@morningljn/mnemo 0.1.3 → 0.2.0

package/docs/superpowers/plans/2026-05-16-retrieval-and-injection-optimization.md

@@ -0,0 +1,770 @@
+# Retrieval and Injection Optimization Implementation Plan
+
+> **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:** Improve mnemo-mcp retrieval accuracy and reduce injection overhead by adding MCP Resources for session warmup, query refinement, adaptive scoring, and a new injection protocol.
+
+**Architecture:** Add a `ResourceManager` to serve per-category memory summaries via MCP Resource protocol, a `refineQuery()` pure function to strip noise tokens from user messages, dynamic FTS/Jaccard weighting based on query length, content-based deduplication, and a relevance score gate. Update CLAUDE.md rules to shift from "search every message" to "Resource warmup + on-demand search".
+
+**Tech Stack:** TypeScript, Node.js, better-sqlite3, @modelcontextprotocol/sdk v1.29, zod v4, vitest
+
+---
+
+## File Structure
+
+| File | Responsibility |
+|------|---------------|
+| `src/refine.ts` | **NEW** — `refineQuery()` pure function: token filtering, action word removal, entity extraction |
+| `src/resources.ts` | **NEW** — `ResourceManager` class: per-category Resource registration, caching, cache invalidation |
+| `src/retriever.ts` | **MODIFY** — Dynamic scoring weights, relevance gate, content-based deduplication, integrate refineQuery |
+| `src/server.ts` | **MODIFY** — Wire ResourceManager, pass refineQuery into search flow |
+| `src/types.ts` | **MODIFY** — Add `RefineResult` type |
+| `tests/refine.test.ts` | **NEW** — Query refinement unit tests |
+| `tests/resource.test.ts` | **NEW** — Resource caching and invalidation tests |
+| `tests/retriever.test.ts` | **MODIFY** — Dynamic weight, relevance gate, dedup tests |
+| `CLAUDE.md` (user-side) | **UPDATE** — New injection protocol rules (provided in docs) |
+
+---
+
+## Task 1: Query Refinement Module
+
+**Files:**
+- Create: `src/refine.ts`
+- Test: `tests/refine.test.ts`
+
+- [ ] **Step 1: Create `src/refine.ts`**
+
+```typescript
+/**
+ * Query refinement: strip noise tokens from user messages before memory search.
+ * Pure function — no side effects, no DB access.
+ */
+
+import type { FactCategory } from './types.js'
+
+// Action words / helper phrases to strip (Chinese)
+const ACTION_WORDS = new Set([
+  '帮我', '看看', '看一下', '做一下', '帮我看看', '能不能', '为什么', '怎么',
+  '是什么', '如何', '请', '麻烦', '可以', '能不能', '能不能帮我', '给我',
+  '给我看看', '给我做', '给我写', '给我查', '给我找', '给我说', '给我讲',
+  '告诉我', '跟我说', '跟我讲', '给我解释', '给我说明', '给我介绍',
+  '运行', '执行', '启动', '停止', '创建', '删除', '修改', '更新', '查看',
+  '检查', '测试', '提交', '推送', '拉取', '合并', '切换', '重置',
+])
+
+// Reuse existing stop words from retriever
+const CN_STOP_WORDS = new Set([
+  '的', '了', '是', '在', '有', '和', '就', '不', '人', '都',
+  '一', '个', '上', '也', '很', '到', '说', '要', '去', '你',
+  '会', '着', '没', '看', '好', '自', '这', '他', '她', '它',
+  '那', '些', '用', '对', '下', '为', '从', '被', '把', '能',
+  '可', '以', '所', '而', '又', '与', '但', '或', '等', '中',
+  '大', '小', '多', '少', '其', '之', '做', '让', '给', '已',
+  '还', '来', '地', '得', '过', '时', '里', '后', '前', '当',
+])
+
+export interface RefineResult {
+  query: string | null
+  tokens: string[]
+  entityTokens: string[]
+}
+
+/**
+ * Refine a raw user message into memory-searchable keywords.
+ * Returns null if the message is a pure operation command with no memory relevance.
