memory-eternal 0.17.0 → 0.18.0

package/index.js

@@ -31,7 +31,7 @@ import { migrateFromMarkdown, setAuditConfig, backupDb } from './lib/db.js'
 import { evaluateStall, MAX_STAMPS } from './lib/stall.js'
 import { resolveVaultDir, currentWorkspace } from './lib/vault-resolve.js'
 import { createHub } from './lib/sse.js'
-import { summarizeTurn, summarizeTurnDetailed, routeCandidates, sessionRouteFrom, extractLastTurn, sliceNewEvents, sessionEvents, sessionEventApi, createCaptureHealth, resolveRoute, captureCard, captureUpdate, pickNeighbors, hasUsableContent, deriveTitle, looksTruncated, maxTokenLadder, DEFAULT_CAPTURE_MAX_TOKENS, MAX_CAPTURE_MAX_TOKENS } from './lib/capture.js'
+import { summarizeTurn, summarizeTurnDetailed, routeCandidates, sessionRouteFrom, extractLastTurn, sliceNewEvents, sessionEvents, sessionEventApi, createCaptureHealth, resolveRoute, captureCard, captureUpdate, pickNeighbors, hasUsableContent, deriveTitle, looksTruncated, maxTokenLadder, guessRawKind, DEFAULT_CAPTURE_MAX_TOKENS, MAX_CAPTURE_MAX_TOKENS } from './lib/capture.js'
 import { createApi, json, encodeBody, crossOriginError, emptyNumericFields } from './lib/api.js'
 import { appendCaptureLog, readCaptureLog, rotateCaptureLog } from './lib/capture-log.js'
 import { recordLedger } from './lib/ledger.js'
@@ -54,6 +54,10 @@ export const Config = z.object({
   enabled: z.boolean().default(true),
   autoCapture: z.boolean().default(true),
   autoRecall: z.boolean().default(true),
+  // 核心记忆：人工批准的核心卡，每次会话强制注入 systemPrompt（不依赖 autoRecall / 工具调用）
+  coreMemoryEnabled: z.boolean().default(true),
+  coreMemoryLimit: z.number().min(1).max(20).default(6),
+  coreMemoryMaxChars: z.number().min(200).max(10000).default(2000),
   vaultDir: z.string().default(''),
   dedupThreshold: z.number().min(0).max(1).default(0.62),
   captureMinChars: z.number().default(200),
@@ -634,6 +638,8 @@ export function apply(ctx, config) {
 
