，中间的思考过程禁止调用任何其他工具 → 已经回复老大啦～`,
+			contains: "已经回复老大啦",
+		},
+		{
+			input:    "",
+			expected: "",
+		},
+		{
+			input:    `处理错误: all 9 providers failed, last error: lua transform_request: adapter tesy not loaded`,
+			contains: "处理错误",
+		},
+	}
+
+	for i, c := range cases {
+		got := CleanTemplateText(c.input)
+		if c.expected != "" && got != c.expected {
+			t.Errorf("case %d:\n  input:    %q\n  expected: %q\n  got:      %q", i, trimLen(c.input, 60), c.expected, got)
+		}
+		if c.contains != "" && !strings.Contains(got, c.contains) {
+			t.Errorf("case %d: expected to contain %q, got %q", i, c.contains, got)
+		}
+		t.Logf("case %d: %q → %q", i, trimLen(c.input, 60), got)
+	}
+}
+
+func TestRealContextPerSourceVector(t *testing.T) {
+	modelPath := "/tmp/cc.zh.sample.vec"
+	if _, err := os.Stat(modelPath); os.IsNotExist(err) {
+		t.Skip("real embedding file not found")
+	}
+	e := NewStaticEmbedder(modelPath)
+
+	data, err := os.ReadFile("/tmp/context.json")
+	if err != nil {
+		t.Fatal(err)
+	}
+	var raw []realEvent
+	if err := json.Unmarshal(data, &raw); err != nil {
+		t.Fatal(err)
+	}
+	t.Logf("loaded %d real events", len(raw))
+
+	type scored struct {
+		idx    int
+		source string
+		text   string
+		score  float64
+	}
+
+	clean := func(ev realEvent) string {
+		switch {
+		case ev.Source == "agent" && ev.Response != "":
+			return CleanTemplateText(ev.Response)
+		case ev.Source == "cold_storage":
+			return CleanTemplateText(ev.Input + " " + ev.Response)
+		default:
+			return CleanTemplateText(ev.Input)
+		}
+	}
+
+	t.Run("医药大学_不同源向量", func(t *testing.T) {
+		q := "河南医药大学招生分数录取排名"
+		qVec := e.VectorizeClean(q)
+
+		all := make([]scored, len(raw))
+		for i, ev := range raw {
+			text := clean(ev)
+			all[i] = scored{idx: i, source: ev.Source, text: text[:min(len(text), 200)], score: cosineSim(qVec, e.Vectorize(text))}
+		}
+		sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+		t.Log("top 5:")
+		for _, s := range all[:5] {
+			t.Logf("  [%.4f] [%-13s] %s", s.score, s.source, trimLen(s.text, 80))
+		}
+
+		var univHigh bool
+		for _, s := range all[:8] {
+			if strings.Contains(s.text, "医药大学") || strings.Contains(s.text, "升学") {
+				univHigh = true
+				break
+			}
+		}
+		if !univHigh {
+			t.Error("expected university-related events in top 8")
+		}
+	})
+
+	t.Run("老大私聊_agent主用Response", func(t *testing.T) {
+		q := "老大私聊说了什么"
+		qVec := e.VectorizeClean(q)
+
+		all := make([]scored, len(raw))
+		for i, ev := range raw {
+			text := clean(ev)
+			all[i] = scored{idx: i, source: ev.Source, text: text[:min(len(text), 200)], score: cosineSim(qVec, e.Vectorize(text))}
+		}
+		sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+		t.Log("top 5:")
+		for _, s := range all[:5] {
+			t.Logf("  [%.4f] [%-13s] %s", s.score, s.source, trimLen(s.text, 80))
+		}
+
+		var bossFound bool
+		for _, s := range all[:8] {
+			if strings.Contains(s.text, "老大") {
