@morningljn/mnemo 0.2.0 → 0.3.0

package/docs/superpowers/plans/2026-05-16-memory-dreaming.md

@@ -0,0 +1,626 @@
+# Memory Dreaming 实现计划
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** 为 mnemo-mcp 添加后台记忆整理（dream）和搜索结果精简功能，保持数据库精炼、减少 token 消耗。
+
+**Architecture:** 在 `MemoryStore` 中新增 `runDream()` 方法，编排合并去重、摘要压缩、分类修正三阶段。新增 `CompactResult` 类型精简搜索返回字段。新增 CLI 入口 `src/dream.ts`。
+
+**Tech Stack:** TypeScript, better-sqlite3, Vitest
+
+---
+
+## File Structure
+
+| File | Responsibility |
+|------|---------------|
+| `src/types.ts` | 新增 `DreamReport` 和 `CompactFactResult` 类型 |
+| `src/store.ts` | 新增 `runDream()`、`mergeOverlappingFacts()`、`compressLongFacts()`、`reclassifyFacts()`、`backupDatabase()` |
+| `src/retriever.ts` | 搜索结果格式化为 `CompactFactResult` |
+| `src/server.ts` | 新增 `dream` action，所有搜索 action 使用精简格式 |
+| `src/dream.ts` | CLI 入口，执行 dream 并输出 report |
+| `package.json` | 新增 `mnemo-dream` bin |
+| `tests/store.test.ts` | dream 相关单元测试 |
+
+---
+
+### Task 1: 新增类型定义
+
+**Files:**
+- Modify: `src/types.ts`
+
+- [ ] **Step 1: 在 types.ts 末尾添加 DreamReport 和 CompactFactResult 类型**
+
+```typescript
+/** Dream 整理报告 */
+export interface DreamReport {
+  merged: number
+  compressed: number
+  reclassified: number
+  deleted: number
+  mergeDetails: Array<{ kept: number; removed: number; similarity: number }>
+  health: {
+    total: number
+    avg_trust: number
+    avg_length: number
+    coverage: Record<FactCategory, number>
+  }
+}
+
+/** 精简搜索结果 */
+export interface CompactFactResult {
+  factId: number
+  display: string
+  category: FactCategory
+  trustScore: number
+  score: number
+}
+```
+
+- [ ] **Step 2: 更新 FactStoreArgs 的 action 联合类型**
+
+在 `src/types.ts` 的 `FactStoreArgs.action` 字段中，把 `'learn' | 'audit'` 改为 `'learn' | 'audit' | 'dream'`：
+
+```typescript
+action: 'add' | 'search' | 'probe' | 'related' | 'reason' | 'contradict' | 'update' | 'remove' | 'list' | 'learn' | 'audit' | 'dream'
+```
+
+- [ ] **Step 3: 运行构建确认类型无报错**
+
+Run: `npm run build`
+Expected: 编译成功，无报错
+
+- [ ] **Step 4: Commit**
+
+```bash
+git add src/types.ts
+git commit -m "feat(types): add DreamReport and CompactFactResult types"
+```
+
+---
+
+### Task 2: Store 层 dream 方法 — 备份与压缩
+
+**Files:**
+- Modify: `src/store.ts`
+
+- [ ] **Step 1: 写失败测试 — backupDatabase**
+
+在 `tests/store.test.ts` 的最后一个 `describe` 块之后添加：
+
+```typescript
+describe('dream - backup', () => {
+  it('creates backup before dream', () => {
+    const id = store.addFact('test fact for backup', 'general')
+    const result = store.backupDatabase()
+    expect(result).toBeTruthy()
+    expect(result).toContain('dream-')
+    expect(result).toContain('.db')
+  })
+})
+```
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: FAIL — `backupDatabase` 不存在
+
+- [ ] **Step 2: 实现 backupDatabase**
+
+在 `src/store.ts` 的 `runAudit()` 方法之后、`get connection()` 之前添加：
+
+```typescript
+/** Dream 前备份数据库 */
+backupDatabase(): string {
+  const { mkdirSync, copyFileSync } = require('node:fs')
+  const { join, dirname } = require('node:path')
+  const { homedir } = require('node:os')