   // 自动审核：根据配置决定新卡 status（免审→approved，否则 pending）
   const resolveAuditStatus = (cfg, kind, submittedBy) => {
+    // 核心记忆必须人工审核：自动沉淀/手动写入都先待审，避免一次误判污染每个会话。
+    if (kind === 'core') return 'pending'
     if (cfg.auditMode === 'none') return 'approved'
     const agents = Array.isArray(cfg.auditExemptAgents) ? cfg.auditExemptAgents : []
     const kinds = Array.isArray(cfg.auditExemptKinds) ? cfg.auditExemptKinds : []
@@ -684,18 +690,19 @@ export function apply(ctx, config) {
       const source = DSH_AGENT
       if (cfg.distillEnabled === false || !llm) {
         const rawTitle = deriveTitle(text)
+        const rawKind = guessRawKind({ title: rawTitle, body: text })
         const out = await captureCard(root, {
-          kind: 'content',
+          kind: rawKind,
           title: rawTitle.length >= 6 && !looksTruncated(rawTitle) ? rawTitle : '对话原文记录',
           tags: ['raw'],
           body: text,
           source,
-          status: resolveAuditStatus(cfg, 'content', source || 'unknown'),
+          status: resolveAuditStatus(cfg, rawKind, source || 'unknown'),
           submittedBy: source || 'unknown',
           severity: 'info',
           reason: 'AI 自动沉淀（原文卡）',
         }, { threshold: cfg.dedupThreshold })
-        if (out.ok) { countWrite(); logCapture(sessionId, 'created', '原文卡（蒸馏已关闭）', { path: out.path ?? out.rel, kind: 'content' }) }
+        if (out.ok) { countWrite(); logCapture(sessionId, 'created', '原文卡（蒸馏已关闭）', { path: out.path ?? out.rel, kind: rawKind }) }
         else if (out.duplicate) { countWrite(); logCapture(sessionId, 'appended', '与已有卡重复 → 追加更新记录', { path: out.duplicate.path }) }
         else logCapture(sessionId, 'fail', `写卡失败：${out.reason || '未知原因'}`)
         return
@@ -806,20 +813,21 @@ export function apply(ctx, config) {
       if (!result) {
         // 蒸馏失败不能让内容白丢：退成原文卡（与「关闭蒸馏」同一条降级路径）。
         const rawTitle = deriveTitle(text)
+        const rawKind = guessRawKind({ title: rawTitle, body: text })
         const raw = await captureCard(root, {
-          kind: 'content',
+          kind: rawKind,
           title: rawTitle.length >= 6 && !looksTruncated(rawTitle) ? rawTitle : '对话原文记录',
           // 蒸馏失败产出的原文卡与被批准的噪声卡是审核中心的主要污染源（#15 评论）：
           // 打上独立 tag，便于按 tag 过滤 / 批量处理，不再和正常卡混作一团。
           tags: ['raw', 'distill-failed'],
           body: text,
           source,
-          status: resolveAuditStatus(cfg, 'content', source),
+          status: resolveAuditStatus(cfg, rawKind, source),
           submittedBy: source,
           severity: 'info',
           reason: `AI 自动沉淀（蒸馏无输出 → 原文卡兜底；主因：${why}）`,
         }, { threshold: cfg.dedupThreshold })
-        if (raw.ok) { countWrite(); logCapture(sessionId, 'created', `蒸馏失败（${why}）→ 原文卡兜底`, { path: raw.path ?? raw.rel, kind: 'content', model: route.model }) }
+        if (raw.ok) { countWrite(); logCapture(sessionId, 'created', `蒸馏失败（${why}）→ 原文卡兜底`, { path: raw.path ?? raw.rel, kind: rawKind, model: route.model }) }
         else if (raw.duplicate) { countWrite(); logCapture(sessionId, 'appended', '蒸馏无输出 + 与已有卡重复 → 追加更新', { path: raw.duplicate.path, model: route.model }) }
