@yottameta/yotta-memory 0.17.2 → 0.17.3

package/CHANGELOG.md

@@ -1,3 +1,8 @@
+## v0.17.3 (2026-09-26)
+
+帮助可读性微调：顶层 `--help` 的每个选项去掉逐项重复的「做什么：」前缀，直接写说明正文；
+「什么时候用：」与「注意：」保留不变。仅影响帮助文案输出，命令行为、参数与退出码零变化。
+
 ## v0.17.2 (2026-09-25)
 
 安全修复：私密维护的 owner 门（explain fail-closed / archive / maintain 跨 owner 跳过）。
@@ -312,8 +317,6 @@
 - SKILL.md：新增「常见问题 FAQ（速查）」小节，指向 references/faq.md。
 - 版本四件对齐 0.8.7（package.json / SKILL.md / CHANGELOG / 引擎 VERSION）。
 
-# 更新日志
-
 ## v0.8.5 (2026-08-29)
 
 安全修复（安全扫描发现高危，修复后三源同步升版）：

package/README.md

@@ -23,6 +23,7 @@
 
 > 📖 The user-facing operations manual lives in [USER_GUIDE.md](USER_GUIDE.md).
 
+> 🆕 **v0.17.3 (help readability)**: the top-level `--help` no longer repeats a "做什么：" label on every command and option — each entry starts straight with its description, while "什么时候用：" and "注意：" stay as-is. Help text only: commands, arguments and exit codes are unchanged.
 > 🆕 **v0.17.2 (owner gate for private maintenance)**: `explain` now refuses entries owned by another agent instead of printing their subject/statement, and `archive` / `maintain` skip private entries the caller does not own (declare your identity with `--agent <id>`; only an explicit user `--unsafe` overrides). Installer hardening: symlinked targets are refused and no whole-directory delete is performed. Previous v0.17.1: a mismatched identity-confirmation ID in `view` now shows an immediate error instead of returning silently; cancel stays silent and the server-side confirmation and destructive-operation gates are unchanged.
 > 🆕 **v0.17.0 (scale: year/month layout + doctor scale check + lazy index + bench)**: new entries are written into year/month folders — public `facts/<yyyy>/<mm>/`, private `private/<owner>/<type>/<yyyy>/<mm>/`. Flat files from v0.16 and earlier stay in place and remain readable; nothing is migrated automatically. Archiving keeps the year/month path, so same-named files from different months no longer overwrite each other. `doctor` gained a scale section (entry count / largest directory / index size / index cold start) with thresholds set through `config set scale_warn_entries` / `scale_warn_files_per_dir` / `scale_warn_index_bytes` / `scale_warn_cold_start_ms`; exceeding a threshold warns only and never locks destructive writes. Large stores can read only the shards of a given year: `recall --year <yyyy>` / `context --year <yyyy>` (repeatable; omitting it loads everything exactly as before). New `bench` command: deterministic auto sampling or `--evalset <file>` (evalset v1), reporting Recall@k / MRR / nDCG@k / HitRate with bootstrap 95% CI plus store and evalset fingerprints; `--gate <metric>=<value>` fits CI, and the same store + evalset + parameters always produce the same output. `bench` is read-only: it never rebuilds the index, never touches access counters and never calls an external embedding plugin. Previous v0.16.7: interactive `yotta-memory migrate --recovery-key-out` prompt, `YOTTA_MEMORY_PASS` for automation, and a `view` memory-home fingerprint before reusing a running server.
 > 🆕 **v0.17.0 (recovery probe + memory-store scan)**: `doctor --baseline [--against <store>] [--template <file>]` and `backup drill --probe` run six read-only baseline probes (identity / recent / source-only diff / CJK / operation rules / owner scope) against the target store or restored copy. Entry identity keys are layout-independent, so an old flat store migrating into the year/month layout never reports false gaps; missing items are listed and exit non-zero (`doctor` exit 2 / `drill` fails). Probes never hard-code any memory and `--template` only tightens the verdict. New `scan` command: seven classes (malicious instruction / prompt injection / credential leak / data exfiltration / guardrail bypass / behaviour manipulation / privilege escalation), five severities and `file:line` evidence; report-only by default, zero network and zero dependencies. `--gate` exits 1 at or above the given severity; `--quarantine` needs `--yes` or interactive confirmation and backs up each original file byte-for-byte into `.memory-scan/quarantine/` before redacting hit lines, with `--restore` to roll back; credential snippets are always masked.

