@avantf/dsh-mem 0.9.0 → 0.10.0

package/README.md

@@ -1,38 +1,34 @@
 # @avantf/dsh-mem
 
-DSH（DeepSeek Harness）原生记忆插件：给 agent 一份可长期检索的**记忆**和一个**文档知识库**——
-它跨会话记得你说过的事，也能把文档收进库里按语义查回来。装上后提供 7 个模型工具、三段用法提示
-和两个主窗口标签页（记忆 / 知识）。
+DSH（DeepSeek Harness）原生记忆插件：给 agent 一份可长期检索的**记忆**——它跨会话记得你说过的事，
+也能按语义、全文、实体、标题与时间把它找回来。装上后提供 3 个模型工具、一段用法提示
+和一个主窗口标签页（记忆）。
+
+> 文档知识库不在本插件里：它已拆到独立插件 **`dsh-knowledges`**（`knowledge.db`、受管文档、`kb_*` 工具、
+> 知识页与文档转换都在那边）。本插件既不读也不删 `~/.avantf/knowledge/**`。
 
 ## 主要功能
 
-- **7 个模型工具**
-  - `mem_remember`：新增 / 改写 / 删除事实（写入后立刻做矛盾检测）
-  - `mem_recall`：混合检索（语义 + 全文 + 实体）、三元组方向问答、实体探查、多实体联合推理、矛盾查询
-  - `mem_admin`：列表 / **导出 / 导入 / 快照重建**。文件路径**只允许本次运行的数据根内**
+- **3 个模型工具**
+  - `mem_remember`：新增 / 改写 / 删除事实（写入后立刻做矛盾检测），分两层——`journal`（做了什么 /
+    聊了什么，必带 `event_date` 与短 `title`，同一事务复用同一 title）与 `note`（稳定事实 / 偏好 /
+    约定 / 坑，可带 `relations` 三元组）
+  - `mem_recall`：`search` 内容检索（先对整句做字面精确匹配，再按查询形态返回事实集合 / 关系结构 /
+    模糊检索；只给 `time` 就是纯时间召回）与 `titles` 标题检索（写 journal 前查重、标题纠错用）
+  - `mem_admin`：`list` 列举 / `contradictions` 列出未处理的矛盾对 / **导出 / 导入 / 快照重建**。
+    文件路径**只允许本次运行的数据根内**
     （默认导出到 `<数据根>/backups/memory-YYYYMMDD-HHMMSS.json.br`，原子写，结果不含文档正文）；
     `import` 按**文件头**自动识别 SQLite（只读合并另一份 `memory.db`，报告 `mode:'db'`）与 JSON 导出文档
     （`mode:'json'`）；`restore_snapshot` 必须显式 `confirm: true`，且只接受**空 facts 表**，恢复后同一宿主
     进程立即可检索。
-  - `kb_add` / `kb_list` / `kb_remove`：新增文档（文本 / 文件 / URL，或批量导入目录，同名拒绝）、
-    列表（带出受管 `.md` 的绝对路径与切片详情）、删除；
-    **改一篇既有文档就是改那份 `.md`**（用文件工具改，索引自动跟随）
-  - `kb_query`：跨「记忆事实 + 文档切片」的统一检索，结果带来源标注，可据此引用原文
-  - 诊断、策展（pin / 归档恢复 / 反馈 / 矛盾裁决）与索引修复（重建索引、向量修复）是**操作者动作**：
-    在「记忆」「知识」页与 CLI 里可用，不占用模型工具面。
-- **记忆与知识一起检索**：对记忆事实与文档切片做统一排序，两类结果处在同一分数刻度上，
-  所以"我之前定过的约定"和"文档里怎么写的"能一次问出来。
-- **记忆是活的**：按活跃使用结算，支持有效期、遗忘、去重与修订链；同一主题的改动会被识别成
+  - 诊断、策展（pin / unpin / 归档恢复 / 矛盾裁决）与索引修复（重建索引、向量修复）是**操作者动作**：
+    在「记忆」页与 CLI 里可用，不占用模型工具面。
+- **混合检索**：语义 + 全文 + 实体三腿融合；`search` 按查询形态给出事实集合（多实体并列）、关系结构
+  （谁管理谁这类方向问句）、纯时间召回或模糊检索。
+- **记忆是活的**：按活跃使用日结算、支持遗忘、去重与修订链；同一主题的改动会被识别成
   对旧事实的**修订**而不是简单叠加，写入时顺带做矛盾检测。
-- **资料随手入库**：文本、本地文件、目录、URL 都能收；Markdown 按标题分块，PDF、`.docx`、`.xlsx`
-  等常见格式也能直接入库（转换过程有问题会明确回报，不静默丢内容）。
-- **知识可维护**：每篇文档在数据目录里留一份可编辑的 Markdown 副本，用你自己的编辑器改，
-  插件会在其后自动把改动同步回索引；删除文档只删这份副本，你原来的文件不动。
-- **两个主窗口标签页**（排在 对话 / 轨迹 之后）：**记忆**（浏览、编辑、归档恢复、反馈、矛盾裁决）
-  与**知识**（文档列表，每行可查看切片 / 编辑 / 打开目录 / 重新摄入 / 删除；另有跨库查询与入库——
-  URL / 本地文件 / 本地目录 / 粘贴文本共用一个输入）。
-- 系统提示词里只加**三段**用法提示（何时该主动记忆、何时该先查记忆、何时该先查知识库 / 怎么改一篇
-  既有文档）。
+- **一个主窗口标签页**（排在 对话 / 轨迹 之后）：**记忆**（浏览、编辑、pin/unpin、归档恢复、矛盾裁决）。
+- 系统提示词里只加**一段**用法提示（何时该主动记忆、journal 与 note 怎么分、同一事务怎么复用 title）。
 
 ## 接入 dsh
 
@@ -64,19 +60,21 @@ dsh plugin --profile <PROFILE> add @avantf/dsh-mem@<version>
 ```
 