+ */
+export function refineQuery(raw: string): RefineResult | null {
+  const trimmed = raw.trim()
+  if (!trimmed) return null
+
+  // Extract high-signal tokens first: quoted content, book titles, capitalized phrases
+  const entityTokens: string[] = []
+
+  // Chinese quotes: 「深色主题」 or "深色主题" or '深色主题'
+  for (const m of trimmed.matchAll(/[「""'']([^「""'']{2,20})[」""'']?/g)) {
+    entityTokens.push(m[1])
+  }
+  // Book titles: 《记忆系统》
+  for (const m of trimmed.matchAll(/《([^》]+)》/g)) {
+    entityTokens.push(m[1])
+  }
+  // Capitalized English phrases: "TypeScript", "Visual Studio Code"
+  for (const m of trimmed.matchAll(/\b([A-Z][a-zA-Z]*(?:\s+[A-Z][a-zA-Z]*)+)\b/g)) {
+    entityTokens.push(m[1])
+  }
+
+  // Tokenize: split by spaces and Chinese character boundaries
+  const tokens: string[] = []
+  const parts = trimmed.split(/\s+/)
+  for (const part of parts) {
+    // English words
+    for (const word of part.match(/[a-zA-Z0-9_\-.]+/g) ?? []) {
+      if (word.length >= 2) tokens.push(word)
+    }
+    // Chinese characters (individual chars and 2-grams)
+    const cnChars = part.match(/[\u4e00-\u9fff]/g) ?? []
+    for (const c of cnChars) {
+      if (!CN_STOP_WORDS.has(c)) tokens.push(c)
+    }
+    // Chinese 2-grams for better matching
+    for (let i = 0; i < cnChars.length - 1; i++) {
+      const bigram = cnChars[i] + cnChars[i + 1]
+      tokens.push(bigram)
+    }
+  }
+
+  // Filter action words and stop words
+  const filtered = tokens.filter(t => {
+    if (ACTION_WORDS.has(t)) return false
+    if (CN_STOP_WORDS.has(t)) return false
+    if (t.length < 2) return false
+    return true
+  })
+
+  // Deduplicate while preserving order
+  const seen = new Set<string>()
+  const deduped: string[] = []
+  for (const t of filtered) {
+    if (!seen.has(t)) {
+      seen.add(t)
+      deduped.push(t)
+    }
+  }
+
+  // If nothing left after filtering, check if we have entity tokens
+  if (deduped.length === 0 && entityTokens.length === 0) {
+    return null
+  }
+
+  // Combine: entity tokens first (higher signal), then deduped tokens
+  const allTokens = [...entityTokens, ...deduped.filter(t => !entityTokens.includes(t))]
+  const query = allTokens.join(' ')
+
+  return { query, tokens: deduped, entityTokens }
+}
+```
+
+- [ ] **Step 2: Create `tests/refine.test.ts`**
+
+```typescript
+import { describe, it, expect } from 'vitest'
+import { refineQuery } from '../src/refine.js'
+
+describe('refineQuery', () => {
+  it('filters action words from Chinese query', () => {
+    const result = refineQuery('帮我用 TypeScript 重构 auth 模块')
+    expect(result).not.toBeNull()
+    expect(result!.query).toContain('TypeScript')
+    expect(result!.query).toContain('auth')
+    expect(result!.query).not.toContain('帮我')
+    expect(result!.query).not.toContain('重构')
+  })
+
+  it('returns null for pure operation commands', () => {
+    expect(refineQuery('运行测试')).toBeNull()
+    expect(refineQuery('git status')).toBeNull()
+    expect(refineQuery('创建文件')).toBeNull()
+  })
+
+  it('extracts quoted Chinese entities', () => {
+    const result = refineQuery('我喜欢「深色主题」')
+    expect(result).not.toBeNull()
+    expect(result!.entityTokens).toContain('深色主题')
+    expect(result!.query).toContain('深色主题')
+  })
+
+  it('extracts book title entities', () => {
+    const result = refineQuery('读了《设计模式》这本书')
+    expect(result).not.toBeNull()
+    expect(result!.entityTokens).toContain('设计模式')
+  })
+
+  it('extracts capitalized English phrases', () => {
+    const result = refineQuery('使用 Visual Studio Code 编辑器')
+    expect(result).not.toBeNull()
+    expect(result!.entityTokens).toContain('Visual Studio Code')
+  })
+
+  it('returns null for empty string', () => {
+    expect(refineQuery('')).toBeNull()
+    expect(refineQuery('   ')).toBeNull()
+  })
+
+  it('preserves meaningful Chinese tokens', () => {
+    const result = refineQuery('用户偏好深色主题')