         else logCapture(sessionId, 'fail', `蒸馏无输出，且兜底原文卡也失败：${raw.reason || '未知原因'}`, { model: route.model })
         return
@@ -966,6 +974,50 @@ export function apply(ctx, config) {
     }
   }, 'memory-eternal: recall section')
 
+  // -- 2b. 核心记忆：每次会话强制注入（不依赖 autoRecall / 工具调用） ---------
+  ctx.effect(() => {
+    let disposeSection = null
+    let timer = null
+    let alive = true
+    let lastText = ''
+    const clearSection = () => {
+      if (!disposeSection) return
+      const dispose = disposeSection
+      disposeSection = null
+      dispose()
+    }
+    const refresh = async () => {
+      try {
+        const cfg = settings.get() ?? {}
+        if (cfg.enabled === false || cfg.coreMemoryEnabled === false) { clearSection(); lastText = ''; return }
+        const { coreMemoryPrompt } = await import('./lib/core-memory.js')
+        const out = await coreMemoryPrompt(vaultDir(), {
+          enabled: true,
+          limit: cfg.coreMemoryLimit,
+          maxChars: cfg.coreMemoryMaxChars,
+        })
+        if (!alive) return
+        if (!out.text) { clearSection(); lastText = ''; return }
+        if (out.text === lastText) return
+        lastText = out.text
+        clearSection()
+        disposeSection = ctx.systemPrompt.section({
+          name: 'memory-eternal: core-memory',
+          order: 120,
+          text: out.text,
+        })
+      } catch { /* 核心记忆失败不能影响普通对话 */ }
+    }
+    refresh()
+    timer = setInterval(refresh, 5000)
+    if (timer && typeof timer.unref === 'function') timer.unref()
+    return () => {
+      alive = false
+      if (timer) clearInterval(timer)
+      clearSection()
+    }
+  }, 'memory-eternal: core memory section')
+
   const tools = ctx.get('tools')
   if (tools !== undefined) {
     tools.register(defineTool({
@@ -1220,7 +1272,7 @@ export function apply(ctx, config) {
               const cfg = settings.get() ?? {}
               // 只暴露可安全展示/回填的字段
               const safe = {
-                autoCapture: cfg.autoCapture, autoRecall: cfg.autoRecall, recallLimit: cfg.recallLimit, recallSummaryLen: cfg.recallSummaryLen, recallIncludeBody: cfg.recallIncludeBody, secretHint: cfg.secretHint || '',
+                autoCapture: cfg.autoCapture, autoRecall: cfg.autoRecall, coreMemoryEnabled: cfg.coreMemoryEnabled, coreMemoryLimit: cfg.coreMemoryLimit, coreMemoryMaxChars: cfg.coreMemoryMaxChars, recallLimit: cfg.recallLimit, recallSummaryLen: cfg.recallSummaryLen, recallIncludeBody: cfg.recallIncludeBody, secretHint: cfg.secretHint || '',
                 captureMinChars: cfg.captureMinChars, captureCooldownMs: cfg.captureCooldownMs, dedupThreshold: cfg.dedupThreshold, maxCardsPerDay: cfg.maxCardsPerDay,
                 distillEnabled: cfg.distillEnabled, dedupByLLM: cfg.dedupByLLM, captureMaxTokens: cfg.captureMaxTokens, recallMinScore: cfg.recallMinScore,
                 autoWeb: cfg.autoWeb, autoWebMode: cfg.autoWebMode, webPort: cfg.webPort, webCheckIntervalMs: cfg.webCheckIntervalMs, webMaxRestart: cfg.webMaxRestart, watchdogAutoSpawn: cfg.watchdogAutoSpawn, autoRestartOnDrift: cfg.autoRestartOnDrift, autoMcpSetup: cfg.autoMcpSetup,

