@morningljn/mnemo 0.1.4 → 0.2.0

package/src/retriever.ts

@@ -106,38 +106,32 @@ export class FactRetriever {
         const inferred = this.categoryInferFallback(searchQuery, minTrust, limit)
         if (inferred.length > 0) return inferred
       }
-      // 个人/身份相关的短查询触发 trust fallback
-      if (this.isPersonalQuery(searchQuery)) {
+      // 个人/身份相关的短查询触发 trust fallback（用原始 query，避免 refineQuery 拆词导致正则失配）
+      if (this.isPersonalQuery(query)) {
         return this.trustFallback(category, minTrust, limit)
       }
       return []
     }
 
-    // Stage 2-4: Jaccard 重排序 + 信任评分 + 时间衰减
-    // 动态权重：短查询偏 FTS，长查询偏 Jaccard
+    // Stage 2-4: Jaccard 重排序 + 信任评分 + 时间衰减 + length penalty
     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
 
     const scored: ScoredFact[] = []
 
     for (const fact of candidates) {
-      const contentTokens = this.tokenize(fact.content)
+      // summary 优先用于匹配
+      const matchText = fact.summary ?? fact.content
+      const matchTokens = this.tokenize(matchText)
       const tagTokens = this.tokenize(fact.tags)
-      const allTokens = new Set([...contentTokens, ...tagTokens])
+      const allTokens = new Set([...matchTokens, ...tagTokens])
 
       const jaccard = this.jaccardSimilarity(queryTokens, allTokens)
-      // Containment: 查询 token 在事实 token 中的覆盖率
       const qInF = this.containmentScore(queryTokens, allTokens)
-
-      // 混合相似度：Jaccard + Containment（简化版，移除 keywordScore）
       const similarity = 0.3 * jaccard + 0.7 * qInF
       const ftsScore = fact.ftsRank
 
-      // 综合评分
-      const relevance = ftsWeight * ftsScore + jaccardWeight * similarity
-
+      // 静态权重 0.5/0.5（回退 v3 动态权重）
+      const relevance = 0.5 * ftsScore + 0.5 * similarity
       let score = relevance * fact.trustScore
 
       // 时间衰减
@@ -145,37 +139,22 @@ export class FactRetriever {
         score *= this.temporalDecay(fact.updatedAt || fact.createdAt)
       }
 
+      // Length penalty：基于 matchText 长度
+      score *= Math.min(1.0, 300 / matchText.length)
+
       scored.push({ ...fact, score })
     }
 
     scored.sort((a, b) => b.score - a.score)
 
-    // 相关性门控：过滤低相关性结果
-    const RELEVANCE_THRESHOLD = 0.15
-    const gated = scored.filter(s => s.score >= RELEVANCE_THRESHOLD)
-    const pool = gated.length > 0 ? gated : scored
-
-    // 内容去重：Jaccard > 0.7 的只保留高分
-    const results: ScoredFact[] = []
-    for (const candidate of pool) {
-      let isDuplicate = false
-      const candidateTokens = this.tokenize(candidate.content)
-      for (const kept of results) {
-        const keptTokens = this.tokenize(kept.content)
-        if (this.jaccardSimilarity(candidateTokens, keptTokens) > 0.7) {
-          isDuplicate = true
-          break
-        }
-      }
-      if (!isDuplicate) {
-        results.push(candidate)
-        if (results.length >= limit) break
-      }
-    }
+    // 取 limit 条（不再做 relevance gate 和 content dedup）
+    const results = scored.slice(0, limit)
 
     // 检索追踪：递增 retrieval_count + top3 信任刷新
     if (results.length > 0) {
       this.trackRetrieval(results)
+      // 记录检索日志
+      this.store.logRetrieval(searchQuery, results.map(r => ({ id: r.factId, score: Math.round(r.score * 1000) / 1000 })))
     }
 