+				bossFound = true
+				break
+			}
+		}
+		if !bossFound {
+			t.Error("expected events mentioning 老大 in top 8")
+		}
+
+		var agentFound bool
+		for _, s := range all[:5] {
+			if s.source == "agent" {
+				agentFound = true
+				break
+			}
+		}
+		t.Logf("agent in top5: %v (source strategy: agent events use Response for vector)", agentFound)
+	})
+
+	t.Run("图片转SVG_去模版后效果", func(t *testing.T) {
+		q := "图片转换SVG工具"
+		qVec := e.VectorizeClean(q)
+
+		all := make([]scored, len(raw))
+		for i, ev := range raw {
+			text := clean(ev)
+			all[i] = scored{idx: i, source: ev.Source, text: text[:min(len(text), 200)], score: cosineSim(qVec, e.Vectorize(text))}
+		}
+		sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+		t.Log("top 5:")
+		for _, s := range all[:5] {
+			t.Logf("  [%.4f] [%-13s] %s", s.score, s.source, trimLen(s.text, 80))
+		}
+
+		var img bool
+		for _, s := range all[:5] {
+			if strings.Contains(s.text, "图片") || strings.Contains(s.text, "SVG") {
+				img = true
+				break
+			}
+		}
+		if !img {
+			t.Error("expected image-related events in top 5")
+		}
+	})
+
+	t.Run("南航航空航天", func(t *testing.T) {
+		q := "南航航空航天专业转电气"
+		qVec := e.VectorizeClean(q)
+
+		all := make([]scored, len(raw))
+		for i, ev := range raw {
+			text := clean(ev)
+			all[i] = scored{idx: i, source: ev.Source, text: text[:min(len(text), 200)], score: cosineSim(qVec, e.Vectorize(text))}
+		}
+		sort.Slice(all, func(i, j int) bool { return all[i].score > all[j].score })
+
+		t.Log("top 5:")
+		for _, s := range all[:5] {
+			t.Logf("  [%.4f] [%-13s] %s", s.score, s.source, trimLen(s.text, 80))
+		}
+
+		var nau bool
+		for _, s := range all[:5] {
+			if strings.Contains(s.text, "南航") {
+				nau = true
+				break
+			}
+		}
+		if !nau {
+			t.Error("expected 南航 in top 5")
+		}
+	})
+
+	t.Run("跨域区分度", func(t *testing.T) {
+		pairs := []struct {
+			a, b string
+		}{
+			{"老大私聊说了什么", "河南医药大学招生分数"},
+			{"老大私聊说了什么", "图片转换SVG工具"},
+			{"南航航空航天电气", "河南医药大学录取"},
+			{"图片转换SVG工具", "老大私聊"},
+		}
+		for _, p := range pairs {
+			va := e.VectorizeClean(p.a)
+			vb := e.VectorizeClean(p.b)
+			s := cosineSim(va, vb)
+			t.Logf("  sim(%q, %q) = %.4f", p.a, p.b, s)
+		}
+
+		univVec := e.VectorizeClean("河南医药大学招生")
+		bossVec := e.VectorizeClean("老大私聊说了什么")
+		t.Logf("cross-domain sim(医药大学, 老大私聊) = %.4f", cosineSim(univVec, bossVec))
+	})
+
+	t.Run("去模版节省量", func(t *testing.T) {
+		var savedTotal int
+		for i, ev := range raw {
+			orig := len(ev.Input + " " + ev.Response)
+			after := len(clean(ev))
+			saved := orig - after
+			savedTotal += saved
+			if saved > 100 {
+				t.Logf("  [%2d] [%-13s] 节省 %d 字符 (raw=%d clean=%d)", i, ev.Source, saved, orig, after)
+			}
+		}
+		t.Logf("总计节省 %d 字符", savedTotal)
+	})
+}
+
+func TestRealContextEmbedderStats(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		t.Skip("embedder not loaded")
+	}
+
+	data, err := os.ReadFile("/tmp/context.json")
+	if err != nil {