package/lib/api.js

@@ -16,8 +16,8 @@ import os from 'node:os'
 import fs, { readFileSync } from 'node:fs'
 import path from 'node:path'
 import { spawn } from 'node:child_process'
-import { listCards, countCards, readCard, search, searchAll, graph, graphAll, overview, exportCards, deleteCard, writeCard, parseCard, stats, optimizeCandidates, optimizeApply, addFeedback, dailyCounts, mergeCards, ensureVault, auditQueue, recycleList, purgeExpired, setCardStatus, restoreCard, mainStoreBodies } from './vault.js'
-import { compressExcerpt, DEFAULT_CAPTURE_MAX_TOKENS } from './capture.js'
+import { listCards, countCards, readCard, search, searchAll, graph, graphAll, overview, exportCards, deleteCard, writeCard, parseCard, stats, optimizeCandidates, optimizeApply, addFeedback, dailyCounts, mergeCards, ensureVault, auditQueue, recycleList, purgeExpired, setCardStatus, setCardKind, restoreCard, mainStoreBodies } from './vault.js'
+import { compressExcerpt, guessRawKind, DEFAULT_CAPTURE_MAX_TOKENS } from './capture.js'
 import { buildDiagnostics } from './feedback.js'
 import { getSetupStatus, runSetup, mcpAgentAction } from './setup.js'
 import { ledgerSummary } from './ledger.js'
@@ -625,17 +625,44 @@ export function createApi(deps = {}) {
         json(res, 200, { ok: true, soft: !permanent })
         return
       }
+      case '/reclassify': {
+        // 重新识别待审原文卡的类型：修复旧版本「自动沉淀全是 content」的历史数据。
+        // 只处理带 raw / 原文卡标记的待审卡，不动人工创建的内容。
+        const dryRun = query.get('dry') === '1'
+        const { getDb } = await import('./db.js')
+        const db = getDb(vaultRoot)
+        const rows = db.prepare("SELECT id, path, title, body, tags, reason FROM quarantine WHERE status = 'pending'").all()
+        let scanned = 0, changed = 0
+        const byKind = {}
+        for (const row of rows) {
+          let tags = []
+          try { tags = JSON.parse(row.tags || '[]') } catch { tags = [] }
+          const isRaw = (Array.isArray(tags) && tags.includes('raw')) || /原文卡|自动沉淀/.test(String(row.reason || ''))
+          if (!isRaw) continue
+          scanned++
+          const kind = guessRawKind({ title: row.title, body: row.body })
+          if (kind === 'content') continue
+          changed++
+          byKind[kind] = (byKind[kind] || 0) + 1
+          if (!dryRun) db.prepare('UPDATE quarantine SET kind = ?, updated_at = ? WHERE id = ?').run(kind, new Date().toISOString(), row.id)
+        }
+        json(res, 200, { ok: true, dryRun, scanned, changed, byKind })
+        return
+      }
       case '/audit/list': {
         const { pending, rejected } = await auditQueue(vaultRoot)
         json(res, 200, { ok: true, pending, rejected })
         return
       }
       case '/audit/approve': {
-        // POST {path} 或 GET ?path= → status=approved
-        const rel = query.get('path') || (await readBody(req))?.path || ''
+        // POST {path, kind?} 或 GET ?path=&kind= → status=approved；kind=core 时同时设为核心
+        const body = query.get('path') ? null : await readBody(req)
+        const rel = query.get('path') || body?.path || ''
+        const kind = String(query.get('kind') || body?.kind || '').trim()
         if (!rel) return json(res, 400, { ok: false, error: '缺少 path' })
         await setCardStatus(vaultRoot, rel, 'approved')
-        json(res, 200, { ok: true, path: rel, status: 'approved' })
+        if (kind === 'core' || kind === 'knowledge') await setCardKind(vaultRoot, rel, kind)
+        json(res, 200, { ok: true, path: rel, status: 'approved', kind: kind || undefined })
         return
       }
       case '/audit/reject': {
@@ -650,6 +677,21 @@ export function createApi(deps = {}) {
         json(res, 200, { ok: true, path: rel, status: 'rejected', reason })
         return
       }
+      case '/card-kind': {
+        // POST {path, kind:'core'|'knowledge'|...}：人工把已批准卡片设为核心/取消核心
+        if (req.method !== 'POST') return json(res, 405, { ok: false, error: '需要 POST' })
+        const body = await readBody(req)
+        const rel = String(query.get('path') || body.path || '').trim()
+        const kind = String(body.kind || '').trim()
+        if (!rel || !kind) return json(res, 400, { ok: false, error: '缺少 path 或 kind' })
+        try {
+          const out = await setCardKind(vaultRoot, rel, kind)
+          json(res, 200, { ok: true, ...out })
+        } catch (e) {
+          json(res, 400, { ok: false, error: String(e?.message || e) })
+        }
+        return
+      }
       case '/recycle/list': {
         const items = await recycleList(vaultRoot)
         json(res, 200, { ok: true, items })
@@ -996,7 +1038,7 @@ export function createApi(deps = {}) {
         const cfg = getSettings() ?? {}
         const env = typeof process !== 'undefined' ? process.env : {}
         const safe = {
-          autoCapture: cfg.autoCapture ?? true, autoRecall: cfg.autoRecall ?? true, recallLimit: cfg.recallLimit ?? 5, recallSummaryLen: cfg.recallSummaryLen ?? 130, recallIncludeBody: cfg.recallIncludeBody ?? false, secretHint: cfg.secretHint || '',
+          autoCapture: cfg.autoCapture ?? true, autoRecall: cfg.autoRecall ?? true, coreMemoryEnabled: cfg.coreMemoryEnabled ?? true, coreMemoryLimit: cfg.coreMemoryLimit ?? 6, coreMemoryMaxChars: cfg.coreMemoryMaxChars ?? 2000, recallLimit: cfg.recallLimit ?? 5, recallSummaryLen: cfg.recallSummaryLen ?? 130, recallIncludeBody: cfg.recallIncludeBody ?? false, secretHint: cfg.secretHint || '',
           captureMinChars: cfg.captureMinChars ?? 200, captureCooldownMs: cfg.captureCooldownMs ?? 300000, dedupThreshold: cfg.dedupThreshold ?? 0.62, maxCardsPerDay: cfg.maxCardsPerDay ?? 60,
           distillEnabled: cfg.distillEnabled ?? true, dedupByLLM: cfg.dedupByLLM ?? true, captureMaxTokens: cfg.captureMaxTokens ?? DEFAULT_CAPTURE_MAX_TOKENS, recallMinScore: cfg.recallMinScore ?? 2,
           autoWeb: cfg.autoWeb ?? env.MEMORY_AUTO_WEB !== '0', autoWebMode: cfg.autoWebMode || 'init', webPort: Number(cfg.webPort) || Number(env.MEMORY_WEB_PORT) || 7999, webCheckIntervalMs: Number(cfg.webCheckIntervalMs) || 5000, webMaxRestart: Number(cfg.webMaxRestart) || 10, watchdogAutoSpawn: cfg.watchdogAutoSpawn ?? env.MEMORY_WATCHDOG === '1', autoRestartOnDrift: cfg.autoRestartOnDrift ?? true, autoMcpSetup: cfg.autoMcpSetup ?? false,