+  const backupDir = join(homedir(), '.mnemo', 'backup')
+  mkdirSync(backupDir, { recursive: true })
+  const timestamp = new Date().toISOString().replace(/[:.]/g, '-').slice(0, 19)
+  const backupPath = join(backupDir, `dream-${timestamp}.db`)
+  copyFileSync(this.db.name, backupPath)
+  return backupPath
+}
+```
+
+注意：需要在文件顶部添加 `import { copyFileSync, mkdirSync } from 'node:fs'`（如果还没有的话）。实际上 `mkdirSync` 已在顶部导入了，只需确认 `copyFileSync` 也已导入。检查 store.ts 顶部的 import，如果没有 `copyFileSync`，添加到现有 `mkdirSync` 的 import 行中。
+
+- [ ] **Step 3: 运行测试确认通过**
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: PASS
+
+- [ ] **Step 4: 写失败测试 — compressLongFacts**
+
+```typescript
+describe('dream - compress', () => {
+  it('generates summary for long facts without summary', () => {
+    const longContent = '用户偏好使用 TypeScript 开发前端项目。偏好 React 框架进行组件化开发。' + '额外补充说明'.repeat(50)
+    store.addFact(longContent, 'coding_style')
+    const result = store.compressLongFacts()
+    expect(result).toBeGreaterThanOrEqual(1)
+    const row = store.connection.prepare('SELECT summary FROM facts WHERE content = ?').get(longContent) as any
+    expect(row.summary).toBeTruthy()
+    expect(row.summary.length).toBeLessThanOrEqual(150)
+    expect(row.summary).toContain('TypeScript')
+  })
+
+  it('skips facts with existing summary', () => {
+    const longContent = 'x'.repeat(300)
+    const id = store.addFact(longContent, 'general')
+    store.connection.prepare('UPDATE facts SET summary = ? WHERE fact_id = ?').run('existing summary', id)
+    const result = store.compressLongFacts()
+    const row = store.connection.prepare('SELECT summary FROM facts WHERE fact_id = ?').get(id) as any
+    expect(row.summary).toBe('existing summary')
+  })
+
+  it('skips short facts', () => {
+    store.addFact('short fact', 'general')
+    const result = store.compressLongFacts()
+    expect(result).toBe(0)
+  })
+})
+```
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: FAIL — `compressLongFacts` 不存在
+
+- [ ] **Step 5: 实现 compressLongFacts**
+
+在 `store.ts` 的 `backupDatabase()` 之后添加：
+
+```typescript
+/** 压缩长 fact：content > 200 字且无 summary 的，自动提取前 2 句作为 summary */
+compressLongFacts(): number {
+  const rows = this.db.prepare(
+    'SELECT fact_id, content FROM facts WHERE length(content) > 200 AND (summary IS NULL OR summary = "")'
+  ).all() as Array<{ fact_id: number; content: string }>
+
+  let compressed = 0
+  for (const row of rows) {
+    const summary = this.extractSummary(row.content)
+    if (summary) {
+      this.db.prepare('UPDATE facts SET summary = ? WHERE fact_id = ?').run(summary, row.fact_id)
+      compressed++
+    }
+  }
+  return compressed
+}
+
+/** 从 content 提取前 2 个完整句子（总长 ≤ 150 字） */
+private extractSummary(content: string): string | null {
+  const sentences = content.split(/[。\n.]/).map(s => s.trim()).filter(s => s.length > 0)
+  if (sentences.length === 0) return null
+  let summary = sentences[0]
+  if (sentences.length > 1 && summary.length + sentences[1].length <= 148) {
+    summary += '。' + sentences[1]
+  }
+  return summary.length <= 150 ? summary : summary.slice(0, 147) + '...'