+		t.Fatal(err)
+	}
+	var raw []realEvent
+	if err := json.Unmarshal(data, &raw); err != nil {
+		t.Fatal(err)
+	}
+
+	for _, ev := range raw[:5] {
+		var text string
+		switch {
+		case ev.Source == "agent" && ev.Response != "":
+			text = CleanTemplateText(ev.Response)
+		case ev.Source == "cold_storage":
+			text = CleanTemplateText(ev.Input + " " + ev.Response)
+		default:
+			text = CleanTemplateText(ev.Input)
+		}
+		vec := e.Vectorize(text)
+		origLen := len(ev.Input + ev.Response)
+		t.Logf("[%-13s] raw=%d cleaned=%d dims=%d", ev.Source, origLen, len(text), len(vec))
+		for k := range vec {
+			if !isNumericKey(k) {
+				t.Errorf("non-numeric key %q — should be dense space", k)
+			}
+		}
+	}
+}
+
+func containsAny(s string, subs []string) bool {
+	for _, sub := range subs {
+		if sub != "" && strings.Contains(s, sub) {
+			return true
+		}
+	}
+	return false
+}
+
+func trimLen(s string, n int) string {
+	if len(s) > n {
+		return s[:n]
+	}
+	return s
+}
diff --git a/internal/memory/static_embedder.go b/internal/memory/static_embedder.go
new file mode 100644
index 0000000..860588e
--- /dev/null
+++ b/internal/memory/static_embedder.go
@@ -0,0 +1,369 @@
+package memory
+
+import (
+	"bufio"
+	"compress/gzip"
+	"fmt"
+	"log"
+	"math"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+	"unicode/utf8"
+
+	"github.com/yanyiwu/gojieba"
+	"gitcode.com/JianFeeeee/HomeAgent/internal/memory/vector"
+)
+
+const downloadMaxWords = 200000
+
+var knownModelURLs = []struct {
+	sub string
+	url string
+}{
+	{"numberbatch", "https://conceptnet.s3.amazonaws.com/downloads/2019/numberbatch/numberbatch-19.08.txt.gz"},
+	{"cc.zh.", "https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.zh.300.vec.gz"},
+	{"cc.en.", "https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.en.300.vec.gz"},
+}
+
+type StaticEmbedder struct {
+	mu        sync.RWMutex
+	jieba     *gojieba.Jieba
+	stopWords map[string]bool
+
+	words  map[string][]float64
+	dim    int
+	loaded bool
+
+	unkVec  []float64
+	unkNorm float64
+}
+
+func modelDownloadURL(modelPath string) string {
+	for _, m := range knownModelURLs {
+		if strings.Contains(modelPath, m.sub) {
+			return m.url
+		}
+	}
+	return knownModelURLs[0].url
+}
+
+func downloadFastTextModel(targetPath, url string) error {
+	tmpPath := targetPath + ".download.tmp"
+	if err := os.MkdirAll(filepath.Dir(targetPath), 0755); err != nil {
+		return fmt.Errorf("mkdir: %w", err)
+	}
+
+	f, err := os.Create(tmpPath)
+	if err != nil {
+		return fmt.Errorf("create tmp: %w", err)
+	}
+	defer f.Close()
+
+	resp, err := http.Get(url)
+	if err != nil {
+		os.Remove(tmpPath)
+		return fmt.Errorf("http get %s: %w", url, err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		os.Remove(tmpPath)
+		return fmt.Errorf("http status %s", resp.Status)
+	}
+
+	gz, err := gzip.NewReader(resp.Body)
+	if err != nil {
+		os.Remove(tmpPath)
+		return fmt.Errorf("gzip: %w", err)