     // 缓存存储 + 指标记录
@@ -295,9 +274,11 @@ export class FactRetriever {
       category: r.category as FactCategory,
       tags: r.tags,
       keywords: r.keywords ?? '[]',
+      summary: (r as any).summary ?? null,
       trustScore: r.trust_score,
       retrievalCount: r.retrieval_count,
       helpfulCount: r.helpful_count,
+      lastRetrievedAt: (r as any).last_retrieved_at ?? null,
       createdAt: r.created_at,
       updatedAt: r.updated_at,
       score: r.trust_score * (1 - i * 0.05),
@@ -409,9 +390,11 @@ export class FactRetriever {
             category: r.category as FactCategory,
             tags: r.tags,
             keywords: r.keywords ?? '[]',
+            summary: (r as any).summary ?? null,
             trustScore: r.trust_score,
             retrievalCount: 0,
             helpfulCount: 0,
+            lastRetrievedAt: (r as any).last_retrieved_at ?? null,
             createdAt: r.created_at,
             updatedAt: r.updated_at,
           })
@@ -509,9 +492,11 @@ export class FactRetriever {
       category: String(row.category) as FactCategory,
       tags: String(row.tags),
       keywords: String(row.keywords ?? '[]'),
+      summary: row.summary != null ? String(row.summary) : null,
       trustScore: Number(row.trust_score),
       retrievalCount: Number(row.retrieval_count),
       helpfulCount: Number(row.helpful_count),
+      lastRetrievedAt: row.last_retrieved_at != null ? String(row.last_retrieved_at) : null,
       createdAt: String(row.created_at),
       updatedAt: String(row.updated_at),
       ftsRank: rawRanks[i] / maxRank,
@@ -619,8 +604,8 @@ export class FactRetriever {
     const conditions: string[] = []
     const params: unknown[] = []
     for (const word of words) {
-      conditions.push('(f.content LIKE ? OR f.tags LIKE ?)')
-      params.push(`%${word}%`, `%${word}%`)
+      conditions.push('(f.content LIKE ? OR f.tags LIKE ? OR f.summary LIKE ?)')
+      params.push(`%${word}%`, `%${word}%`, `%${word}%`)
     }
 
     // 中文子串分解：将中文查询拆为 2~3 字滑动窗口，追加 LIKE 条件
@@ -632,14 +617,14 @@ export class FactRetriever {
         // 2-gram
         for (let i = 0; i < seg.length - 1; i++) {
           const bigram = seg.slice(i, i + 2)
-          conditions.push('(f.content LIKE ? OR f.tags LIKE ?)')
-          params.push(`%${bigram}%`, `%${bigram}%`)
+          conditions.push('(f.content LIKE ? OR f.tags LIKE ? OR f.summary LIKE ?)')
+          params.push(`%${bigram}%`, `%${bigram}%`, `%${bigram}%`)
         }
         // 3-gram（覆盖更长的短语匹配）
         for (let i = 0; i < seg.length - 2; i++) {
           const trigram = seg.slice(i, i + 3)
-          conditions.push('(f.content LIKE ? OR f.tags LIKE ?)')
-          params.push(`%${trigram}%`, `%${trigram}%`)
+          conditions.push('(f.content LIKE ? OR f.tags LIKE ? OR f.summary LIKE ?)')
+          params.push(`%${trigram}%`, `%${trigram}%`, `%${trigram}%`)
         }
       }
     }
@@ -656,7 +641,7 @@ export class FactRetriever {
 
     const sql = `
       SELECT f.fact_id, f.content, f.category, f.tags, f.keywords,
-             f.trust_score, f.retrieval_count, f.helpful_count,
+             f.summary, f.trust_score, f.retrieval_count, f.helpful_count,
              f.created_at, f.updated_at
       FROM facts f
       WHERE (${conditionsSql})
@@ -668,6 +653,7 @@ export class FactRetriever {
 
     const rows = this.db.prepare(sql).all(...params) as Array<{
       fact_id: number; content: string; category: string; tags: string; keywords: string;
+      summary: string | null;
       trust_score: number; retrieval_count: number; helpful_count: number;
       created_at: string; updated_at: string;
     }>
@@ -679,9 +665,11 @@ export class FactRetriever {
       category: r.category as FactCategory,
       tags: r.tags,
       keywords: r.keywords ?? '[]',
+      summary: r.summary ?? null,
       trustScore: r.trust_score,
       retrievalCount: r.retrieval_count,
       helpfulCount: r.helpful_count,
+      lastRetrievedAt: (r as any).last_retrieved_at ?? null,
       createdAt: r.created_at,
       updatedAt: r.updated_at,
       ftsRank: 0.5,

package/src/schema.ts

@@ -6,9 +6,11 @@ CREATE TABLE IF NOT EXISTS facts (
   category        TEXT DEFAULT 'general',
   tags            TEXT DEFAULT '',
   keywords        TEXT DEFAULT '[]',
+  summary         TEXT DEFAULT NULL,
   trust_score     REAL DEFAULT 0.5,
   retrieval_count INTEGER DEFAULT 0,
   helpful_count   INTEGER DEFAULT 0,
+  last_retrieved_at TEXT DEFAULT NULL,
   created_at      TEXT DEFAULT (datetime('now', 'localtime')),
   updated_at      TEXT DEFAULT (datetime('now', 'localtime'))
 );
@@ -29,33 +31,42 @@ CREATE TABLE IF NOT EXISTS fact_entities (
   PRIMARY KEY (fact_id, entity_id)
 );
 