 ```text
-把 ~/docs/cgroup-v2.pdf 收到知识库的 ops 域下。
+上周做了什么？（按事件时间召回 journal 条目）
 ```
 
 ```text
-看看知识库里关于「内存保护」的原文。
+谁负责发布流程？（关系问句直接返回关系结构）
 ```
 
 ```text
 把「每周四发布」改成「每周三发布」，并说明依据。
 ```
 
-改一篇已经入库的文档：让模型"更新这篇文档"即可；也可以用任意文件工具直接编辑那份受管
-`.md`（先用 `kb_list` 拿到路径），改完会自动同步，「知识」页的「重新摄入」是手动兜底。
+查重与改标题：写 journal 前先让模型用 `mem_recall` 的 `titles` 找候选，读到候选就 `update`、
+读不到才 `add`；标题写错时用 `update` 带 `retitle: true` 纠正一次（旧标题保留为别名，仍能检索到）。
+
+要改一篇**文档**：那是 `dsh-knowledges` 的事（`kb_*` 工具与「知识」页在那边），本插件不提供文档入口。
 
 常见维护动作（在「记忆」页或 CLI 里）：
 
@@ -87,8 +85,7 @@ dsh plugin --profile <PROFILE> add @avantf/dsh-mem@<version>
 
 ### 相关性门槛：严格 / 宽松
 
-「记忆」「知识」页的查询面板有**严格 / 宽松**选择（工具面的 `mem_recall` / `kb_query` 也接受
-`floors: 'strict' | 'loose'`）：
+「记忆」页的查询面板有**严格 / 宽松**选择（工具面的 `mem_recall` 也接受 `floors: 'strict' | 'loose'`）：
 
 - **严格**（面板默认）：只有明确够相关的才返回，所以"没有结果"是一个可信的答案。
 - **宽松**：门槛更低，用来找回边角相关的条目——仍有一条底线，不是无门槛。
@@ -97,17 +94,19 @@ dsh plugin --profile <PROFILE> add @avantf/dsh-mem@<version>
 
 ## 配置、数据与提示词
 
-- **数据目录**：默认 `~/.avantf`。记忆库、知识库（含受管文档副本）与模型缓存都在其中；
+- **数据目录**：默认 `~/.avantf`。记忆库（`memory/memory.db`）、配置与提示词文件、备份与模型缓存都在其中；
   环境变量 `AVANTF_HOME` 可以改变它。
-- **配置**：`~/.avantf/configs/` 下的三份 YAML —— `common.yaml`（公共项，缺失时按注释模板自动创建）、
-  `memory.yaml`、`knowledge.yaml`。只写你要改的项，其余取内置默认。
-- **提示词文件**：`~/.avantf/prompts/` 下的 `mem-memory-usage.md`、`mem-knowledge-usage.md`、
-  `mem-kb-edit.md`。缺失或为空时插件写入默认内容，**只在启动时读一次**，改完重启生效。
-  可以按自己的领域改写：内置默认刻意不出现任何领域名词，只给判据与两种形状——
-  - 事实型：`用户不吃香菜；订餐默认备注不要香菜。`（谁 / 什么范围 / 偏好或要求或术语是什么）
-  - 规则型：`只有「注销式回购」减少总股本，库存股回购不减少——比较 EPS 影响时要区分。`
-  - 不适合写进记忆：一次调查的过程叙述、数据表、命令与链接——依据归知识库文档，记忆里只留
-    **结论 + 位置**（路径 / 来源 / 名称）；敏感信息同理，只记存放位置，不记明文。
+- **配置**：`~/.avantf/configs/` 下的两份 YAML —— `common.yaml`（公共项，缺失时按注释模板自动创建）与
+  `memory.yaml`。只写你要改的项，其余取内置默认。寿命按**活跃使用日**表达（`trust.base_lifetime_days` 是新事实的
+  寿命，默认 **90**；`trust.max_retention_days` 是召回加强能换来的最长寿命，默认 **174**；
+  `lifecycle.purge_after_archived_days` 是归档后的回收窗口）；没有单条到期字段，也没有日历兜底。
+- **提示词文件**：`~/.avantf/prompts/` 下的 `mem-memory-usage.md`。缺失或为空时插件写入默认内容，
+  **只在启动时读一次**，改完重启生效。
+  可以按自己的领域改写：内置默认刻意不出现任何领域名词，只给判据与记忆的两层——
+  - `journal`：一项实质性工作 / 话题完成或阶段性收敛时主动记一条，必须带 `event_date` 与短 `title`；
+  - `note`：稳定的偏好 / 约定 / 坑，直接记录，规则写成「当……应……」。
+  - 不适合写进记忆：临时聊天、一次性中间状态、容易查到的内容——依据归文档（`dsh-knowledges` 的文档库），
+    记忆里只留 **结论 + 位置**（路径 / 来源 / 名称）；敏感信息同理，只记存放位置，不记明文。
 - **升级前建议先备份 `~/.avantf`**：数据库迁移是单向的，用旧版本打开新库会被明确拒绝。
 
 ## 升级 / 卸载
@@ -120,8 +119,8 @@ dsh plugin --profile <PROFILE> add @avantf/dsh-mem@<version>
 dsh plugin --profile <PROFILE> remove @avantf/dsh-mem
 ```
 
-两者都要重启 dsh 生效。卸载**不动 `~/.avantf`**：记忆库、知识库、模型缓存全部保留。DSH Desktop 在
-「插件」页上卸载。
+两者都要重启 dsh 生效。卸载**不动 `~/.avantf`**：记忆库与模型缓存全部保留（`~/.avantf/knowledge/**` 属
+`dsh-knowledges`，同样不动）。DSH Desktop 在「插件」页上卸载。
 