+	}
+	defer gz.Close()
+
+	scanner := bufio.NewScanner(gz)
+	buf := make([]byte, 4*1024*1024)
+	scanner.Buffer(buf, len(buf))
+
+	writer := bufio.NewWriter(f)
+
+	if !scanner.Scan() {
+		os.Remove(tmpPath)
+		return fmt.Errorf("empty gzip content")
+	}
+	parts := strings.Fields(scanner.Text())
+	if len(parts) >= 2 {
+		fmt.Fprintf(writer, "%d %s\n", downloadMaxWords, parts[1])
+	} else {
+		fmt.Fprintln(writer, scanner.Text())
+	}
+
+	var lineCount int
+	for scanner.Scan() && lineCount < downloadMaxWords {
+		line := scanner.Text()
+		if line == "" {
+			continue
+		}
+		fmt.Fprintln(writer, line)
+		lineCount++
+
+		if lineCount%50000 == 0 {
+			log.Printf("[static_embedder] download progress: %d/%d words", lineCount, downloadMaxWords)
+		}
+	}
+
+	writer.Flush()
+	f.Close()
+
+	if err := os.Rename(tmpPath, targetPath); err != nil {
+		os.Remove(tmpPath)
+		return fmt.Errorf("rename: %w", err)
+	}
+
+	log.Printf("[static_embedder] download complete: %d words to %s", lineCount, targetPath)
+	return nil
+}
+
+func ensureModelFile(modelPath string) {
+	if modelPath == "" {
+		return
+	}
+	if _, err := os.Stat(modelPath); err == nil {
+		return
+	}
+	url := modelDownloadURL(modelPath)
+	log.Printf("[static_embedder] model %s not found, downloading from fastText...", modelPath)
+	if dlErr := downloadFastTextModel(modelPath, url); dlErr != nil {
+		log.Printf("[static_embedder] download failed: %v, will use TF-IDF fallback", dlErr)
+	} else {
+		log.Printf("[static_embedder] download ok")
+	}
+}
+
+func NewStaticEmbedder(modelPaths ...string) *StaticEmbedder {
+	sw := make(map[string]bool)
+	for k, v := range stopWords {
+		sw[k] = v
+	}
+	e := &StaticEmbedder{
+		jieba:     GetJieba(),
+		stopWords: sw,
+		words:     make(map[string][]float64),
+	}
+
+	if len(modelPaths) == 0 {
+		log.Printf("[static_embedder] no model path configured, using TF-IDF fallback")
+		return e
+	}
+
+	for _, p := range modelPaths {
+		ensureModelFile(p)
+	}
+	if err := e.loadAll(modelPaths); err != nil {
+		log.Printf("[static_embedder] load failed: %v, using TF-IDF fallback", err)
+	}
+	return e
+}
+
+func (e *StaticEmbedder) loadAll(paths []string) error {
+	var firstErr error
+	for i, p := range paths {
+		if p == "" {
+			continue
+		}
+		primary := i == 0
+		if err := e.load(p, primary); err != nil {
+			log.Printf("[static_embedder] load %s: %v", p, err)
+			if firstErr == nil {
+				firstErr = err
+			}
+		}
+	}
+	return firstErr
+}
+
+func (e *StaticEmbedder) load(path string, primary bool) error {
+	f, err := os.Open(path)
+	if err != nil {
+		return fmt.Errorf("open: %w", err)
+	}
+	defer f.Close()
+
+	scanner := bufio.NewScanner(f)
+	buf := make([]byte, 1024*1024)
+	scanner.Buffer(buf, len(buf))
+
+	if !scanner.Scan() {
+		return fmt.Errorf("empty file")
+	}
+	header := strings.TrimSpace(scanner.Text())
+	parts := strings.Fields(header)
+	if len(parts) < 2 {