+-- 检索日志表
+CREATE TABLE IF NOT EXISTS retrieval_log (
+  id        INTEGER PRIMARY KEY AUTOINCREMENT,
+  query     TEXT NOT NULL,
+  results   TEXT DEFAULT NULL,
+  timestamp TEXT DEFAULT (datetime('now', 'localtime'))
+);
+
 -- 索引
 CREATE INDEX IF NOT EXISTS idx_facts_trust    ON facts(trust_score DESC);
 CREATE INDEX IF NOT EXISTS idx_facts_category ON facts(category);
 CREATE INDEX IF NOT EXISTS idx_entities_name  ON entities(name);
 CREATE INDEX IF NOT EXISTS idx_fact_entities_entity ON fact_entities(entity_id);
+CREATE INDEX IF NOT EXISTS idx_retrieval_log_ts ON retrieval_log(timestamp);
 
--- FTS5 全文索引
+-- FTS5 全文索引（含 summary 列）
 CREATE VIRTUAL TABLE IF NOT EXISTS facts_fts
-  USING fts5(content, tags, content=facts, content_rowid=fact_id);
+  USING fts5(content, tags, summary, content=facts, content_rowid=fact_id);
 
 -- FTS5 同步触发器：插入
 CREATE TRIGGER IF NOT EXISTS facts_ai AFTER INSERT ON facts BEGIN
-  INSERT INTO facts_fts(rowid, content, tags)
-    VALUES (new.fact_id, new.content, new.tags);
+  INSERT INTO facts_fts(rowid, content, tags, summary)
+    VALUES (new.fact_id, new.content, new.tags, COALESCE(new.summary, ''));
 END;
 
 -- FTS5 同步触发器：删除
 CREATE TRIGGER IF NOT EXISTS facts_ad AFTER DELETE ON facts BEGIN
-  INSERT INTO facts_fts(facts_fts, rowid, content, tags)
-    VALUES ('delete', old.fact_id, old.content, old.tags);
+  INSERT INTO facts_fts(facts_fts, rowid, content, tags, summary)
+    VALUES ('delete', old.fact_id, old.content, old.tags, COALESCE(old.summary, ''));
 END;
 
 -- FTS5 同步触发器：更新
 CREATE TRIGGER IF NOT EXISTS facts_au AFTER UPDATE ON facts BEGIN
-  INSERT INTO facts_fts(facts_fts, rowid, content, tags)
-    VALUES ('delete', old.fact_id, old.content, old.tags);
-  INSERT INTO facts_fts(rowid, content, tags)
-    VALUES (new.fact_id, new.content, new.tags);
+  INSERT INTO facts_fts(facts_fts, rowid, content, tags, summary)
+    VALUES ('delete', old.fact_id, old.content, old.tags, COALESCE(old.summary, ''));
+  INSERT INTO facts_fts(rowid, content, tags, summary)
+    VALUES (new.fact_id, new.content, new.tags, COALESCE(new.summary, ''));
 END;
 `

package/src/server.ts

@@ -27,8 +27,9 @@ const FACT_STORE_DESCRIPTION = `结构化事实记忆系统（SQLite+FTS5 索引
 写入时先 search 检查是否已存在相似事实。identity/coding_style/tool_pref/workflow/general → 全局库，project → 项目库。`
 
 const factStoreSchema = {
-  action: z.enum(['add', 'search', 'probe', 'related', 'reason', 'contradict', 'update', 'remove', 'list']),
+  action: z.enum(['add', 'search', 'probe', 'related', 'reason', 'contradict', 'update', 'remove', 'list', 'learn', 'audit']),
   content: z.union([z.string(), z.array(z.string())]).optional().describe("事实内容（'add' 必需，支持批量）"),
+  summary: z.string().optional().describe('超长事实的摘要（检索用 summary 匹配）'),
   query: z.string().optional().describe("搜索查询（'search' 必需）"),
   entity: z.string().optional().describe("实体名（'probe'/'related' 使用）"),
   entities: z.array(z.string()).optional().describe("实体列表（'reason' 使用）"),
@@ -62,6 +63,18 @@ const retriever = new FactRetriever(store, { temporalDecayHalfLife: 30 })
 store.decayTrustScores()
 store.auditContradictions()
 