 ## 环境要求
 

package/lib/client.js

@@ -4160,11 +4160,9 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 				content: optionalText.optional(),
 				fact_id: count.optional(),
 				kind: optionalText.optional(),
-				ttl_days: count.optional(),
 				reason: optionalText.optional(),
 				source_ref: optionalText.optional(),
 				event_date: optionalText.optional(),
-				valid_until: optionalText.optional(),
 				title: optionalText.optional(),
 				retitle: bool.optional(),
 				relations: array(object({
@@ -4314,6 +4312,86 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 			return violations && violations.length > 0 ? violations.join("；") : "";
 		}
 		//#endregion
+		//#region ../contract/lib/time.js
+		/**
+		* The ONE place a stored timestamp turns into a string a human (or a model) reads.
+		*
+		* ── THE BUG THIS MODULE EXISTS FOR ───────────────────────────────────────────────────────────────
+		*
+		* `facts.created_at` / `updated_at` / `archived_at` (and `contradiction_log.detected_at`, and the
+		* retrieval-health stamp) are written by SQLite's `DEFAULT CURRENT_TIMESTAMP`, which is **UTC**, and
+		* `store/memory.ts` used to hand the raw text up to every surface. Two consequences, both reported as
+		* "记忆时间显示不对":
+		*
+		*  1. the UI printed `2026-10-28 11:58:48` as if it were a local wall clock (Asia/Shanghai really
+		*     meant `19:58:48`), because a zone-less string cannot say otherwise;
+		*  2. the MODEL received the same zone-less string next to `event_date`, which is a calendar DATE
+		*     with no time and no zone. Two different kinds of value, one shape — the model had to guess
+		*     which one it was looking at, and it cannot.
+		*
+		* ── THE SPLIT THIS MODULE ENFORCES ───────────────────────────────────────────────────────────────
+		*
+		* Storage stays UTC and stays byte-identical (`facts_io` export/import/snapshot is a compatibility
+		* surface and is deliberately NOT routed through here). The fix is in the two DISPLAY faces:
+		*
+		*  - {@link formatOffsetIso} — the MODEL face (`mem_recall` hits, `mem_admin list/detail/
+		*    contradictions`, `mem_admin stats`, the CLI and MCP that print those payloads): an ISO-8601
+		*    instant **with an explicit offset**, e.g. `2026-10-28T19:58:48+08:00`, rendered in the host's
+		*    local zone. The offset is part of the string, so the model needs no conversion and the value is
+		*    never ambiguous again — while `event_date` stays `2026-10-28`, a calendar date with no zone.
+		*  - {@link formatLocalHuman} — the HUMAN face (the 记忆 tab): `2026-10-28 19:58`, the same instant,
+		*    local, to the minute.
+		*
+		* Both read the input through the SAME parser ({@link parseUtcTs}) and the same instant, so the two
+		* faces cannot disagree. The UI deliberately formats the model face's OWN string rather than
+		* re-deriving one from the raw column: the input it gets already carries an offset, `parseUtcTs`
+		* leaves an explicit offset alone, and there is therefore exactly ONE parse rule and ONE pair of
+		* renderers in the repository — not a second opinion in the browser.
+		*
+		* Dependency-free on purpose: this module is reachable from the DSH client bundle through the
+		* `@avantf/mem-contract/time` subpath (the bundle must not pull in the tool schemas / zod). It has no
+		* DSH host and no DB, so the engine, the CLI and the MCP face can all use it.
+		*
+		* @module contract/time
+		*/
+		/** Two-digit, zero-padded — the only formatting primitive the two renderers share. */
+		function pad2(value) {
+			return String(value).padStart(2, "0");
+		}
+		/**
+		* `'YYYY-MM-DD HH:MM:SS'` (SQLite UTC) or ISO → epoch ms; null/garbage → null.
+		*
+		* A stamp with NO zone designator is read as UTC, not as local time. `Date.parse` treats a zone-less
+		* ISO string as LOCAL, and everything in this store is written in UTC (`CURRENT_TIMESTAMP`, and
+		* {@link formatUtcTs}), so handing one over as-is would shift every quota window and every decay
+		* measurement by the host's UTC offset — a fact reinforced at 23:00 UTC+8 would look reinforced 8
+		* hours earlier, and the 24h cap would open or close at the wrong moment. An explicit `Z` or
+		* `±HH:MM` is left alone (that is what makes a `formatOffsetIso` output safe to feed back in).
+		*/
+		function parseUtcTs(ts) {
+			if (ts === null || ts === void 0 || ts === "") return null;
+			const iso = ts.includes("T") ? ts : ts.replace(" ", "T");
+			const zoned = /(?:Z|[+-]\d{2}:?\d{2})$/i.test(iso) ? iso : `${iso}Z`;
+			const ms = Date.parse(zoned);
+			return Number.isFinite(ms) ? ms : null;
+		}
+		/**
+		* The HUMAN face: the same instant as local wall-clock text to the minute (`2026-10-28 19:58`).
+		*
+		* Seconds are deliberately dropped: the list/detail lines and the health line are read at a glance,
+		* and `HH:MM` is the whole useful resolution there. `null` for an absent/unreadable value (same
+		* contract as {@link formatOffsetIso}), so a caller prints its own 「未知」/`—` placeholder.
+		*
+		* NOT for `event_date` / `valid_from`: those are calendar dates the CALLER wrote (`2026-10-28`), with
+		* no time and no zone. Running one through here would invent a midnight in the host's zone.
+		*/
+		function formatLocalHuman(ts) {
+			const ms = parseUtcTs(ts);
+			if (ms === null) return null;
+			const at = new Date(ms);
+			return `${String(at.getFullYear())}-${pad2(at.getMonth() + 1)}-${pad2(at.getDate())} ${pad2(at.getHours())}:${pad2(at.getMinutes())}`;
+		}
+		//#endregion
 		//#region src/client/summarize.ts
 		/**
 		* Short, page-friendly rendering of a tool result — never a JSON wall.
@@ -4775,7 +4853,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 		* 2. Every remote call is normalized and surfaced: the raw response is logged
 		*    (prefixed `[avantf-mem]`) and rendered as an explicit status/error line, so
 		*    an empty result is never indistinguishable from a failed call.
-		* 3. Full DESIGN §12 surface: the tab does CRUD (edit/archive/restore/feedback) + fact detail
+		* 3. Full DESIGN §12 surface: the tab does CRUD (edit/archive/restore) + fact detail
 		*    (entities/triples) + open contradictions. Generation stays external (Plan A) — no answer
 		*    synthesis.
 		*
@@ -4792,6 +4870,19 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 		* panels cannot drift apart.
 		*/
 		const BTN = `${controls_module_css_default.button} ${controls_module_css_default.sm} ${controls_module_css_default.outline}`;
+		/**
+		* Every timestamp the 记忆 tab prints goes through here: local wall-clock text to the minute
+		* (`2026-10-28 19:58`), or the caller's placeholder when the host sent nothing readable.
+		*