+}
+```
+
+- [ ] **Step 6: 运行测试确认通过**
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: PASS
+
+- [ ] **Step 7: Commit**
+
+```bash
+git add src/store.ts tests/store.test.ts
+git commit -m "feat(store): add backupDatabase and compressLongFacts for dream cycle"
+```
+
+---
+
+### Task 3: Store 层 dream 方法 — 合并与分类修正
+
+**Files:**
+- Modify: `src/store.ts`
+- Modify: `tests/store.test.ts`
+
+- [ ] **Step 1: 写失败测试 — mergeOverlappingFacts**
+
+```typescript
+describe('dream - merge', () => {
+  it('merges overlapping facts in same category', () => {
+    store.addFact('用户偏好使用 TypeScript 编写前端代码', 'coding_style')
+    store.addFact('用户偏好使用 TypeScript 编写后端代码', 'coding_style')
+    const result = store.mergeOverlappingFacts()
+    expect(result.merged).toBeGreaterThanOrEqual(1)
+    expect(result.details.length).toBeGreaterThanOrEqual(1)
+  })
+
+  it('protects high frequency facts from deletion', () => {
+    const id1 = store.addFact('用户偏好使用 TypeScript 编写前端代码', 'coding_style')
+    const id2 = store.addFact('用户偏好使用 TypeScript 编写前端代码扩展', 'coding_style')
+    store.connection.prepare('UPDATE facts SET retrieval_count = 200 WHERE fact_id = ?').run(id1)
+    const result = store.mergeOverlappingFacts()
+    const kept = store.connection.prepare('SELECT fact_id FROM facts WHERE fact_id = ?').get(id1) as any
+    expect(kept).toBeTruthy()
+  })
+
+  it('does not merge facts across categories', () => {
+    store.addFact('用户偏好使用 TypeScript 编写前端代码', 'coding_style')
+    store.addFact('用户偏好使用 TypeScript 编写前端代码', 'general')
+    const result = store.mergeOverlappingFacts()
+    expect(result.merged).toBe(0)
+  })
+})
+```
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: FAIL — `mergeOverlappingFacts` 不存在
+
+- [ ] **Step 2: 实现 mergeOverlappingFacts**
+
+在 `store.ts` 的 `compressLongFacts()` 之后添加：
+
+```typescript
+/** 合并同 category 内 Jaccard > 0.6 的重叠 fact */
+mergeOverlappingFacts(): { merged: number; details: Array<{ kept: number; removed: number; similarity: number }> } {
+  const categories: FactCategory[] = ['identity', 'coding_style', 'tool_pref', 'workflow', 'general']
+  let merged = 0
+  const details: Array<{ kept: number; removed: number; similarity: number }> = []
+
+  for (const cat of categories) {
+    const rows = this.db.prepare(
+      'SELECT fact_id, content, retrieval_count FROM facts WHERE category = ? ORDER BY trust_score DESC'
+    ).all(cat) as Array<{ fact_id: number; content: string; retrieval_count: number }>
+
+    const removed = new Set<number>()
+
+    for (let i = 0; i < rows.length; i++) {
+      if (removed.has(rows[i].fact_id)) continue
+      const tokensA = this.tokenizeForDedup(rows[i].content)
+
+      for (let j = i + 1; j < rows.length; j++) {
+        if (removed.has(rows[j].fact_id)) continue
+        const tokensB = this.tokenizeForDedup(rows[j].content)
+        const sim = this.jaccardSimilarity(tokensA, tokensB)
+
+        if (sim > 0.6) {
+          // 高频保护：retrieval_count > 100 的不能被删除
+          const aHighFreq = rows[i].retrieval_count > 100
+          const bHighFreq = rows[j].retrieval_count > 100
+
+          if (aHighFreq && bHighFreq) continue
+
+          let keptId: number, removedId: number
+          if (bHighFreq) {
+            keptId = rows[j].fact_id
+            removedId = rows[i].fact_id
+          } else {
+            keptId = rows[i].fact_id
+            removedId = rows[j].fact_id
+          }
+
+          this.removeFact(removedId)
+          removed.add(removedId)
+          details.push({ kept: keptId, removed: removedId, similarity: Math.round(sim * 100) / 100 })
+          merged++
+        }
+      }
+    }