+// Auto-learn on startup (non-blocking)
+process.nextTick(() => {
+  try {
+    const result = store.runLearning()
+    if (result.demoted > 0 || result.aged > 0 || result.long_facts.length > 0) {
+      console.error(`[mnemo:auto-learn] promoted=${result.promoted} demoted=${result.demoted} aged=${result.aged} long_facts=${result.long_facts.length}`)
+    }
+  } catch (err) {
+    console.error('[mnemo:auto-learn] error:', err)
+  }
+})
+
 // -- MCP Server --
 const server = new McpServer({ name: 'mnemo-mcp', version: '0.1.0' })
 
@@ -93,13 +106,22 @@ server.tool(
             let warnings: string[] | undefined
             const scan = fullSecurityScan(content)
             if (scan.warnings.length > 0 || scan.hasPii) warnings = [...scan.warnings]
+            if (content.length > 500 && !a.summary) {
+              warnings = [...(warnings ?? []), 'content 超过 500 字，建议提供 summary 或拆分为多条 fact']
+            }
 
             if (similar) {
               store.updateFact(similar.factId, { content, tags: a.tags, trustDelta: 0.05 })
+              if (a.summary) {
+                store.connection.prepare('UPDATE facts SET summary = ? WHERE fact_id = ?').run(a.summary, similar.factId)
+              }
               const demoted = store.demoteContradictingFacts(similar.factId, content, category)
               results.push({ fact_id: similar.factId, status: 'updated', reason: 'similar_fact_merged', ...(demoted > 0 ? { contradicted_demoted: demoted } : {}), ...(warnings ? { warnings } : {}) })
             } else {
               const factId = store.addFact(content, category, a.tags ?? '')
+              if (a.summary) {
+                store.connection.prepare('UPDATE facts SET summary = ? WHERE fact_id = ?').run(a.summary, factId)
+              }
               const demoted = store.demoteContradictingFacts(factId, content, category)
               results.push({ fact_id: factId, status: 'added', category, ...(demoted > 0 ? { contradicted_demoted: demoted } : {}), ...(warnings ? { warnings } : {}) })
             }
@@ -144,6 +166,9 @@ server.tool(
         case 'update': {
           if (!a.fact_id) return { content: [{ type: 'text' as const, text: JSON.stringify({ error: 'Missing required argument: fact_id' }) }] }
           const updated = store.updateFact(a.fact_id as number, { content: a.content as string | undefined, tags: a.tags, category, trustDelta: a.trust_delta })
+          if (a.summary !== undefined) {
+            store.connection.prepare('UPDATE facts SET summary = ? WHERE fact_id = ?').run(a.summary, a.fact_id as number)
+          }
           retriever.getCache().clear()
           resourceManager.invalidate()
           return { content: [{ type: 'text' as const, text: JSON.stringify({ updated }) }] }
@@ -159,6 +184,16 @@ server.tool(
           return { content: [{ type: 'text' as const, text: JSON.stringify(response) }] }
         }
 
+        case 'learn': {
+          const result = store.runLearning()
+          return { content: [{ type: 'text' as const, text: JSON.stringify(result) }] }
+        }
+
+        case 'audit': {
+          const report = store.runAudit()
+          return { content: [{ type: 'text' as const, text: JSON.stringify(report) }] }
+        }
+
         case 'list': {
           const facts = store.listFacts(category, a.min_trust ?? 0.0, a.limit ?? 10)
           return { content: [{ type: 'text' as const, text: JSON.stringify({ facts, count: facts.length }) }] }

package/src/store.ts

@@ -63,9 +63,11 @@ interface FactRow {
   category: string
   tags: string
   keywords: string
+  summary?: string | null
   trust_score: number
   retrieval_count: number
   helpful_count: number
+  last_retrieved_at?: string | null
   created_at: string
   updated_at: string
 }
@@ -110,10 +112,58 @@ export class MemoryStore {
 