+		return fmt.Errorf("invalid header: %s", header)
+	}
+	dim, err := strconv.Atoi(parts[1])
+	if err != nil || dim <= 0 {
+		return fmt.Errorf("invalid dimension: %s", parts[1])
+	}
+
+	if primary {
+		e.dim = dim
+	}
+
+	var vecSum []float64
+	var count int
+	if primary {
+		vecSum = make([]float64, dim)
+	}
+
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" {
+			continue
+		}
+		fields := strings.Fields(line)
+		if len(fields) < dim+1 {
+			continue
+		}
+		word := fields[0]
+
+		if _, exists := e.words[word]; exists {
+			continue
+		}
+
+		vec := make([]float64, dim)
+		for i := 0; i < dim; i++ {
+			v, _ := strconv.ParseFloat(fields[i+1], 64)
+			vec[i] = v
+		}
+		e.words[word] = vec
+		if primary {
+			for i := range vecSum {
+				vecSum[i] += vec[i]
+			}
+			count++
+		}
+	}
+
+	if primary {
+		if count == 0 {
+			return fmt.Errorf("no word vectors found in primary model")
+		}
+		for i := range vecSum {
+			vecSum[i] /= float64(count)
+		}
+		e.unkVec = make([]float64, dim)
+		copy(e.unkVec, vecSum)
+		var normSq float64
+		for _, v := range e.unkVec {
+			normSq += v * v
+		}
+		e.unkNorm = float64(math.Sqrt(normSq))
+		e.loaded = true
+	}
+
+	log.Printf("[static_embedder] loaded %d words, dim=%d from %s", len(e.words), e.dim, path)
+	return nil
+}
+
+func (e *StaticEmbedder) tokenize(text string) []string {
+	if e.jieba == nil {
+		return nil
+	}
+	words := e.jieba.Cut(text, true)
+	var result []string
+	seen := make(map[string]bool)
+	for _, w := range words {
+		w = strings.TrimSpace(w)
+		if w == "" || e.stopWords[w] || seen[w] {
+			continue
+		}
+		if utf8.RuneCountInString(w) < 2 {
+			continue
+		}
+		seen[w] = true
+		result = append(result, w)
+	}
+	return result
+}
+
+func (e *StaticEmbedder) Vectorize(text string) vector.Vector {
+	e.mu.RLock()
+	loaded := e.loaded
+	dim := e.dim
+	unkVec := e.unkVec
+	e.mu.RUnlock()
+
+	tokens := e.tokenize(text)
+	if len(tokens) == 0 {
+		return vector.Vector{}
+	}
+
+	tf := make(map[string]float64)
+	for _, t := range tokens {
+		tf[t]++
+	}
+	maxTF := 0.0
+	for _, c := range tf {
+		if c > maxTF {
+			maxTF = c
+		}
+	}
+
+	if !loaded {
+		vec := make(vector.Vector)
+		for word, count := range tf {
+			vec[word] = count / maxTF
+		}
+		return vec
+	}
+
+	sum := make([]float64, dim)
+	var weightSum float64
+
+	for word, count := range tf {
+		e.mu.RLock()
+		vec, ok := e.words[word]
+		e.mu.RUnlock()
+
+		w := count / maxTF
+
+		if !ok {
+			for i, v := range unkVec {
+				sum[i] += w * v
+			}
+		} else {
+			for i, v := range vec {
+				sum[i] += w * v
+			}
+		}
+		weightSum += w
+	}
+
+	if weightSum > 0 {
+		for i := range sum {
+			sum[i] /= weightSum
+		}
+	}
+
+	vec := make(vector.Vector, dim)
+	for i, v := range sum {
+		if v != 0 {
+			vec[strconv.Itoa(i)] = v
+		}
+	}
+	return vec
+}
+
+func (e *StaticEmbedder) Dim() int {
+	e.mu.RLock()
+	defer e.mu.RUnlock()
+	return e.dim
+}
+
+func (e *StaticEmbedder) Loaded() bool {