+    expect(result).not.toBeNull()
+    expect(result!.query).toContain('用户')
+    expect(result!.query).toContain('偏好')
+    expect(result!.query).toContain('深色')
+    expect(result!.query).toContain('主题')
+  })
+})
+```
+
+- [ ] **Step 3: Run tests**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && npx vitest run tests/refine.test.ts
+```
+
+Expected: 7 tests PASS
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && git add src/refine.ts tests/refine.test.ts && git commit -m "feat(refine): add query refinement module with action word filtering"
+```
+
+---
+
+## Task 2: Resource Manager Module
+
+**Files:**
+- Create: `src/resources.ts`
+- Test: `tests/resource.test.ts`
+
+- [ ] **Step 1: Create `src/resources.ts`**
+
+```typescript
+/**
+ * MCP Resource manager for mnemo-mcp.
+ * Exposes per-category memory summaries as MCP Resources for session warmup injection.
+ */
+
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+import { ResourceTemplate } from '@modelcontextprotocol/sdk/server/mcp.js'
+import type { MemoryStore } from './store.js'
+import type { FactCategory } from './types.js'
+
+const CATEGORIES: FactCategory[] = ['identity', 'coding_style', 'tool_pref', 'workflow', 'general']
+const RESOURCE_LIMIT = 10
+
+export interface ResourceFact {
+  fact_id: number
+  content: string
+  trust_score: number
+}
+
+export class ResourceManager {
+  private cache = new Map<FactCategory, ResourceFact[]>()
+
+  constructor(
+    private server: McpServer,
+    private store: MemoryStore,
+  ) {}
+
+  /** Register all category resources with the MCP server */
+  registerResources(): void {
+    for (const category of CATEGORIES) {
+      const template = new ResourceTemplate(
+        `mnemo://global/{category}`,
+        { list: undefined },
+      )
+      this.server.registerResource(
+        `mnemo-global-${category}`,
+        template,
+        {
+          description: `${category} category global facts (top ${RESOURCE_LIMIT} by trust)`,
+          mimeType: 'application/json',
+        },
+        (uri, variables) => this.readResource(uri, variables),
+      )
+    }
+  }
+
+  /** Read handler for resource template */
+  private readResource(uri: URL, variables: Record<string, string | string[]>): { contents: Array<{ uri: string; mimeType: string; text: string }> } {
+    const category = (Array.isArray(variables.category) ? variables.category[0] : variables.category) as FactCategory
+    if (!CATEGORIES.includes(category)) {
+      return { contents: [{ uri: uri.toString(), mimeType: 'application/json', text: '[]' }] }
+    }
+
+    const facts = this.getFacts(category)
+    return {
+      contents: [{
+        uri: uri.toString(),
+        mimeType: 'application/json',
+        text: JSON.stringify(facts, null, 2),
+      }],
+    }
+  }
+
+  /** Get facts for a category — with caching */
+  getFacts(category: FactCategory): ResourceFact[] {
+    const cached = this.cache.get(category)
+    if (cached) return cached
+
+    const facts = this.store.listFacts(category, 0.0, RESOURCE_LIMIT).map(f => ({
+      fact_id: f.factId,
+      content: f.content,
+      trust_score: f.trustScore,
+    }))
+
+    this.cache.set(category, facts)
+    return facts
+  }
+
+  /** Invalidate all caches — call after any write operation */
+  invalidate(): void {
+    this.cache.clear()
+  }
+
+  /** Get cache size for debugging */
+  cacheSize(): number {
+    return this.cache.size
+  }
+}
+```
+
+- [ ] **Step 2: Create `tests/resource.test.ts`**
+
+```typescript
+import { describe, it, expect, beforeEach, afterEach } from 'vitest'
+import { MemoryStore } from '../src/store.js'
+import { ResourceManager } from '../src/resources.js'
+import { mkdtempSync, rmSync } from 'node:fs'
+import { join } from 'node:path'
+import { tmpdir } from 'node:os'
+import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+
+let store: MemoryStore