+		* This is a PLACEHOLDER policy, not a formatter — the formatting itself is
+		* `formatLocalHuman` from `@avantf/mem-contract/time`, the same module the host renders the model
+		* face with. The value arriving here is an instant WITH an offset (the host's `created_at` etc.), so
+		* `Date.parse` inside that one renderer is unambiguous; nothing in this file parses a time itself.
+		*
+		* A calendar date (`event_date` / `valid_from`) must NOT come through here: it has no time and no
+		* zone, and rendering it would invent a midnight in the viewer's zone.
+		*/
+		const localTime = (ts, fallback = "—") => formatLocalHuman(ts) ?? fallback;
 		const NOT_MOUNTED = "avantfMem Remote 未挂载：本插件自带的 typert contribution 挂载失败（详见控制台错误），暂时无法读取数据。";
 		/** One 记忆 tab; the host `admin` list default is the same. */
 		const MEMORY_PAGE_SIZE = 50;
@@ -4814,7 +4905,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 		* How many `admin.detail` calls one search answer may trigger, to give its results the list's action
 		* row (status / pin state / full text). A hard cap so a large exact page or a wide relation answer
 		* cannot turn one keystroke into an unbounded fan-out; anything past the cap keeps the id-only
-		* actions (详情 / 编辑 when the excerpt is complete / 归档 / 反馈), which still work.
+		* actions (详情 / 编辑 when the excerpt is complete / 归档), which still work.
 		*/
 		const SEARCH_FACT_LOOKUPS_MAX = 20;
 		/**
@@ -5053,7 +5144,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 		*/
 		function FactDetailView(props) {
 			const fact = props.fact;
-			return h("div", { className: controls_module_css_default.meta }, `标题：${fact.title ?? "（无标题）"}`, h("br"), `创建于 ${fact.created_at} / 更新于 ${fact.updated_at ?? "—"}`, h("br"), `实体：${(fact.entities ?? []).join("、") || "（无）"}`, h("br"), `三元组：${(fact.triples ?? []).map((t) => `${t.subj} —${t.pred}→ ${t.obj}（${t.confidence.toFixed(2)}）`).join("；") || "（无）"}`, h("br"), `检索次数：${String(fact.retrieval_count ?? 0)}${fact.supersedes_id ? ` · 修订自 #${String(fact.supersedes_id)}` : ""} · 被断言 ${String(fact.assert_count ?? 1)} 次`, h("br"), `有效期：${fact.valid_from ?? "未知"} → ${fact.valid_to ?? (fact.status === "active" ? "仍在有效" : "未知")}` + (fact.superseded_by === null || fact.superseded_by === void 0 ? "" : ` · 被 #${String(fact.superseded_by)} 取代`), h("br"), `来源：${(fact.sources ?? []).length === 0 ? "（未记录）" : (fact.sources ?? []).map((s) => `${s.kind}:${s.ref}`).join("；")}`, h("br"), `关系：${(fact.relations ?? []).length === 0 ? "（无）" : (fact.relations ?? []).map((r) => `${r.subject} —${r.predicate}→ ${r.object}`).join("；")}`, fact.kind === "note" && fact.status === "active" ? h(RelationEditor, {
+			return h("div", { className: controls_module_css_default.meta }, `标题：${fact.title ?? "（无标题）"}`, h("br"), `创建于 ${localTime(fact.created_at)} / 更新于 ${localTime(fact.updated_at)}`, h("br"), `实体：${(fact.entities ?? []).join("、") || "（无）"}`, h("br"), `三元组：${(fact.triples ?? []).map((t) => `${t.subj} —${t.pred}→ ${t.obj}（${t.confidence.toFixed(2)}）`).join("；") || "（无）"}`, h("br"), `检索次数：${String(fact.retrieval_count ?? 0)}${fact.supersedes_id ? ` · 修订自 #${String(fact.supersedes_id)}` : ""} · 被断言 ${String(fact.assert_count ?? 1)} 次`, h("br"), `事件时间：${fact.valid_from ?? "未知"}` + (fact.superseded_by === null || fact.superseded_by === void 0 ? "" : ` · 被 #${String(fact.superseded_by)} 取代`), h("br"), `来源：${(fact.sources ?? []).length === 0 ? "（未记录）" : (fact.sources ?? []).map((s) => `${s.kind}:${s.ref}`).join("；")}`, h("br"), `关系：${(fact.relations ?? []).length === 0 ? "（无）" : (fact.relations ?? []).map((r) => `${r.subject} —${r.predicate}→ ${r.object}`).join("；")}`, fact.kind === "note" && fact.status === "active" ? h(RelationEditor, {
 				fact,
 				disabled: props.disabled,
 				onSave: (relations) => {
@@ -5076,7 +5167,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 				title: f.title ?? (f.kind === "journal" ? "（无标题）" : void 0),
 				text: f.content,
 				truncated: false,
-				meta: `#${String(f.fact_id)} · ${f.kind} · trust ${f.trust_score.toFixed(2)}` + (f.pinned === true ? " · 永久" : typeof f.remaining_days === "number" ? ` · 剩余 ≈${String(Math.ceil(f.remaining_days))} 活跃日` : "") + ` · 有用 ${String(f.helpful_count)} · ${f.created_at}` + (f.status === "archived" && f.archive_reason ? ` · 归档原因 ${f.archive_reason}` : "") + (f.status === "archived" && f.pinned === true ? " · 永久保留（不会被清理）" : ""),
+				meta: `#${String(f.fact_id)} · ${f.kind} · trust ${f.trust_score.toFixed(2)}` + (f.pinned === true ? " · 永久" : typeof f.remaining_days === "number" ? ` · 剩余 ≈${String(Math.ceil(f.remaining_days))} 活跃日` : "") + ` · ${localTime(f.created_at)}` + (f.status === "archived" && f.archive_reason ? ` · 归档原因 ${f.archive_reason}` : "") + (f.status === "archived" && f.pinned === true ? " · 永久保留（不会被清理）" : ""),
 				action: {
 					fact_id: f.fact_id,
 					status: f.status,
@@ -5109,7 +5200,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 				title: entry.title,
 				text: entry.content,
 				truncated: entry.truncated === true,
-				meta: `#${String(entry.fact_id)} · ${entry.memory_kind} · ${entry.status} · ${entry.event_date ?? entry.created_at}` + (entry.source_ref === void 0 ? "" : ` · ${entry.source_ref}`),
+				meta: `#${String(entry.fact_id)} · ${entry.memory_kind} · ${entry.status} · ${entry.event_date ?? localTime(entry.created_at)}` + (entry.source_ref === void 0 ? "" : ` · ${entry.source_ref}`),
 				action,
 				fromSearch: true
 			};
@@ -5123,7 +5214,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 				title: entry.title,
 				text: entry.text,
 				truncated: entry.truncated === true,
-				meta: `#${String(entry.fact_id)} · ${entry.memory_kind} · ${entry.created_at}`,
+				meta: `#${String(entry.fact_id)} · ${entry.memory_kind} · ${localTime(entry.created_at)}`,
 				action,
 				fromSearch: true
 			};
@@ -5331,7 +5422,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 			* A CONTINUATION page calls it too (REVIEW-2026-10-27 P2-2): without that, every row a scroll brought
 			* in kept the id-only action set until the query was run again. The lookup is per page, so the bound
 			* above still holds — one page can resolve at most {@link SEARCH_FACT_LOOKUPS_MAX} of its rows and the
-			* rest keep the id-only actions (详情 / 编辑 / 归档 / 反馈), exactly like the first page's overflow.
+			* rest keep the id-only actions (详情 / 编辑 / 归档), exactly like the first page's overflow.
 			* The record is MERGED, so an append never drops what the first page resolved.
 			*
 			* The guard is the SEARCH's, not the page request's (`searchFactsGuard`): a page landing while the
@@ -5733,7 +5824,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 			* The per-fact action buttons — ONE implementation for the admin list and every result shape.