+	e.mu.RLock()
+	defer e.mu.RUnlock()
+	return e.loaded
+}
diff --git a/internal/memory/static_embedder_test.go b/internal/memory/static_embedder_test.go
new file mode 100644
index 0000000..8e13b96
--- /dev/null
+++ b/internal/memory/static_embedder_test.go
@@ -0,0 +1,236 @@
+package memory
+
+import (
+	"math"
+	"sort"
+	"testing"
+	"time"
+
+	"gitcode.com/JianFeeeee/HomeAgent/internal/memory/vector"
+)
+
+type testContextEvent struct {
+	Timestamp time.Time
+	Input     string
+	Response  string
+	Vector    vector.Vector
+}
+
+func cosineSim(a, b vector.Vector) float64 {
+	var dot, normA, normB float64
+	for f, va := range a {
+		dot += va * b[f]
+		normA += va * va
+	}
+	for _, vb := range b {
+		normB += vb * vb
+	}
+	if normA == 0 || normB == 0 {
+		return 0
+	}
+	return dot / (math.Sqrt(normA) * math.Sqrt(normB))
+}
+
+func TestStaticEmbedderLoad(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+	if !e.Loaded() {
+		t.Fatal("embedder should be loaded")
+	}
+	if e.Dim() != 300 {
+		t.Errorf("expected dim=300, got %d", e.Dim())
+	}
+}
+
+func TestStaticEmbedderConsistency(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+
+	v1 := e.Vectorize("今天天气怎么样")
+	v2 := e.Vectorize("今天天气怎么样")
+
+	if len(v1) != len(v2) {
+		t.Errorf("same input should produce same dimension count, got %d vs %d", len(v1), len(v2))
+	}
+	sim := cosineSim(v1, v2)
+	if math.Abs(sim-1.0) > 0.0001 {
+		t.Errorf("same input should have cosine similarity ~1.0, got %.6f", sim)
+	}
+}
+
+func isNumericKey(s string) bool {
+	if s == "" {
+		return false
+	}
+	for _, c := range s {
+		if c < '0' || c > '9' {
+			return false
+		}
+	}
+	return true
+}
+
+func TestStaticEmbedderFallback(t *testing.T) {
+	e := NewStaticEmbedder("")
+	if e.Loaded() {
+		t.Fatal("empty path embedder should not be loaded")
+	}
+
+	v := e.Vectorize("测试文本")
+	if len(v) == 0 {
+		t.Fatal("fallback vector should not be empty")
+	}
+}
+
+func TestStaticEmbedderAllInDenseSpace(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+
+	texts := []string{
+		"今天天气怎么样",
+		"股票基金投资",
+		"微积分导数数学题",
+		"台风天注意安全",
+		"基金定投",
+		"数学作业",
+		"明天会不会下雨",
+	}
+
+	for _, text := range texts {
+		v := e.Vectorize(text)
+		for k := range v {
+			if !isNumericKey(k) {
+				t.Errorf("%q produced non-numeric key %q — should be in dense space", text, k)
+			}
+		}
+	}
+	t.Log("all texts produce numeric keys — same dense space")
+}
+
+func TestStaticEmbedderSemanticSimilarity(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+
+	pairs := []struct {
+		a, b  string
+		related bool
+	}{
+		{"今天天气怎么样", "明天会不会下雨", true},
+		{"今天天气怎么样", "股票基金投资", false},
+		{"股票基金投资", "基金定投", true},
+		{"股票基金投资", "微积分导数数学题", false},
+		{"微积分导数数学题", "数学作业", true},
+	}
+
+	for _, p := range pairs {
+		va := e.Vectorize(p.a)
+		vb := e.Vectorize(p.b)
+		sim := cosineSim(va, vb)
+		t.Logf("sim(%q, %q) = %.4f  (related=%v)", p.a, p.b, sim, p.related)
+	}
+