+let server: McpServer
+let manager: ResourceManager
+let tmpDir: string
+
+beforeEach(() => {
+  tmpDir = mkdtempSync(join(tmpdir(), 'mnemo-test-'))
+  store = new MemoryStore(join(tmpDir, 'test.db'))
+  server = new McpServer({ name: 'test', version: '0.1.0' })
+  manager = new ResourceManager(server, store)
+})
+
+afterEach(() => {
+  store.close()
+  rmSync(tmpDir, { recursive: true, force: true })
+})
+
+describe('ResourceManager', () => {
+  it('returns empty array for empty category', () => {
+    const facts = manager.getFacts('identity')
+    expect(facts).toEqual([])
+  })
+
+  it('returns facts ordered by trust score', () => {
+    store.addFact('用户偏好深色主题', 'tool_pref')
+    store.addFact('用户喜欢 VS Code', 'tool_pref')
+    const facts = manager.getFacts('tool_pref')
+    expect(facts.length).toBe(2)
+    expect(facts[0].content).toBe('用户偏好深色主题')
+  })
+
+  it('caches results', () => {
+    store.addFact('测试事实', 'general')
+    manager.getFacts('general')
+    expect(manager.cacheSize()).toBe(1)
+    // Second call should hit cache
+    const facts2 = manager.getFacts('general')
+    expect(facts2.length).toBe(1)
+  })
+
+  it('invalidates cache on write', () => {
+    store.addFact('测试事实', 'general')
+    manager.getFacts('general')
+    expect(manager.cacheSize()).toBe(1)
+    manager.invalidate()
+    expect(manager.cacheSize()).toBe(0)
+  })
+
+  it('limits to top 10 facts', () => {
+    for (let i = 0; i < 15; i++) {
+      store.addFact(`事实 ${i}`, 'general')
+    }
+    const facts = manager.getFacts('general')
+    expect(facts.length).toBe(10)
+  })
+})
+```
+
+- [ ] **Step 3: Run tests**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && npx vitest run tests/resource.test.ts
+```
+
+Expected: 5 tests PASS
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && git add src/resources.ts tests/resource.test.ts && git commit -m "feat(resources): add MCP Resource manager for per-category memory warmup"
+```
+
+---
+
+## Task 3: Update Retriever with Dynamic Scoring, Relevance Gate, and Deduplication
+
+**Files:**
+- Modify: `src/retriever.ts`
+- Modify: `src/types.ts`
+- Test: `tests/retriever.test.ts`
+
+- [ ] **Step 1: Update `src/types.ts` — add `RefineResult`**
+
+Add to `src/types.ts` after `SecurityScanResult`:
+
+```typescript
+/** 查询提炼结果 */
+export interface RefineResult {
+  query: string | null
+  tokens: string[]
+  entityTokens: string[]
+}
+```
+
+- [ ] **Step 2: Modify `src/retriever.ts` — integrate refineQuery, dynamic weights, relevance gate, dedup**
+
+Replace the `search()` method (lines 68-171) with:
+
+```typescript
+  /** 主搜索：FTS5 → LIKE → 字符交叉 → 分类推断 → Jaccard → 信任评分 → 时间衰减 */
+  search(query: string, options?: SearchOptions & { skipRefine?: boolean }): ScoredFact[] {
+    const startTime = performance.now()
+    const minTrust = options?.minTrust ?? 0.3
+    const limit = options?.limit ?? 10
+    const category = options?.category
+
+    // Query refinement (unless explicitly skipped)
+    let searchQuery = query
+    if (!options?.skipRefine) {
+      const refined = refineQuery(query)
+      if (refined?.query) {
+        searchQuery = refined.query
+      }
+    }
+
+    // Cache check using refined query
+    const cacheKey = this.cache.makeKey({ action: 'search', query: searchQuery, category, minTrust, limit })
+    const cached = this.cache.get(cacheKey)
+    if (cached) {
+      this.metrics.record({ action: 'search', durationMs: performance.now() - startTime, resultCount: cached.length, cacheHit: true })
+      return cached
+    }
+
+    // Query bilingual expansion
+    const expandedQuery = this.expandQueryBilingually(searchQuery)
+
+    // Stage 1: FTS5 candidates with fallback chain
+    let candidates = this.ftsCandidates(expandedQuery, category, minTrust, limit * 3)
+    if (candidates.length === 0) candidates = this.likeFallback(expandedQuery, category, minTrust, limit * 3)