   /** 增量迁移：添加新列（已存在则跳过） */
   private migrateSchema(): void {
+    const addColumn = (table: string, column: string, def: string): void => {
+      try {
+        this.db.exec(`ALTER TABLE ${table} ADD COLUMN ${column} ${def}`)
+      } catch { /* 列已存在 */ }
+    }
+
     // keywords 列（v2）
-    try {
-      this.db.exec('ALTER TABLE facts ADD COLUMN keywords TEXT DEFAULT \'[]\'')
-    } catch { /* 列已存在 */ }
+    addColumn('facts', 'keywords', "TEXT DEFAULT '[]'")
+    // summary 列
+    addColumn('facts', 'summary', 'TEXT DEFAULT NULL')
+    // last_retrieved_at 列
+    addColumn('facts', 'last_retrieved_at', 'TEXT DEFAULT NULL')
+
+    // FTS5 重建：检查 facts_fts 是否包含 summary 列
+    const ftsCols = this.db.pragma('table_info(facts_fts)') as Array<{ name: string }>
+    const hasSummary = ftsCols.some(c => c.name === 'summary')
+    if (!hasSummary) {
+      // DROP 旧 FTS5 表和触发器，重建含 summary 的新版本
+      this.db.exec(`
+        DROP TABLE IF EXISTS facts_fts;
+        DROP TRIGGER IF EXISTS facts_ai;
+        DROP TRIGGER IF EXISTS facts_ad;
+        DROP TRIGGER IF EXISTS facts_au;
+      `)
+      // 重建 FTS5 虚拟表（含 summary）
+      this.db.exec(`
+        CREATE VIRTUAL TABLE facts_fts
+          USING fts5(content, tags, summary, content=facts, content_rowid=fact_id);
+      `)
+      // 重填充
+      this.db.exec(`
+        INSERT INTO facts_fts(rowid, content, tags, summary)
+          SELECT fact_id, content, tags, COALESCE(summary, '') FROM facts;
+      `)
+      // 重建触发器
+      this.db.exec(`
+        CREATE TRIGGER facts_ai AFTER INSERT ON facts BEGIN
+          INSERT INTO facts_fts(rowid, content, tags, summary)
+            VALUES (new.fact_id, new.content, new.tags, COALESCE(new.summary, ''));
+        END;
+        CREATE TRIGGER facts_ad AFTER DELETE ON facts BEGIN
+          INSERT INTO facts_fts(facts_fts, rowid, content, tags, summary)
+            VALUES ('delete', old.fact_id, old.content, old.tags, COALESCE(old.summary, ''));
+        END;
+        CREATE TRIGGER facts_au AFTER UPDATE ON facts BEGIN
+          INSERT INTO facts_fts(facts_fts, rowid, content, tags, summary)
+            VALUES ('delete', old.fact_id, old.content, old.tags, COALESCE(old.summary, ''));
+          INSERT INTO facts_fts(rowid, content, tags, summary)
+            VALUES (new.fact_id, new.content, new.tags, COALESCE(new.summary, ''));
+        END;
+      `)
+    }
   }
 
   private prepareStatements(): void {
@@ -361,8 +411,8 @@ export class MemoryStore {
     params.push(limit)
 
     const sql = `
-      SELECT fact_id, content, category, tags, keywords, trust_score,
-             retrieval_count, helpful_count, created_at, updated_at
+      SELECT fact_id, content, category, tags, keywords, summary, trust_score,
+             retrieval_count, helpful_count, last_retrieved_at, created_at, updated_at
       FROM facts
       WHERE trust_score >= ?
         ${categoryClause}
@@ -410,8 +460,8 @@ export class MemoryStore {
     params.push(limit)
 
     const sql = `
-      SELECT f.fact_id, f.content, f.category, f.tags, f.keywords, f.trust_score,
-             f.retrieval_count, f.helpful_count, f.created_at, f.updated_at
+      SELECT f.fact_id, f.content, f.category, f.tags, f.keywords, f.summary, f.trust_score,
+             f.retrieval_count, f.helpful_count, f.last_retrieved_at, f.created_at, f.updated_at
       FROM facts f
       JOIN fact_entities fe ON f.fact_id = fe.fact_id
       JOIN entities e ON fe.entity_id = e.entity_id
@@ -445,8 +495,8 @@ export class MemoryStore {
     params.push(limit)
 
     const sql = `
-      SELECT f.fact_id, f.content, f.category, f.tags, f.keywords, f.trust_score,
-             f.retrieval_count, f.helpful_count, f.created_at, f.updated_at
+      SELECT f.fact_id, f.content, f.category, f.tags, f.keywords, f.summary, f.trust_score,
+             f.retrieval_count, f.helpful_count, f.last_retrieved_at, f.created_at, f.updated_at
       FROM facts f
       WHERE f.fact_id IN (${intersects})
         ${categoryClause}
@@ -635,6 +685,160 @@ export class MemoryStore {
     return row.count
   }
 