package/README.zh-CN.md

@@ -23,6 +23,7 @@
 
 > 📖 面向用户的操作手册见 [USER_GUIDE.md](USER_GUIDE.md)。
 
+> 🆕 **v0.17.3（帮助可读性）**：顶层 `--help` 不再逐项重复「做什么：」前缀——每条命令与选项直接写说明正文，「什么时候用：」与「注意：」保留不变。仅帮助文案变化，命令、参数与退出码零改动。
 > 🆕 **v0.17.2（私密维护的 owner 门）**：`explain` 对其它 owner 的私密条目改为直接拒绝，不再打印 subject / statement；`archive` / `maintain` 跳过不属于当前身份的私密条目（用 `--agent <id>` 声明身份，只有用户显式 `--unsafe` 才越界）。安装器加固：拒绝对符号链接目标写入，不做整目录删除。上一版 v0.17.1：`view` 删除 AI 身份时确认 ID 输入错误会立即提示「ID 不匹配，请输入完整 agent ID」，不再静默返回；用户取消仍保持静默，后端确认校验与破坏性操作闸门不变。
 > 🆕 **v0.17.0（规模：文件分层 + doctor 规模体检 + 索引按需加载 + bench 基准评测）**：新写入按年/月分层——公共 `facts/<年>/<月>/`、私密 `private/<owner>/<type>/<年>/<月>/`；v0.16 及更早的平铺文件留在原位继续可读，不做自动迁移。归档保留分层，不同月份的同名文件不会互相覆盖。`doctor` 新增规模段（记忆条数 / 单目录最大文件数 / 索引总体积 / 索引冷启动耗时），阈值用 `config set scale_warn_entries` / `scale_warn_files_per_dir` / `scale_warn_index_bytes` / `scale_warn_cold_start_ms` 调整，超阈值只告警，不锁定破坏性写入。大库检索可以按年份只读对应分片：`recall --year <yyyy>` / `context --year <yyyy>`（可重复传多次，不传即全量、行为与旧版一致）。新增 `bench` 可复算基准评测：默认按库内条目确定性抽样，`--evalset <文件>` 也可指定评测集；输出 Recall@k / MRR / nDCG@k / HitRate + 95% 置信区间、库指纹与评测集指纹，`--gate <指标>=<数值>` 可接 CI，同库同评测集同参数必得同结果；全程只读，不重建索引、不写访问计数、不调用外部 embedding。上一版 v0.16.7：明文库转加密推荐交互式 `yotta-memory migrate --recovery-key-out "$env:USERPROFILE\yotta-memory-recovery.key"`，自动化使用 `YOTTA_MEMORY_PASS`；`view` 复用端口前校验 memory_home 指纹。
 > 🆕 **v0.17.0（恢复探针 + 记忆库安全扫描）**：`doctor --baseline [--against <库路径>] [--template <文件>]` 与 `backup drill --probe` 在目标库 / 恢复副本上跑六类只读基线探针（身份 / 近期 / 仅源库独有 / CJK / 操作规则 / owner 范围）；条目身份键与布局无关，旧平铺 → 新年/月分层的迁移不会误报「缺失」；缺失项列出清单并非零退出（`doctor` exit 2 / `drill` 判失败），探针不硬编码任何记忆，`--template` 只加严判定。新增 `scan` 记忆库安全扫描：七类（恶意指令 / Prompt 注入 / 凭证泄漏 / 数据外泄 / 护栏绕过 / 行为操纵 / 权限提升）+ 五级分级 + `file:line` 证据，默认只报告、零网络零依赖；`--gate` 命中该级别及以上 exit 1；`--quarantine` 需 `--yes` 或交互确认，先把原文件逐字节备份到 `.memory-scan/quarantine/` 再替换命中行，`--restore` 还原；凭证片段一律打码不回显。