 			*
 			* The regression this replaced: the actions were written inline in the list row, so when a search
-			* owned the view, a hit's 「详情」 had nowhere to render and there was no way to edit/archive/feedback
+			* owned the view, a hit's 「详情」 had nowhere to render and there was no way to edit/archive
 			* a fact found by search.
 			*
 			* `fromSearch` only decides what happens AFTER the write: a result set is what is on screen, so it
@@ -5794,28 +5885,6 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 							setDetail(null);
 						}
 					}) : null,
-					fact.status === "active" ? h(Btn, {
-						key: "helpful",
-						label: "有用 +1",
-						disabled: busy,
-						onClick: () => {
-							run(`反馈有用 #${String(fact.fact_id)}`, (remote) => remote.remember({
-								action: "helpful",
-								fact_id: fact.fact_id
-							}));
-						}
-					}) : null,
-					fact.status === "active" ? h(Btn, {
-						key: "unhelpful",
-						label: "没用 −1",
-						disabled: busy,
-						onClick: () => {
-							run(`反馈没用 #${String(fact.fact_id)}`, (remote) => remote.remember({
-								action: "unhelpful",
-								fact_id: fact.fact_id
-							}));
-						}
-					}) : null,
 					fact.status === "active" ? h(Btn, {
 						key: "archive",
 						label: "归档",
@@ -6158,7 +6227,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 			}))), error !== void 0 ? h("div", { className: controls_module_css_default.error }, `错误：${error}`) : null, notice !== void 0 ? h("div", { className: controls_module_css_default.status }, notice) : null, health !== null ? h(HealthBlock, { stats: health }) : null, contradictions !== null ? contradictions.length === 0 ? h("div", { className: controls_module_css_default.status }, "没有未处理的矛盾。") : h("ul", { className: controls_module_css_default.list }, contradictions.map((c) => h("li", {
 				key: c.contradiction_id,
 				className: controls_module_css_default.item
-			}, `#${String(c.fact_a)}「${c.content_a}」 ↔ #${String(c.fact_b)}「${c.content_b}」`, h("span", { className: controls_module_css_default.meta }, `score ${c.score.toFixed(2)} · ${c.detected_at ?? ""}`), props.remote !== void 0 ? h("div", { className: pages_module_css_default.row }, h(Btn, {
+			}, `#${String(c.fact_a)}「${c.content_a}」 ↔ #${String(c.fact_b)}「${c.content_b}」`, h("span", { className: controls_module_css_default.meta }, `score ${c.score.toFixed(2)} · ${localTime(c.detected_at, "")}`), props.remote !== void 0 ? h("div", { className: pages_module_css_default.row }, h(Btn, {
 				label: `判 #${String(c.fact_a)} 错`,
 				danger: true,
 				disabled: busy,
@@ -6205,7 +6274,7 @@ Set the \`cycles\` parameter to \`"ref"\` to resolve cyclical schemas with defs.
 			const pct = (v) => `${String(Math.round(v * 100))}%`;
 			const kinds = Object.entries(r.by_kind).map(([kind, v]) => `${kind} ${String(v.queries)} 次（空 ${String(v.zero_results)}）`).join(" · ");
 			const truncated = r.embedding_truncated + r.output_truncated;
-			return h("div", { className: pages_module_css_default.detail }, pendingVectors === 0 ? null : h("div", { className: controls_module_css_default.status }, `⚠ 向量空间待迁移：${String(pendingVectors)} 条 ACTIVE 向量来自旧嵌入空间（宽度不符 ${String(vectors?.stale ?? 0)} · 其他表示 ${String(vectors?.space_stale ?? 0)}）——语义腿对它们失效，正在后台分批重算；手动入口 \`avantf-mem vectors --fix\``), wal?.warning === null || wal?.warning === void 0 ? null : h("div", { className: controls_module_css_default.status }, `⚠ ${wal.warning}（当前 ${String(wal.bytes)} 字节）`), h("div", { className: controls_module_css_default.status }, `检索 ${String(r.queries)} 次 · 空结果 ${pct(r.zero_result_rate)} · 平均 ${r.avg_latency_ms.toFixed(1)}ms · 峰值 ${r.max_latency_ms.toFixed(1)}ms`), h("div", { className: controls_module_css_default.meta }, `语义腿在线 ${pct(r.semantic_live_rate)} · 平均命中 ${String(r.avg_results_per_query)} 条`), kinds === "" ? null : h("div", { className: controls_module_css_default.meta }, kinds), props.stats.sources === void 0 && props.stats.validity === void 0 ? null : h("div", { className: controls_module_css_default.meta }, `来源覆盖 ${props.stats.sources === void 0 ? "—" : pct(props.stats.sources.coverage)}（${String(props.stats.sources?.facts_with_source ?? 0)}/${String(props.stats.sources?.active ?? 0)}） · 事件时间覆盖 ${props.stats.validity === void 0 ? "—" : pct(props.stats.validity.coverage)}（${String(props.stats.validity?.facts_with_valid_from ?? 0)}/${String(props.stats.validity?.active ?? 0)}）`), h("div", { className: controls_module_css_default.meta }, truncated === 0 ? "未被任何预算截断。" : `截断：嵌入 ${String(r.embedding_truncated)} · 输出 ${String(r.output_truncated)}${r.output_truncated > 0 ? "（调大 retrieval.max_output_tokens 或用 max_tokens 参数）" : ""}`), r.updated_at === null ? null : h("div", { className: controls_module_css_default.meta }, `更新于 ${r.updated_at}`));
+			return h("div", { className: pages_module_css_default.detail }, pendingVectors === 0 ? null : h("div", { className: controls_module_css_default.status }, `⚠ 向量空间待迁移：${String(pendingVectors)} 条 ACTIVE 向量来自旧嵌入空间（宽度不符 ${String(vectors?.stale ?? 0)} · 其他表示 ${String(vectors?.space_stale ?? 0)}）——语义腿对它们失效，正在后台分批重算；手动入口 \`avantf-mem vectors --fix\``), wal?.warning === null || wal?.warning === void 0 ? null : h("div", { className: controls_module_css_default.status }, `⚠ ${wal.warning}（当前 ${String(wal.bytes)} 字节）`), h("div", { className: controls_module_css_default.status }, `检索 ${String(r.queries)} 次 · 空结果 ${pct(r.zero_result_rate)} · 平均 ${r.avg_latency_ms.toFixed(1)}ms · 峰值 ${r.max_latency_ms.toFixed(1)}ms`), h("div", { className: controls_module_css_default.meta }, `语义腿在线 ${pct(r.semantic_live_rate)} · 平均命中 ${String(r.avg_results_per_query)} 条`), kinds === "" ? null : h("div", { className: controls_module_css_default.meta }, kinds), props.stats.sources === void 0 && props.stats.validity === void 0 ? null : h("div", { className: controls_module_css_default.meta }, `来源覆盖 ${props.stats.sources === void 0 ? "—" : pct(props.stats.sources.coverage)}（${String(props.stats.sources?.facts_with_source ?? 0)}/${String(props.stats.sources?.active ?? 0)}） · 事件时间覆盖 ${props.stats.validity === void 0 ? "—" : pct(props.stats.validity.coverage)}（${String(props.stats.validity?.facts_with_valid_from ?? 0)}/${String(props.stats.validity?.active ?? 0)}）`), h("div", { className: controls_module_css_default.meta }, truncated === 0 ? "未被任何预算截断。" : `截断：嵌入 ${String(r.embedding_truncated)} · 输出 ${String(r.output_truncated)}${r.output_truncated > 0 ? "（调大 retrieval.max_output_tokens 或用 max_tokens 参数）" : ""}`), r.updated_at === null ? null : h("div", { className: controls_module_css_default.meta }, `更新于 ${localTime(r.updated_at)}`));
 		}
 		/** Client cordis-plugin name shown in boot diagnostics. */
 		const name = "avantf-mem-client";