+  /** 记录检索日志并更新 last_retrieved_at */
+  logRetrieval(query: string, results: Array<{ id: number; score: number }>): void {
+    const resultsJson = JSON.stringify(results)
+    this.db.prepare(
+      "INSERT INTO retrieval_log (query, results) VALUES (?, ?)"
+    ).run(query, resultsJson)
+
+    // 更新返回 fact 的 last_retrieved_at
+    if (results.length > 0) {
+      const ids = results.map(r => r.id)
+      const placeholders = ids.map(() => '?').join(',')
+      this.db.prepare(
+        `UPDATE facts SET last_retrieved_at = datetime('now', 'localtime') WHERE fact_id IN (${placeholders})`
+      ).run(...ids)
+    }
+
+    // 自动清理日志
+    this.pruneRetrievalLog(5000)
+  }
+
+  /** 清理检索日志，保留最近 maxEntries 条 */
+  pruneRetrievalLog(maxEntries = 5000): void {
+    this.db.prepare(
+      `DELETE FROM retrieval_log WHERE id NOT IN (
+        SELECT id FROM retrieval_log ORDER BY id DESC LIMIT ?
+      )`
+    ).run(maxEntries)
+  }
+
+  /** 自学习：基于检索统计自动调整 trust_score */
+  runLearning(): {
+    promoted: number
+    demoted: number
+    aged: number
+    unchanged: number
+    long_facts: Array<{ id: number; content_length: number; penalty: number; has_summary: boolean }>
+  } {
+    const rows = this.db.prepare(
+      'SELECT fact_id, content, summary, retrieval_count, helpful_count, trust_score, last_retrieved_at FROM facts'
+    ).all() as Array<{
+      fact_id: number; content: string; summary: string | null;
+      retrieval_count: number; helpful_count: number; trust_score: number; last_retrieved_at: string | null
+    }>
+
+    let promoted = 0
+    let demoted = 0
+    let aged = 0
+    let unchanged = 0
+    const longFacts: Array<{ id: number; content_length: number; penalty: number; has_summary: boolean }> = []
+
+    const now = Date.now()
+
+    for (const row of rows) {
+      let changed = false
+      const rate = row.retrieval_count > 0 ? row.helpful_count / row.retrieval_count : 0
+
+      // Rate-based adjustment (需要 30+ 次检索)
+      if (row.retrieval_count > 30) {
+        if (rate < 0.05) {
+          const newTrust = clampTrust(row.trust_score * 0.9)
+          this.db.prepare('UPDATE facts SET trust_score = ? WHERE fact_id = ?').run(newTrust, row.fact_id)
+          demoted++
+          changed = true
+        } else if (rate > 0.3) {
+          const newTrust = clampTrust(row.trust_score + 0.05)
+          this.db.prepare('UPDATE facts SET trust_score = ? WHERE fact_id = ?').run(newTrust, row.fact_id)
+          promoted++
+          changed = true
+        }
+      }
+
+      // Aging (60 天未检索)
+      if (row.last_retrieved_at) {
+        const lastRetrieved = new Date(row.last_retrieved_at + 'Z').getTime()
+        const daysSinceRetrieval = (now - lastRetrieved) / 86_400_000
+        if (daysSinceRetrieval > 60) {
+          const currentTrust = this.db.prepare('SELECT trust_score FROM facts WHERE fact_id = ?').get(row.fact_id) as any
+          const newTrust = clampTrust(currentTrust.trust_score * 0.95)
+          this.db.prepare('UPDATE facts SET trust_score = ? WHERE fact_id = ?').run(newTrust, row.fact_id)
+          aged++
+          changed = true
+        }
+      }
+      // last_retrieved_at 为 NULL = 新 fact，不老化
+
+      if (!changed) unchanged++
+
+      // Long facts report (content > 300 字无 summary)
+      const matchLength = row.summary ? row.summary.length : row.content.length
+      if (matchLength > 300) {
+        longFacts.push({
+          id: row.fact_id,
+          content_length: row.content.length,
+          penalty: Math.min(1.0, 300 / matchLength),
+          has_summary: !!row.summary,