+    if (candidates.length === 0) candidates = this.charOverlapFallback(expandedQuery, category, minTrust, limit * 3)
+    if (candidates.length === 0) {
+      if (!category) {
+        const inferred = this.categoryInferFallback(searchQuery, minTrust, limit)
+        if (inferred.length > 0) return inferred
+      }
+      if (this.isPersonalQuery(searchQuery)) {
+        return this.trustFallback(category, minTrust, limit)
+      }
+      return []
+    }
+
+    // Dynamic weighting based on query token count
+    const queryTokens = this.tokenize(searchQuery)
+    const tokenCount = queryTokens.size
+    const ftsWeight = tokenCount <= 3 ? 0.7 : 0.3
+    const jaccardWeight = tokenCount <= 3 ? 0.3 : 0.7
+
+    // Stage 2-4: Score candidates
+    const scored: ScoredFact[] = []
+    for (const fact of candidates) {
+      const contentTokens = this.tokenize(fact.content)
+      const tagTokens = this.tokenize(fact.tags)
+      const allTokens = new Set([...contentTokens, ...tagTokens])
+
+      const jaccard = this.jaccardSimilarity(queryTokens, allTokens)
+      const qInF = this.containmentScore(queryTokens, allTokens)
+      const similarity = 0.3 * jaccard + 0.7 * qInF
+      const ftsScore = fact.ftsRank
+
+      // Dynamic relevance
+      const relevance = ftsWeight * ftsScore + jaccardWeight * similarity
+      let score = relevance * fact.trustScore
+
+      if (this.halfLifeDays > 0) {
+        score *= this.temporalDecay(fact.updatedAt || fact.createdAt)
+      }
+
+      scored.push({ ...fact, score })
+    }
+
+    scored.sort((a, b) => b.score - a.score)
+
+    // Relevance gate: filter out low-relevance results
+    const RELEVANCE_THRESHOLD = 0.15
+    const gated = scored.filter(s => s.score >= RELEVANCE_THRESHOLD)
+    const resultsToDedup = gated.length > 0 ? gated : scored
+
+    // Content-based deduplication (Jaccard > 0.7) instead of category-per-top1
+    const deduped: ScoredFact[] = []
+    for (const candidate of resultsToDedup) {
+      let isDuplicate = false
+      const candidateTokens = this.tokenize(candidate.content)
+      for (const kept of deduped) {
+        const keptTokens = this.tokenize(kept.content)
+        if (this.jaccardSimilarity(candidateTokens, keptTokens) > 0.7) {
+          isDuplicate = true
+          break
+        }
+      }
+      if (!isDuplicate) {
+        deduped.push(candidate)
+        if (deduped.length >= limit) break
+      }
+    }
+
+    const results = deduped
+
+    // Track retrieval
+    if (results.length > 0) {
+      this.trackRetrieval(results)
+    }
+
+    this.cache.set(cacheKey, results)
+    this.metrics.record({ action: 'search', durationMs: performance.now() - startTime, resultCount: results.length, cacheHit: false, retrievalPath: 'FTS5' })
+    return results
+  }
+```
+
+Add import at top of `src/retriever.ts`:
+
+```typescript
+import { refineQuery } from './refine.js'
+```
+
+- [ ] **Step 3: Update `tests/retriever.test.ts` — add new test cases**
+
+Append to existing test file:
+
+```typescript
+  describe('dynamic scoring', () => {
+    it('uses high FTS weight for short queries', () => {
+      store.addFact('用户偏好深色主题', 'tool_pref')
+      const results = retriever.search('深色主题')
+      expect(results.length).toBeGreaterThan(0)
+    })
+
+    it('uses high Jaccard weight for long queries', () => {
+      store.addFact('用户偏好使用 TypeScript 开发后端 API', 'coding_style')
+      const results = retriever.search('为什么 TypeScript 编译报错找不到模块')
+      // Should still return something due to Jaccard overlap on "TypeScript"
+      expect(results.some(r => r.content.includes('TypeScript'))).toBe(true)
+    })
+  })
+
+  describe('relevance gate', () => {
+    it('filters out low relevance results', () => {
+      store.addFact('完全不相关的内容关于天气和食物', 'general')
+      store.addFact('用户偏好深色主题', 'tool_pref')
+      const results = retriever.search('深色主题')