package/lib/engine/mem/config/config_files.d.ts

@@ -15,7 +15,7 @@ import type { AvantfLogger } from '../../mem-contract/index.js';
  * UNCOMMENTED line is warned about by the loader, so a wrong name here would be a trap rather than a
  * hint. `config_files.spec.ts` pins the ones this template names.
  */
-export declare const DEFAULT_COMMON_CONFIG = "# avantf-mem \u516C\u5171\u914D\u7F6E\uFF08\u5206\u5C42\u7B2C\u2461\u5C42\uFF09\u3002\u8FD9\u91CC\u5168\u662F\u6CE8\u91CA\uFF1A\u5220\u6389/\u7559\u7A7A = \u7528\u5185\u5EFA\u9ED8\u8BA4\uFF0C\u53D6\u6D88\u6CE8\u91CA\u5373\u751F\u6548\u3002\n# \u5206\u5C42\uFF08\u4F4E\u2192\u9AD8\uFF09\uFF1A\u2460 \u5185\u5EFA\u9ED8\u8BA4 \u2192 \u2461 \u672C\u6587\u4EF6 \u2192 \u2462 \u540C\u76EE\u5F55\u7684 memory.yaml\uFF08\u7F3A\u5931=\u4E0D\u8986\u76D6\uFF09\n#               \u2192 \u2463 \u73AF\u5883\u53D8\u91CF \u2192 \u2464 \u63D2\u4EF6/CLI \u663E\u5F0F\u4F20\u53C2\n# \u6240\u6709\u53EF\u7F16\u8F91\u6587\u672C\u90FD\u5728 ~/.avantf \u4E0B\uFF1Aconfigs/*.yaml\uFF08\u672C\u76EE\u5F55\uFF09\u4E0E prompts/*.md\uFF08\u6A21\u578B\u63D0\u793A\u8BCD\uFF09\u3002\n\n# \u6570\u636E\u6839\u4E0D\u5728\u8FD9\u91CC\u914D\uFF1A\u5B83\u7531 \u2463 AVANTF_HOME\uFF08\u5BB6\u65CF\u6839\uFF1A\u540C\u65F6\u51B3\u5B9A\u6570\u636E\u6839\u4E0E models/tools\uFF09\u51B3\u5B9A \u2014\u2014\n# \u53EA\u60F3\u6362\u6570\u636E\u6839\u3001\u4E0D\u642C\u6A21\u578B\u7F13\u5B58\u65F6\u7528 CLI \u7684 \u2464 \u663E\u5F0F\u53C2\u6570 avantf-mem --data-home <path>\uFF08\u63D2\u4EF6\u5219\u662F profile \u7684\n# dataHome\uFF0C\u5C5E \u2461\uFF09\u3002\u672C\u6587\u4EF6\u5C31\u5728\u6570\u636E\u6839\u91CC\u9762\uFF0C\u89E3\u6790\u6839\u65F6\u8FD8\u6CA1\u8BFB\u5230\u5B83\uFF0C\u6240\u4EE5\u5199\u5728\u8FD9\u91CC\u7684 dataHome \u5BF9\u6570\u636E\u6839\u65E0\u6548\uFF08DESIGN \u00A73\uFF09\u3002\n\n# semantic:                           # \u5D4C\u5165\u540E\u7AEF\uFF08\u8BB0\u5FC6\u4E0E\u77E5\u8BC6\u5171\u7528\uFF0C\u8DE8\u5E93\u5206\u6570\u624D\u53EF\u6BD4\uFF09\n#   backend: local_bge\n#   local_model: Xenova/bge-base-zh-v1.5    # \u5FC5\u987B\u662F ONNX \u4ED3\u5E93\n#   dim: 768\n#   auto_download: true\n#   auto_migrate: true                # \u6362\u6A21\u578B/\u5BBD\u5EA6\u540E\u81EA\u52A8\u5206\u6279\u91CD\u7B97\u65E7\u7A7A\u95F4\u7684\u5411\u91CF\uFF08\u9ED8\u8BA4\u5F00\uFF09\uFF1B\u5173\u6389\u53EA\u5269\u542F\u52A8\u544A\u8B66 + vectors --fix\n#   self_heal:                        # \u8BED\u4E49\u4F9D\u8D56\u81EA\u6108\uFF08\u9ED8\u8BA4\u5173\uFF09\uFF1A\u5141\u8BB8\u542F\u52A8\u671F\u5B89\u88C5\u7F3A\u5931\u7684 @huggingface/transformers\n#     enabled: false                  # \u6253\u5F00\u540E\u8D70\u5E95\u5EA7 runtime-deps \u5171\u4EAB\u6839\uFF08base \u2265 0.7.0 / interface v6\uFF09\uFF1B\u5931\u8D25\u53EA\u964D\u7EA7\u5230 FTS+\u5B9E\u4F53\n#                                     # \u88C5\u54EA\u4E2A\u5305 / \u4ECE\u54EA\u88C5 / \u63D0\u793A\u8BF4\u4EC0\u4E48\uFF1A\u89C1\u63D2\u4EF6\u81EA\u5E26\u7684 res/runtime-deps.yaml\uFF08\u968F\u5305\u53D1\u5E03\uFF09\n#   native_self_heal:                 # \u539F\u751F\u53EF\u9009\u4F9D\u8D56\u81EA\u6108\uFF1Anodejieba / hnswlib-node\uFF08\u9ED8\u8BA4\u5173\uFF09\n#     enabled: false                  # \u88C5\u54EA\u4E2A\u5305\u3001\u4ECE\u54EA\u88C5\u3001\u63D0\u793A\u8BF4\u4EC0\u4E48\uFF0C\u90FD\u7531\u63D2\u4EF6\u81EA\u5E26\u7684 res/runtime-deps.yaml \u51B3\u5B9A\n#   \u26A0\uFE0F cache_dir / mirror \u7531\u5BB6\u65CF\u5E95\u5EA7 @avantf/dsh-plugin-base \u7BA1\u7406\uFF1A\u5199\u5728\u8FD9\u91CC\u4F1A\u88AB\u5FFD\u7565\u5E76\u544A\u8B66\u3002\n#      \u7F13\u5B58\u76EE\u5F55\u7528\u73AF\u5883\u53D8\u91CF