+        })
+      }
+    }
+
+    return { promoted, demoted, aged, unchanged, long_facts: longFacts }
+  }
+
+  /** 数据质量审计（只读，不修改数据） */
+  runAudit(): {
+    total_facts: number
+    long_without_summary: Array<{ id: number; content_length: number }>
+    low_helpful_rate: Array<{ id: number; rate: number; retrieval_count: number }>
+    aging_candidates: Array<{ id: number; last_retrieved_at: string | null }>
+  } {
+    const rows = this.db.prepare(
+      'SELECT fact_id, content, summary, retrieval_count, helpful_count, last_retrieved_at FROM facts'
+    ).all() as Array<{
+      fact_id: number; content: string; summary: string | null;
+      retrieval_count: number; helpful_count: number; last_retrieved_at: string | null
+    }>
+
+    const longWithoutSummary: Array<{ id: number; content_length: number }> = []
+    const lowHelpfulRate: Array<{ id: number; rate: number; retrieval_count: number }> = []
+    const agingCandidates: Array<{ id: number; last_retrieved_at: string | null }> = []
+
+    const now = Date.now()
+
+    for (const row of rows) {
+      // 超 500 字无 summary
+      if (row.content.length > 500 && !row.summary) {
+        longWithoutSummary.push({ id: row.fact_id, content_length: row.content.length })
+      }
+
+      // 低 helpful 率（>30 次检索，rate < 5%）
+      if (row.retrieval_count > 30) {
+        const rate = row.helpful_count / row.retrieval_count
+        if (rate < 0.05) {
+          lowHelpfulRate.push({ id: row.fact_id, rate: Math.round(rate * 1000) / 1000, retrieval_count: row.retrieval_count })
+        }
+      }
+
+      // 老化候选（>60 天未检索）
+      if (row.last_retrieved_at) {
+        const lastRetrieved = new Date(row.last_retrieved_at + 'Z').getTime()
+        const daysSince = (now - lastRetrieved) / 86_400_000
+        if (daysSince > 60) {
+          agingCandidates.push({ id: row.fact_id, last_retrieved_at: row.last_retrieved_at })
+        }
+      }
+    }
+
+    return {
+      total_facts: rows.length,
+      long_without_summary: longWithoutSummary,
+      low_helpful_rate: lowHelpfulRate,
+      aging_candidates: agingCandidates,
+    }
+  }
+
   /** 获取数据库连接（供 FactRetriever 直接使用） */
   get connection(): Database.Database {
     return this.db
@@ -795,9 +999,11 @@ export class MemoryStore {
       category: row.category as FactCategory,
       tags: row.tags,
       keywords: row.keywords,
+      summary: (row as any).summary ?? null,
       trustScore: row.trust_score,
       retrievalCount: row.retrieval_count,
       helpfulCount: row.helpful_count,
+      lastRetrievedAt: (row as any).last_retrieved_at ?? null,
       createdAt: row.created_at,
       updatedAt: row.updated_at,
     }

package/src/types.ts

@@ -8,9 +8,11 @@ export interface Fact {
   category: FactCategory
   tags: string
   keywords: string
+  summary: string | null
   trustScore: number
   retrievalCount: number
   helpfulCount: number
+  lastRetrievedAt: string | null
   createdAt: string
   updatedAt: string
 }
@@ -53,7 +55,7 @@ export interface RetrieverOptions {
 
 /** fact_store 工具调用参数 */
 export interface FactStoreArgs {
-  action: 'add' | 'search' | 'probe' | 'related' | 'reason' | 'contradict' | 'update' | 'remove' | 'list'
+  action: 'add' | 'search' | 'probe' | 'related' | 'reason' | 'contradict' | 'update' | 'remove' | 'list' | 'learn' | 'audit'
   content?: string | string[]
   query?: string
   entity?: string
@@ -61,6 +63,7 @@ export interface FactStoreArgs {
   fact_id?: number | number[]
   category?: string
   tags?: string
+  summary?: string
   trust_delta?: number
   min_trust?: number
   limit?: number