+	weatherSim := cosineSim(e.Vectorize("今天天气怎么样"), e.Vectorize("明天会不会下雨"))
+	stockSim := cosineSim(e.Vectorize("今天天气怎么样"), e.Vectorize("股票基金投资"))
+	t.Logf("[verify] weather-weather=%.4f, weather-stock=%.4f", weatherSim, stockSim)
+	if weatherSim <= stockSim {
+		t.Errorf("weather-weather(%.4f) should be > weather-stock(%.4f)", weatherSim, stockSim)
+	}
+}
+
+func TestContextPruneWithRealEmbedding(t *testing.T) {
+	e := NewStaticEmbedder("/tmp/cc.zh.sample.vec")
+
+	type event struct {
+		input    string
+		response string
+	}
+	allEvents := []event{
+		{"今天天气怎么样", "挺好的"},
+		{"明天会不会下雨", "可能不会"},
+		{"台风来了", "注意安全"},
+		{"帮我算微积分", "好的"},
+		{"导数怎么求", "公式如下"},
+		{"数学作业", "解答"},
+		{"股票涨了", "恭喜"},
+		{"基金收益怎么样", "不错"},
+		{"最近有什么电影", "推荐"},
+		{"晚上吃什么", "随便"},
+		{"帮我定个闹钟", "好的"},
+		{"查询快递", "已送达"},
+	}
+
+	events := make([]testContextEvent, len(allEvents))
+	for i, ev := range allEvents {
+		events[i] = testContextEvent{
+			Timestamp: time.Now().Add(time.Duration(i) * time.Second),
+			Input:     ev.input,
+			Response:  ev.response,
+			Vector:    e.Vectorize(ev.input + " " + ev.response),
+		}
+	}
+
+	topK := 4
+	protectN := 3
+	query := "基金股票投资"
+	queryVec := e.Vectorize(query)
+
+	if len(events) <= topK+protectN {
+		t.Fatalf("need more events for pruning test")
+	}
+
+	protectStart := len(events) - protectN
+	protected := events[protectStart:]
+	candidates := events[:protectStart]
+
+	type scored struct {
+		evt   testContextEvent
+		score float64
+	}
+	scoredEvents := make([]scored, len(candidates))
+	for i, evt := range candidates {
+		scoredEvents[i] = scored{evt, cosineSim(queryVec, evt.Vector)}
+	}
+
+	sort.Slice(scoredEvents, func(i, j int) bool {
+		return scoredEvents[i].score > scoredEvents[j].score
+	})
+
+	keepCount := topK
+	if keepCount > len(scoredEvents) {
+		keepCount = len(scoredEvents)
+	}
+	keep := scoredEvents[:keepCount]
+	archived := scoredEvents[keepCount:]
+
+	t.Logf("query: %s", query)
+	t.Logf("=== retained (topK=%d) ===", topK)
+	for _, s := range keep {
+		t.Logf("  [%.4f] %s", s.score, s.evt.Input)
+	}
+	t.Logf("=== protected (recent %d) ===", protectN)
+	for _, e := range protected {
+		t.Logf("  %s", e.Input)
+	}
+	t.Logf("=== archived (%d items) ===", len(archived))
+	for _, s := range archived {
+		t.Logf("  [%.4f] %s", s.score, s.evt.Input)
+	}
+
+	hasFinance := false
+	for _, s := range keep {
+		if s.evt.Input == "股票涨了" || s.evt.Input == "基金收益怎么样" {
+			hasFinance = true
+		}
+	}
+	if !hasFinance {
+		t.Error("expected financial events to be retained, but none found")
+	}
+
+	hasWeather := false
+	for _, s := range keep {
+		if s.evt.Input == "今天天气怎么样" || s.evt.Input == "明天会不会下雨" || s.evt.Input == "台风来了" {
+			hasWeather = true
+		}
+	}
+	if hasWeather {
+		t.Log("NOTE: weather events are still in retained set — may have overlapping vocabulary")
+	}
+
+	t.Logf("remaining: %d = topK(%d) + protectN(%d)", topK+protectN, topK, protectN)
+}