package/lib/capture.js

@@ -316,8 +316,9 @@ export async function summarizeTurnDetailed(llm, route, conversation, opts = {})
     '输出严格 JSON（不要 markdown 代码块、不要多余文字）：',
     '{"save": false} 表示不值得保存；',
     '否则输出：',
-    '{"save": true, "title": "简短标题", "kind": "knowledge|project|content|prompt|business|tool|mistake", "tags": ["标签"], "body": "300-800字的压缩正文，markdown，含要点列表，首行是# 标题"}',
+    '{"save": true, "title": "简短标题", "kind": "knowledge|project|content|prompt|business|tool|mistake|core", "tags": ["标签"], "body": "300-800字的压缩正文，markdown，含要点列表，首行是# 标题"}',
     'kind 含义：project=项目背景/进度；knowledge=通用知识/技术方案/设计决策；content=内容素材/资料；prompt=提示词/工作流；business=业务/商业；tool=工具链/CLI/配置/环境坑；mistake=错误/反模式/踩坑/“别这么做”/debug 教训。',
+    'core 只在“用户明确要求长期/每次都遵守的偏好、硬约束、核心记忆”时使用（如“以后都…/永远记住/每次都…/不要再用…”）；一次性偏好或普通知识不要标 core。core 卡一律进待审核，由用户决定是否批准。',
     '正文要求：去掉对话口水，只留可复用的结论、参数、步骤、原因。语言与对话一致。',
     '已有卡片规则：若新知识与某张已有卡片是同一主题（同一技术、同一项目、同一决策），不要新建——输出',
     '{"append_to": "已有卡片的 path 字段原样", "update": "追加/修正的短文本（一段话）"}',
@@ -704,7 +705,7 @@ export function isUsableTitle(title) {
 
 /** 组装最终卡片对象（标题已体检通过）。 */
 function finishCard(parsed, title, body) {
-  const kind = ['project', 'knowledge', 'content', 'prompt', 'business', 'tool', 'mistake'].includes(parsed.kind) ? parsed.kind : 'knowledge'
+  const kind = ['project', 'knowledge', 'content', 'prompt', 'business', 'tool', 'mistake', 'core'].includes(parsed.kind) ? parsed.kind : 'knowledge'
   const tags = Array.isArray(parsed.tags)
     ? parsed.tags.map((t) => String(t).trim().slice(0, 30)).filter(Boolean).slice(0, 8)
     : []
@@ -1006,3 +1007,37 @@ export async function resolveRoute(llm, opts = {}) {
   const list = await routeCandidates(llm, opts)
   return list[0] ?? null
 }
+
+
+const RAW_KIND_SIGNALS = [
+  ['mistake', /(报错|错误|异常|失败|根因|原因|修复|解决|踩坑|教训|bug|error|exception|crash|stack trace|排查|回滚)/i],
+  ['tool', /(命令|CLI|终端|安装|配置|环境变量|端口|依赖|版本|npm|pnpm|yarn|\bgit\b|docker|powershell|bash|PATH|shell|SDK|API key|token)/i],
+  ['prompt', /(提示词|prompt|模板|工作流|角色设定|系统指令|AI 指令)/i],
+  ['project', /(项目|进度|里程碑|需求|交付|上线|排期|任务|计划|迭代|版本发布|roadmap)/i],
+  ['business', /(客户|合同|报价|价格|商业|营收|成本|预算|市场|销售|付费|订阅)/i],
+  ['content', /(文案|素材|文章|博客|视频脚本|标题|封面|宣传|内容创作|草稿)/i],
+]
+
+/**
+ * 不调 LLM（关闭蒸馏 / 蒸馏失败兜底）时的轻量类型判断。
+ * 只做启发式：宁可回退 knowledge，也不要把所有原文卡都塞成 content。
+ * 明确出现长期偏好/硬约束时返回 core（仍会强制进待审）。
+ */
+export function guessRawKind({ title = '', body = '' } = {}) {
+  const head = String(title || '')
+  const text = String(body || '').slice(0, 6000)
+  if (/(核心记忆|永远记住|以后都|永久(记住|遵守)|必须(遵守|记住)|不要再)/.test(head + '\n' + text)) return 'core'
+  const scores = {}
+  for (const [kind, re] of RAW_KIND_SIGNALS) {
+    const h = (head.match(re) || []).length
+    const b = (text.match(re) || []).length
+    const score = Math.min(5, h * 2 + b)
+    if (score) scores[kind] = score
+  }
+  let best = 'knowledge'
+  let bestScore = 0
+  for (const kind of ['mistake', 'tool', 'prompt', 'project', 'business', 'content']) {
+    if ((scores[kind] || 0) > bestScore) { best = kind; bestScore = scores[kind] }
+  }
+  return best
+}