+  }
+  return { merged, details }
+}
+```
+
+- [ ] **Step 3: 运行测试确认通过**
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: PASS
+
+- [ ] **Step 4: 写失败测试 — reclassifyFacts**
+
+```typescript
+describe('dream - reclassify', () => {
+  it('moves miscategorized facts by keywords', () => {
+    const id = store.addFact('编码规范：文件不超过 500 行', 'identity')
+    const result = store.reclassifyFacts()
+    expect(result).toBeGreaterThanOrEqual(1)
+    const row = store.connection.prepare('SELECT category FROM facts WHERE fact_id = ?').get(id) as any
+    expect(row.category).toBe('coding_style')
+  })
+
+  it('skips correctly categorized facts', () => {
+    store.addFact('用户偏好使用 VS Code 编辑器', 'tool_pref')
+    const result = store.reclassifyFacts()
+    expect(result).toBe(0)
+  })
+})
+```
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: FAIL — `reclassifyFacts` 不存在
+
+- [ ] **Step 5: 实现 reclassifyFacts**
+
+在 `store.ts` 的 `mergeOverlappingFacts()` 之后添加：
+
+```typescript
+/** 分类修正：按关键词规则表将误分类的 fact 挪到正确 category */
+reclassifyFacts(): number {
+  const rules: Array<{ keywords: string[]; target: FactCategory }> = [
+    { keywords: ['角色设定', '暖暖', '身份', '编程女朋友', '暖宝宝'], target: 'identity' },
+    { keywords: ['编码规范', '代码风格', 'pytest', '文件不超过', '方法不超过'], target: 'coding_style' },
+    { keywords: ['工作流', 'OpenSpec', 'writing-plans', 'subagent'], target: 'workflow' },
+    { keywords: ['偏好', 'VS Code', '编辑器', 'IDE', '快捷键'], target: 'tool_pref' },
+  ]
+
+  const rows = this.db.prepare(
+    'SELECT fact_id, content, category FROM facts'
+  ).all() as Array<{ fact_id: number; content: string; category: string }>
+
+  let reclassified = 0
+  for (const row of rows) {
+    for (const rule of rules) {
+      if (rule.target === row.category) continue
+      if (rule.keywords.some(kw => row.content.includes(kw))) {
+        this.db.prepare('UPDATE facts SET category = ? WHERE fact_id = ?').run(rule.target, row.fact_id)
+        reclassified++
+        break
+      }
+    }
+  }
+  return reclassified
+}
+```
+
+- [ ] **Step 6: 运行测试确认通过**
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: PASS
+
+- [ ] **Step 7: 写端到端测试 — runDream**
+
+```typescript
+describe('dream - runDream', () => {
+  it('runs full dream cycle and returns report', () => {
+    // 长文 fact（触发压缩）
+    store.addFact('用户偏好使用 TypeScript 开发前端。使用 React 框架。' + 'x'.repeat(250), 'coding_style')
+    // 重叠 fact（触发合并）
+    store.addFact('用户偏好使用 TypeScript 开发前端代码', 'coding_style')
+    // 分类错误 fact（触发重分类）
+    store.addFact('编码规范：文件不超过 500 行', 'identity')
+
+    const report = store.runDream({ skipBackup: true })
+    expect(report.compressed).toBeGreaterThanOrEqual(0)
+    expect(report.merged).toBeGreaterThanOrEqual(0)
+    expect(report.reclassified).toBeGreaterThanOrEqual(0)
+    expect(report.health.total).toBeGreaterThanOrEqual(1)
+    expect(report.health.coverage).toBeTruthy()
+  })
+})
+```
+
+Run: `npx vitest run tests/store.test.ts`
+Expected: FAIL — `runDream` 不存在
+
+- [ ] **Step 8: 实现 runDream**
+
+在 `store.ts` 的 `reclassifyFacts()` 之后添加：
+
+```typescript
+/** 执行完整 dream cycle：备份 → 压缩 → 合并 → 重分类 → 报告 */
+runDream(options?: { skipBackup?: boolean }): DreamReport {
+  // 备份
+  if (!options?.skipBackup) {
+    this.backupDatabase()
+  }
+
+  // 阶段 1：压缩长文
+  const compressed = this.compressLongFacts()
+
+  // 阶段 2：合并重叠
+  const mergeResult = this.mergeOverlappingFacts()
+
+  // 阶段 3：分类修正
+  const reclassified = this.reclassifyFacts()
+
+  // 健康统计
+  const stats = this.db.prepare(`
+    SELECT COUNT(*) as total,
+           AVG(trust_score) as avg_trust,
+           AVG(length(content)) as avg_length
+    FROM facts