+      expect(results.every(r => r.content.includes('深色') || r.content.includes('主题'))).toBe(true)
+    })
+  })
+
+  describe('content deduplication', () => {
+    it('allows multiple general facts if content differs', () => {
+      store.addFact('用户喜欢蓝色', 'general')
+      store.addFact('用户偏好深色主题', 'general')
+      store.addFact('用户使用 VS Code', 'general')
+      const results = retriever.search('用户偏好')
+      expect(results.filter(r => r.category === 'general').length).toBeGreaterThan(1)
+    })
+
+    it('deduplicates highly similar facts', () => {
+      store.addFact('用户偏好深色主题', 'tool_pref')
+      store.addFact('用户偏好深色主题和蓝色', 'tool_pref')
+      const results = retriever.search('深色主题')
+      // Both are similar but not identical; Jaccard should be high
+      const similarCount = results.filter(r => r.content.includes('深色主题')).length
+      expect(similarCount).toBeLessThanOrEqual(2)
+    })
+  })
+```
+
+- [ ] **Step 4: Run all retriever tests**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && npx vitest run tests/retriever.test.ts
+```
+
+Expected: All tests PASS (existing + new)
+
+- [ ] **Step 5: Commit**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && git add src/retriever.ts src/types.ts tests/retriever.test.ts && git commit -m "feat(retriever): dynamic scoring, relevance gate, content dedup, query refinement"
+```
+
+---
+
+## Task 4: Wire Resources and Refinement into Server
+
+**Files:**
+- Modify: `src/server.ts`
+
+- [ ] **Step 1: Update `src/server.ts` — import and wire ResourceManager**
+
+Add imports:
+
+```typescript
+import { ResourceManager } from './resources.js'
+```
+
+After retriever initialization, add:
+
+```typescript
+const resourceManager = new ResourceManager(server, store)
+resourceManager.registerResources()
+```
+
+In write operations (add/update/remove cases), after `retriever.getCache().clear()`, add:
+
+```typescript
+resourceManager.invalidate()
+```
+
+For the `search` case, integrate refineQuery:
+
+```typescript
+case 'search': {
+  if (!a.query) return { content: [{ type: 'text' as const, text: JSON.stringify({ error: 'Missing required argument: query' }) }] }
+  // Skip refinement for explicit tool calls (user knows what they're searching)
+  const results = retriever.search(a.query, { category: a.category ? category : undefined, minTrust: a.min_trust ?? minTrust, limit: a.limit ?? 10, skipRefine: true })
+  return { content: [{ type: 'text' as const, text: JSON.stringify({ results, count: results.length }) }] }
+}
+```
+
+- [ ] **Step 2: Verify build**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && npm run build
+```
+
+Expected: BUILD OK (no TypeScript errors)
+
+- [ ] **Step 3: Run all tests**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && npm run test
+```
+
+Expected: All tests PASS
+
+- [ ] **Step 4: Commit**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && git add src/server.ts && git commit -m "feat(server): wire ResourceManager and query refinement into MCP server"
+```
+
+---
+
+## Task 5: Update CLAUDE.md Rules and Documentation
+
+**Files:**
+- Modify: `README.md` (in project)
+- Provide: CLAUDE.md rule template (for user to copy)
+
+- [ ] **Step 1: Update `README.md` — add Resource and injection protocol sections**
+
+Add after the existing Tools section:
+
+```markdown
+## MCP Resources
+
+mnemo-mcp exposes 5 global category resources for session warmup injection:
+
+| Resource URI | Description |
+|-------------|-------------|
+| `mnemo://global/identity` | Identity facts (top 10 by trust) |
+| `mnemo://global/coding_style` | Coding style preferences |
+| `mnemo://global/tool_pref` | Tool preferences |
+| `mnemo://global/workflow` | Workflow preferences |
+| `mnemo://global/general` | General facts |
+
+Clients (Claude Code / Codex) automatically fetch these resources at session start,
+injecting memory into system context without any tool calls.