package/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: yotta-memory
 description: 元忆 —— 有权限边界的文件式智能体记忆。文件式、零依赖、可 diff/可回滚：让任何 AI 智能体活过会话，开工 recall 恢复上下文、重要信息 remember 落盘、收工归档。类型体系 FACT（公共共享）/ PREF / BOUND / COMMIT（私密隔离）。触发：记住、别忘了、记一笔、记忆、remember、recall、跨会话、上次说到、续测、交接、归档、记忆盘、共享记忆、局域网记忆、画像、开工上下文、长期理解摘要、近期走廊、会话闭环、记忆守则、profile、context、越用越懂、语义检索、反馈、维护、蒸馏、feedback、maintain、distill、explain、自我学习、自我进化、自我提升、查看平台分页、recall 候选预过滤、任务相关记忆、--focus、--embedding、压缩遗忘、consolidate、周期摘要、自动合并、分类型衰减、回滚、备份、backup、防误删、doctor、事务快照、基线探针、恢复演练、记忆库安全扫描、scan
-version: 0.17.2
+version: 0.17.3
 license: MIT
 ---
 

package/bin/yotta-memory.js

@@ -25,7 +25,7 @@ const net = require('net');
 const child_process = require('child_process');
 const { AsyncLocalStorage } = require('async_hooks');
 
-const VERSION = '0.17.2';
+const VERSION = '0.17.3';
 const CLI_VALUE_OPTS = new Set(['--type', '--limit', '--days', '--out', '--owner', '--agent', '--agent-id', '--agent-key', '--agent-key-file', '--plugin-data', '--threshold', '--scope', '--host', '--port', '--dir', '--name', '--user', '--relationship', '--source', '--weight', '--budget', '--password', '--new-password', '--recovery-key', '--recovery-key-out', '--reason', '--merge', '--model', '--subject', '--embedding', '--focus', '--embedding-timeout', '--min-age', '--min-idle', '--max-utility', '--min-group', '--period', '--to', '--id', '--time', '--tools', '--mcp-config', '--skill-dir', '--year', '--evalset', '--k', '--seed', '--bootstrap', '--gate', '--against', '--template', '--path']);
 const CLI_FLAG_OPTS = new Set(['--project', '--all', '--unsafe', '--no-auth', '--stdio', '--onstart', '--from-current', '--restart', '--force', '--attach', '--allow-same-volume', '--verify', '--no-hint', '--encrypt', '--no-encrypt', '--password-stdin', '--json', '--manual', '--skip-schedule', '--useful', '--useless', '--undo', '--dry-run', '--apply', '--purge', '--dedup', '--batches', '--explain', '--semantic', '--runtime', '--ablate', '--timing', '--baseline', '--probe', '--quarantine', '--restore', '--yes', '--keep-memories', '--keep-identity']);
 
@@ -9341,7 +9341,8 @@ function helpPad(text, width) {
 }
 function helpText(item) {
   const parts = [];
-  if (item.what) parts.push('做什么：' + item.what);
+  // v0.17.3：去掉逐项重复的「做什么：」前缀，直接写说明；「什么时候用 / 注意」保留。
+  if (item.what) parts.push(item.what);
   if (item.when) parts.push('什么时候用：' + item.when);
   if (item.caution) parts.push('注意：' + item.caution);
   return parts.join('；');

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yottameta/yotta-memory",
-  "version": "0.17.2",
+  "version": "0.17.3",
   "description": "Yuanyi (元忆) — boundary-aware, file-based memory for AI agents. File-based, zero-dependency, diff/rollback-able; FACT/PREF/BOUND/COMMIT types (public shared / private isolated), user-level + project-level storage; v0.12 reliability baseline (init guard, trash-based deletion, independent-volume backup/list/doctor/restore, start-of-work doctor, transactional snapshots before destructive writes), v0.10 consolidation (periodic summaries with provenance, near-duplicate auto-merge, per-type decay, batch audit + rollback), v0.9 recall quality + context focus + optional local embedding plugin, plus v0.8 semantic search, feedback loop, self-organization and distillation.",
   "license": "MIT",
   "keywords": [

package/skill-manifest.json

@@ -3,7 +3,7 @@
   "slug": "yotta-memory",
   "name": "元忆",
   "package": "@yottameta/yotta-memory",
-  "version": "0.17.2",
+  "version": "0.17.3",
   "trust": "yottameta",
   "install": {
     "idempotent": true