+  `).get() as { total: number; avg_trust: number; avg_length: number }
+
+  const categories: FactCategory[] = ['identity', 'coding_style', 'tool_pref', 'workflow', 'general']
+  const coverage: Record<string, number> = {}
+  for (const cat of categories) {
+    const row = this.db.prepare('SELECT COUNT(*) as c FROM facts WHERE category = ?').get(cat) as { c: number }
+    coverage[cat] = row.c
+  }
+
+  return {
+    merged: mergeResult.merged,
+    compressed,
+    reclassified,
+    deleted: mergeResult.merged,
+    mergeDetails: mergeResult.details,
+    health: {
+      total: stats.total,
+      avg_trust: Math.round((stats.avg_trust ?? 0) * 100) / 100,
+      avg_length: Math.round(stats.avg_length ?? 0),
+      coverage: coverage as Record<FactCategory, number>,
+    },
+  }
+}
+```
+
+需要在 `store.ts` 顶部添加 `import type { DreamReport } from './types.js'`（如果还没有的话）。
+
+- [ ] **Step 9: 运行全部测试**
+
+Run: `npx vitest run`
+Expected: ALL PASS
+
+- [ ] **Step 10: Commit**
+
+```bash
+git add src/store.ts tests/store.test.ts
+git commit -m "feat(store): add mergeOverlappingFacts, reclassifyFacts, runDream for dream cycle"
+```
+
+---
+
+### Task 4: Server 层 — dream action + 精简搜索格式
+
+**Files:**
+- Modify: `src/server.ts`
+- Modify: `src/retriever.ts`
+
+- [ ] **Step 1: 在 server.ts 添加 dream case**
+
+在 `server.ts` 的 `case 'audit'` 块之后、`case 'list'` 之前添加：
+
+```typescript
+        case 'dream': {
+          const report = store.runDream()
+          retriever.getCache().clear()
+          resourceManager.invalidate()
+          return { content: [{ type: 'text' as const, text: JSON.stringify(report) }] }
+        }
+```
+
+同时在 `factStoreSchema` 的 `action` 枚举中添加 `'dream'`。
+
+- [ ] **Step 2: 更新 factStoreSchema 的 action 枚举**
+
+在 `src/server.ts` 中，把 `action: z.enum([... 'learn', 'audit'])` 改为 `action: z.enum([... 'learn', 'audit', 'dream'])`。
+
+- [ ] **Step 3: 在 server.ts 添加 toCompactResult 辅助函数**
+
+在 server.ts 的 `resolveCategory` 函数之后添加：
+
+```typescript
+function toCompactResult(f: ScoredFact): CompactFactResult {
+  return {
+    factId: f.factId,
+    display: f.summary ?? (f.content.length > 100 ? f.content.slice(0, 100) + '...' : f.content),
+    category: f.category,
+    trustScore: Math.round(f.trustScore * 100) / 100,
+    score: Math.round(f.score * 1000) / 1000,
+  }
+}
+```
+
+需要在 `server.ts` 顶部添加 `import type { CompactFactResult } from './types.js'`。
+
+- [ ] **Step 4: 更新 search/probe/related/reason 响应使用精简格式**
+
+把 `case 'search'` 中的：
+```typescript
+return { content: [{ type: 'text' as const, text: JSON.stringify({ results, count: results.length }) }] }
+```
+改为：
+```typescript
+const compact = results.map(toCompactResult)
+return { content: [{ type: 'text' as const, text: JSON.stringify({ results: compact, count: compact.length }) }] }
+```
+
+对 `probe`、`related`、`reason` 做同样的改动。
+
+- [ ] **Step 5: 构建并运行测试**
+
+Run: `npm run build && npx vitest run`
+Expected: ALL PASS
+
+- [ ] **Step 6: Commit**
+
+```bash
+git add src/server.ts
+git commit -m "feat(server): add dream action, compact search result format"
+```
+
+---
+
+### Task 5: CLI dream 命令
+
+**Files:**
+- Create: `src/dream.ts`
+- Modify: `package.json`
+
+- [ ] **Step 1: 创建 src/dream.ts**
+
+```typescript
+#!/usr/bin/env node
+
+import { MemoryStore } from './store.js'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
+
+const dbPath = join(homedir(), '.mnemo', 'facts.db')
+const store = new MemoryStore(dbPath)
+
+try {
+  console.log('[mnemo dream] 开始整理记忆库...\n')
+  const report = store.runDream()
+  console.log(JSON.stringify(report, null, 2))