+
+## Injection Protocol
+
+### For Claude Code Users
+
+Update your `CLAUDE.md` rules to use the new protocol:
+
+```markdown
+# 记忆系统使用规则
+
+你有 mnemo 记忆工具（fact_store / fact_feedback），必须按以下规则使用：
+
+## 规则 1：会话预热（自动）
+会话启动时，mnemo-mcp 的 MCP Resource 会自动注入全局记忆到 system context。
+你不需要主动调用 fact_store(search) 来获取高频记忆。
+
+## 规则 2：按需补充查询
+仅在以下情况调用 fact_store(action="search")：
+- 用户消息涉及个人偏好/习惯/工具选择且预热中未覆盖
+- 用户明确查询记忆（"我之前说过什么""按我的习惯"）
+- 技术选型时需要确认用户偏好
+
+不触发查询的情况：
+- 纯操作指令（"运行测试""git commit"）
+- 通用技术问题（"Promise 怎么用"）
+- 代码审查/解释请求
+
+## 规则 3：写入记忆
+用户说"记住""记下来"时，调用 fact_store(action="add", content="...", category="...")。
+- 先 search 检查是否已有相似事实，有则 update
+- category：identity / coding_style / tool_pref / workflow / general
+
+## 规则 4：反馈强化
+成功使用某条记忆时，调用 fact_feedback(action="helpful", fact_id=...)。
+```
+```
+
+- [ ] **Step 2: Commit docs**
+
+```bash
+cd /Users/ljn/Documents/demo/ocean/mnemo-mcp && git add README.md && git commit -m "docs: add MCP Resource and injection protocol documentation"
+```
+
+---
+
+## Self-Review
+
+### Spec Coverage
+
+| Spec Requirement | Task |
+|-----------------|------|
+| MCP Resource 暴露 5 个 category | Task 2 |
+| Resource 缓存 + 写操作失效 | Task 2 |
+| 查询提炼过滤动作词/虚词 | Task 1 |
+| 纯操作指令返回 null | Task 1 |
+| 引号/书名号实体提取 | Task 1 |
+| 动态 FTS/Jaccard 权重 | Task 3 |
+| 相关性评分门控 (0.15) | Task 3 |
+| 内容相似度去重 (Jaccard > 0.7) | Task 3 |
+| 会话预热注入协议 | Task 4 (Resource wiring) + Task 5 (docs) |
+| 按需补充查询规则 | Task 5 (CLAUDE.md template) |
+
+### Placeholder Scan
+
+- [x] No "TBD", "TODO", "implement later"
+- [x] No vague "add error handling" without code
+- [x] No "write tests for the above" without test code
+- [x] All file paths are exact
+- [x] All code blocks contain complete implementations
+
+### Type Consistency
+
+- [x] `refineQuery()` returns `RefineResult | null` consistently
+- [x] `ResourceManager.getFacts()` returns `ResourceFact[]`
+- [x] `search()` accepts `skipRefine?: boolean` in options
+- [x] Dynamic weights use `ftsWeight`/`jaccardWeight` variables (same names as constructor params)