feat(memory): add LLM-driven memory extraction from conversations

- memory/extractor: async LLM-based memory extraction from chat history
- Parses JSON lines from LLM response as memory entries
- Wire into orchestrator pipeline via PostProcess hook
- Non-blocking: runs in goroutine with 30s timeout
- Automatic dedup via existing keyword overlap logic

backend/services/memory/extractor_test.go

@@ -0,0 +1,99 @@
+package memory
+
+import (
+	"context"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+)
+
+func TestParseExtractedMemories(t *testing.T) {
+	tests := []struct {
+		input    string
+		minCount int
+	}{
+		{`{"content":"用户喜欢咖啡","category":"preference","priority":0.8}
+{"content":"用户每天跑步","category":"fact","priority":0.7}`, 2},
+		{`这是提取的记忆：
+{"content":"用户名叫张三","category":"fact","priority":0.9}
+{"content":"用户在北京工作","category":"fact","priority":0.85}`, 2},
+		{`no json here`, 0},
+		{``, 0},
+	}
+
+	for _, tt := range tests {
+		t.Run("", func(t *testing.T) {
+			result := parseExtractedMemories(tt.input)
+			if len(result) < tt.minCount {
+				t.Errorf("expected >= %d memories, got %d: %+v", tt.minCount, len(result), result)
+			}
+			for _, m := range result {
+				if m.Content == "" {
+					t.Error("memory content should not be empty")
+				}
+				if m.Category == "" {
+					t.Error("memory category should not be empty")
+				}
+			}
+		})
+	}
+}
+
+func TestExtractorDisabled(t *testing.T) {
+	e := NewExtractor(nil, nil)
+	if e.enabled {
+		t.Error("extractor should be disabled when llm is nil")
+	}
+
+	// Should not panic
+	e.ExtractFromConversation(nil, "user-1", nil)
+}
+
+func TestExtractorParseEdgeCases(t *testing.T) {
+	// Test JSON with whitespace
+	result := parseExtractedMemories(`  {"content":"test","category":"fact","priority":0.5}  `)
+	if len(result) != 1 {
+		t.Errorf("expected 1, got %d", len(result))
+	}
+
+	// Test empty JSON object
+	result = parseExtractedMemories(`{}`)
+	if len(result) != 0 {
+		t.Errorf("expected 0 for empty object, got %d", len(result))
+	}
+
+	// Test invalid JSON line mixed with valid
+	result = parseExtractedMemories(`invalid
+{"content":"ok","category":"fact","priority":0.5}
+not json either`)
+	if len(result) != 1 {
+		t.Errorf("expected 1 valid, got %d", len(result))
+	}
+}
+
+// mockLLM implements LLMClient for testing.
+type mockLLM struct {
+	response string
+	err      error
+}
+
+func (m *mockLLM) Chat(_ context.Context, _ []model.LLMMessage) (*model.LLMResponse, error) {
+	if m.err != nil { return nil, m.err }
+	return &model.LLMResponse{Content: m.response, Model: "mock"}, nil
+}
+
+func TestExtractorWithMockLLM(t *testing.T) {
+	mock := &mockLLM{
+		response: `{"content":"用户喜欢喝咖啡","category":"preference","priority":0.8}
+{"content":"用户在2024年去过日本","category":"event","priority":0.9}`,
+	}
+
+	// We can't easily test the full extraction without a real DB
+	// But verify mock LLM works
+	resp, err := mock.Chat(nil, nil)
+	if err != nil { t.Fatalf("mock LLM: %v", err) }
+	memories := parseExtractedMemories(resp.Content)
+	if len(memories) != 2 {
+		t.Errorf("expected 2 memories, got %d", len(memories))
+	}
+}