+  console.log(`\n[mnemo dream] 完成: merged=${report.merged} compressed=${report.compressed} reclassified=${report.reclassified} deleted=${report.deleted}`)
+} catch (err) {
+  console.error('[mnemo dream] error:', err)
+  process.exit(1)
+} finally {
+  store.close()
+}
+```
+
+- [ ] **Step 2: 在 package.json 添加 bin 入口**
+
+在 `package.json` 的 `bin` 字段中添加 `"mnemo-dream": "dist/dream.js"`：
+
+```json
+"bin": {
+  "mnemo": "dist/server.js",
+  "mnemo-init": "dist/init.js",
+  "mnemo-dream": "dist/dream.js"
+}
+```
+
+- [ ] **Step 3: 构建并验证**
+
+Run: `npm run build && node dist/dream.js`
+Expected: 输出 dream report JSON
+
+- [ ] **Step 4: Commit**
+
+```bash
+git add src/dream.ts package.json
+git commit -m "feat(cli): add mnemo dream CLI command"
+```
+
+---
+
+### Task 6: 最终验证
+
+- [ ] **Step 1: 运行全部测试**
+
+Run: `npx vitest run`
+Expected: ALL PASS
+
+- [ ] **Step 2: 构建并验证 MCP 协议**
+
+Run: `npm run build`
+
+验证 dream action：
+```bash
+printf '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}\n{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"fact_store","arguments":{"action":"dream"}}}\n' | node dist/server.js 2>/dev/null
+```
+Expected: 返回 dream report JSON
+
+验证精简搜索格式：
+```bash
+printf '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}\n{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"fact_store","arguments":{"action":"search","query":"编码规范"}}}\n' | node dist/server.js 2>/dev/null
+```
+Expected: 返回包含 `display` 字段（而非完整 `content`）的精简结果
+
+- [ ] **Step 3: 版本号更新并最终 Commit**
+
+Run:
+```bash
+npm version patch --no-git-tag-version
+git add -A
+git commit -m "feat: memory dreaming — auto merge, compress, reclassify + compact search results"
+```

package/openspec/changes/archive/2026-05-16-memory-dreaming/.openspec.yaml

@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-05-16

package/openspec/changes/archive/2026-05-16-memory-dreaming/design.md

@@ -0,0 +1,71 @@
+## Context
+
+mnemo-mcp 是一个基于 SQLite + FTS5 的结构化事实记忆系统。长期使用后，fact 库会出现以下问题：
+- 多条 fact 内容重叠（如 5 条 Python 编码规范偏好，内容有交叉）
+- 单条 fact 过长（超 500 字的"万能条"，占 token、降低检索精度）
+- 分类错误（identity 类里混了 workflow 内容）
+- trust 评分不准确（`runLearning` 只看 feedback rate，忽略高频检索信号）
+
+当前 `runLearning()` 仅调 trust 分，`runAudit()` 只读不改。没有内容层面的整理能力。
+
+## Goals / Non-Goals
+
+**Goals:**
+- 定期自动整理 fact 库：合并重叠、压缩长文、修正分类
+- 整理后搜索结果更精准、token 消耗更少
+- 支持 CLI 命令 `mnemo dream` 手动触发，也支持 cron 定时
+- 整理操作安全：合并前确认、可回滚
+
+**Non-Goals:**
+- 不做 AI 生成 summary（用规则提取关键句，不调 LLM）
+- 不做跨项目记忆迁移
+- 不做可视化 dashboard
+- 不改变 MCP Resource 注入方式
+
+## Decisions
+
+### 1. 合并策略：Jaccard 相似度 > 0.6 且同 category
+
+用现有的 `tokenizeForDedup()` + `jaccardSimilarity()` 计算两两相似度。同 category 内 Jaccard > 0.6 的 fact 对，合并为一条（保留 content 更长或 trust 更高的，删除另一条）。
+
+**不跨 category 合并**：不同 category 的 fact 即使内容相似也保留（可能是不同语境）。
+
+### 2. 长文压缩：规则提取关键句
+
+content > 200 字且无 summary 的 fact，自动提取前 2 个完整句子作为 summary。不调 LLM，纯规则：
+- 按中文句号（。）、英文句号（.）、换行符分割
+- 取前 2 个句子（总长 ≤ 150 字）
+
+### 3. 分类修正：关键词规则表
+
+硬编码规则表，dream 时扫描并修正：
+- 包含"角色设定/暖暖/身份" → identity
+- 包含"编码规范/代码风格/pytest" → coding_style
+- 包含"工作流/OpenSpec/工作流" → workflow
+- 包含"偏好/VS Code/编辑器" → tool_pref
+
+### 4. Dream report 输出格式
+
+```
+{
+  merged: 3,
+  compressed: 5,
+  reclassified: 2,
+  deleted: 4,
+  health: { total: 65, avg_trust: 0.62, avg_length: 180, coverage: { identity: 8, coding_style: 12, ... } }
+}
+```
+
+### 5. 搜索结果精简
+
+`search` 返回时：
+- 有 summary → 返回 summary
+- 无 summary → 返回 content 前 100 字 + "..."
+- 始终返回 factId、category、trustScore、score
+- 不返回完整 content（减少 token）
+
+## Risks / Trade-offs
+
+- **[误合并]** → Jaccard > 0.6 可能误判相似内容。缓解：合并前 log 记录，dream report 列出所有合并对
+- **[summary 质量低]** → 规则提取的前 2 句可能不是最关键的。缓解：用户可手动 update summary
+- **[不可逆删除]** → 合并后删除冗余 fact 是不可逆的。缓解：dream 前自动备份数据库