AVANTF_MEM_MODEL_CACHE\uFF0C\u955C\u50CF\u7528 AVANTF_MEM_MODEL_MIRROR\uFF08\u6216 HF_ENDPOINT\uFF09\u3002\n\n# vectorStore:\n#   backend: auto                     # auto = \u8D8A\u8FC7\u9608\u503C\u81EA\u52A8\u4ECE local_numpy \u8FC1\u5230 hnswlib\n#   auto_thresholds:\n#     hnswlib: 2000\n#   hnswlib_ef_search: 256            # ANN \u7684\u53EC\u56DE/\u901F\u5EA6\u65CB\u94AE\uFF0C\u8BED\u6599\u53D8\u5927\u5148\u8C03\u5B83\n\n# retriever:                          # \u4E09\u6761\u68C0\u7D22\u817F\u5728\u878D\u5408\u91CC\u7684\u6743\u91CD\u4E0E\"\u76F8\u5173\u6027\u95E8\u69DB\"\n#   weight_semantic: 0.55\n#   weight_fts: 0.30\n#   weight_jaccard: 0.15\n#   # \u817F\u95E8\u69DB\u662F\u7EDD\u5BF9\u95E8\u69DB\uFF0C\u6253\u5728\u6BCF\u6761\u817F\u7684\u539F\u59CB\u5206\u4E0A\u3001fuse() \u4E4B\u524D\uFF1B0 = \u5173\u95ED\u8BE5\u95E8\u69DB\uFF0C\u5206\u6570\u7B49\u4E8E\u95E8\u69DB\u4FDD\u7559\u3002\n#   # \u8BED\u4E49\u540E\u7AEF\u4E0D\u53EF\u7528\u65F6 min_fts_terms \u7684\u751F\u6548\u503C\u653E\u5BBD\u5230 1\u3002\n#   min_semantic_similarity: 0.5      # \u8BED\u4E49\u817F\u4F59\u5F26\uFF1B0.5 \u53EA\u5BF9 bge-base-zh-v1.5/768\uFF08mean pooling\uFF09\u6807\u5B9A\u8FC7\uFF0C\u6362\u6A21\u578B\u8981\u91CD\u6807\n#   min_fts_terms: 2                  # FTS \u817F\uFF1A\u8FD9\u4E00\u884C\u547D\u4E2D\u51E0\u4E2A\u4E0D\u540C\u7684\u67E5\u8BE2\u8BCD\u5143\uFF08\u62C9\u4E01\u8BCD\u22655\u5B57\u7B26 + CJK 3-gram\uFF09\n#   min_jaccard: 0.2                  # \u5B9E\u4F53\u817F\uFF1A\u951A\u70B9 Jaccard \u7684**\u4E0B\u754C**\uFF1B\u5B9E\u9645\u8BC1\u636E\u95E8\u69DB\u518D\u6309\u951A\u70B9\u7A00\u6709\u5EA6\u52A0\u6743\uFF08\u00A720.19 \u7EED\uFF09\n#   # min_score \u662F\u552F\u4E00\u7684**\u663E\u793A\u5206**\u95E8\u69DB\uFF0C\u4E5F\u662F UI \u663E\u793A\u7684\u90A3\u4E2A score\uFF1A\u5408\u5E76\u6C60\u6309\u6700\u5927\u878D\u5408\u5206\u5F52\u4E00\u5316\u540E\u7684\u503C\uFF0C\n#   # \u662F**\u76F8\u5BF9\u91CF**\u2014\u2014\u540C\u4E00\u4E8B\u5B9E\u5728\u4E0D\u540C\u67E5\u8BE2\u91CC\u53EF\u80FD\u901A\u8FC7\u3001\u4E5F\u53EF\u80FD\u4E0D\u901A\u8FC7\uFF08\u53D6\u51B3\u4E8E\u540C\u4E00\u67E5\u8BE2\u91CC\u6709\u6CA1\u6709\u66F4\u5F3A\u7684\u5019\u9009\uFF09\u3002\n#   # \u5B83\u53EA\u5E94\u7528\u4E00\u6B21\uFF1A\u5728\u5408\u5E76\u540E\u7684\u5019\u9009\u6C60\u4E0A\u3001\u6309 kind/source \u8FC7\u6EE4\u4E0E slice \u4E4B\u524D\uFF08mem_recall \u4E0E UI\n#   # \u56E0\u6B64\u662F\u540C\u4E00\u4E2A\u523B\u5EA6\uFF09\uFF1B0 = \u5173\u95ED\uFF0C\u5206\u6570\u7B49\u4E8E\u95E8\u69DB\u4FDD\u7559\uFF1B\u653E\u5BBD\u6863\u53EA\u964D\u817F\u95E8\u69DB\uFF0C\u4E0D\u964D\u5B83\u3002\n#   # \u26A0\uFE0F \u88AB\u5B83\u4E22\u5F03\u7684\u6761\u76EE\u4E0D\u8FDB `dropped_by_floor`\uFF0C\u56E0\u6B64 UI \u7684\u300C\u4E25\u683C\u95E8\u69DB\u4E22\u5F03\u4E86 N \u6761\u300D**\u4E0D\u542B**\u8FD9\u4E00\u7C7B\n#   #    \uFF08\u4FE1\u5C01\u91CC\u5355\u72EC\u6709 dropped_by_score\uFF1B\u9762\u677F\u8FD9\u4E00\u7248\u8FD8\u6CA1\u6E32\u67D3\u5B83\uFF09\u3002\n#   min_score: 0.5\n#   # P-11 \u4E2D\u6587\u65F6\u95F4\u7A97\u817F\uFF08\u9ED8\u8BA4\u5173\uFF09\uFF1A\u5F00\u542F\u540E\uFF0C\u5E26\u4E2D\u6587\u65F6\u95F4\u8868\u8FBE\u7684\u67E5\u8BE2\uFF08\u4E0A\u4E2A\u6708 / \u6700\u8FD1\u4E09\u5929 / 2026\u5E749\u6708 \u2026\uFF09\n#   # \u4F1A\u989D\u5916\u628A\"\u4E8B\u4EF6\u65F6\u95F4 valid_from \u843D\u5728\u8BE5\u7A97\u53E3\"\u7684\u6D3B\u8DC3\u4E8B\u5B9E\u4F5C\u4E3A\u4E00\u6761\u817F\u53C2\u4E0E\u878D\u5408\u3002\u5B83\u662F**\u52A0\u4E00\u6761\u817F\u3001\u4E0D\u662F\u8FC7\u6EE4\u5668**\uFF1A\n#   # \u53EA\u589E\u52A0\u5019\u9009\uFF0C\u7EDD\u4E0D\u5220\u6389\u8BED\u4E49 / FTS / \u5B9E\u4F53\u817F\u5DF2\u7ECF\u627E\u5230\u7684\u4E8B\u5B9E\uFF1B\u6CA1\u6709\u4E8B\u4EF6\u65F6\u95F4\u7684\u4E8B\u5B9E\u4E0D\u4F5C\u4E3A\u8BE5\u817F\u5019\u9009\uFF0C\n#   # \u4E5F**\u4E0D\u56DE\u9000\u5230\u5199\u5165\u65F6\u95F4 created_at**\u3002\u5B83\u5171\u7528 jaccard \u6743\u91CD\u4F4D\uFF08\u5BF9\u5916\u7684\u4E09\u952E\u6743\u91CD\u4E0D\u53D8\uFF09\u3002\n#   # \u4EC0\u4E48\u65F6\u5019\u8BE5\u5F00\uFF1A\u5148\u7528 mem_admin stats \u770B validity.coverage\uFF08\u6709\u591A\u5C11\u6D3B\u8DC3\u4E8B\u5B9E\u5E26\u4E8B\u4EF6\u65F6\u95F4\uFF09\uFF0C\n#   # \u518D\u7528\u56DE\u5F52\u7F51\u786E\u8BA4\u5B83\u786E\u5B9E\u6539\u5584\u65F6\u95F4\u7C7B\u67E5\u8BE2\u2014\u2014\u8986\u76D6\u7387\u22480 \u65F6\u5B83\u7B49\u4E8E\u6CA1\u5F00\u3002\u5199\u5165\u4E8B\u4EF6\u65F6\u95F4\u7528 mem_remember \u7684 event_date\u3002\n#   time_window: false\n\n# lifecycle:\n#   purge_after_archived_days: 365    # \u5F52\u6863\u4E8B\u5B9E\u7684\u7269\u7406\u6E05\u7406\u7A97\u53E3\uFF08\u6D3B\u8DC3\u65E5\uFF09\n#   contradiction_threshold: 0.6\n\n# trust:\n#   decay_per_day: 0.0055\u2026\n#   presence:\n#     heartbeat_minutes: 60           # \u5E38\u9A7B\u8FDB\u7A0B\u7684\u5FC3\u8DF3\uFF08\u6D3B\u8DC3\u65E5\u65F6\u949F\u4E0E\u751F\u547D\u5468\u671F\u626B\u63CF\uFF09\n\n# tools:                              # \u5916\u90E8\u4E8C\u8FDB\u5236\uFF08pandoc\uFF09\u7684\u53D7\u7BA1\u76EE\u5F55\n#   dir: ''                           # \u7559\u7A7A = \u5BB6\u65CF\u6839 tools\uFF08~/.avantf/env/tools\uFF09\uFF1BAVANTF_TOOLS_DIR \u53EF\u8986\u76D6\n#   auto_install: true\n";
+export declare const DEFAULT_COMMON_CONFIG = "# avantf-mem \u516C\u5171\u914D\u7F6E\uFF08\u5206\u5C42\u7B2C\u2461\u5C42\uFF09\u3002\u8FD9\u91CC\u5168\u662F\u6CE8\u91CA\uFF1A\u5220\u6389/\u7559\u7A7A = \u7528\u5185\u5EFA\u9ED8\u8BA4\uFF0C\u53D6\u6D88\u6CE8\u91CA\u5373\u751F\u6548\u3002\n# \u5206\u5C42\uFF08\u4F4E\u2192\u9AD8\uFF09\uFF1A\u2460 \u5185\u5EFA\u9ED8\u8BA4 \u2192 \u2461 \u672C\u6587\u4EF6 \u2192 \u2462 \u540C\u76EE\u5F55\u7684 memory.yaml\uFF08\u7F3A\u5931=\u4E0D\u8986\u76D6\uFF09\n#               \u2192 \u2463 \u73AF\u5883\u53D8\u91CF \u2192 \u2464 \u63D2\u4EF6/CLI \u663E\u5F0F\u4F20\u53C2\n# \u6240\u6709\u53EF\u7F16\u8F91\u6587\u672C\u90FD\u5728 ~/.avantf \u4E0B\uFF1Aconfigs/*.yaml\uFF08\u672C\u76EE\u5F55\uFF09\u4E0E prompts/*.md\uFF08\u6A21\u578B\u63D0\u793A\u8BCD\uFF09\u3002\n\n# \u6570\u636E\u6839\u4E0D\u5728\u8FD9\u91CC\u914D\uFF1A\u5B83\u7531 \u2463 AVANTF_HOME\uFF08\u5BB6\u65CF\u6839\uFF1A\u540C\u65F6\u51B3\u5B9A\u6570\u636E\u6839\u4E0E models/tools\uFF09\u51B3\u5B9A \u2014\u2014\n# \u53EA\u60F3\u6362\u6570\u636E\u6839\u3001\u4E0D\u642C\u6A21\u578B\u7F13\u5B58\u65F6\u7528 CLI \u7684 \u2464 \u663E\u5F0F\u53C2\u6570 avantf-mem --data-home <path>\uFF08\u63D2\u4EF6\u5219\u662F profile \u7684\n# dataHome\uFF0C\u5C5E \u2461\uFF09\u3002\u672C\u6587\u4EF6\u5C31\u5728\u6570\u636E\u6839\u91CC\u9762\uFF0C\u89E3\u6790\u6839\u65F6\u8FD8\u6CA1\u8BFB\u5230\u5B83\uFF0C\u6240\u4EE5\u5199\u5728\u8FD9\u91CC\u7684 dataHome \u5BF9\u6570\u636E\u6839\u65E0\u6548\uFF08DESIGN \u00A73\uFF09\u3002\n\n# semantic:                           # \u5D4C\u5165\u540E\u7AEF\uFF08\u8BB0\u5FC6\u4E0E\u77E5\u8BC6\u5171\u7528\uFF0C\u8DE8\u5E93\u5206\u6570\u624D\u53EF\u6BD4\uFF09\n#   backend: local_bge\n#   local_model: Xenova/bge-base-zh-v1.5    # \u5FC5\u987B\u662F ONNX \u4ED3\u5E93\n#   dim: 768\n#   auto_download: true\n#   auto_migrate: true                # \u6362\u6A21\u578B/\u5BBD\u5EA6\u540E\u81EA\u52A8\u5206\u6279\u91CD\u7B97\u65E7\u7A7A\u95F4\u7684\u5411\u91CF\uFF08\u9ED8\u8BA4\u5F00\uFF09\uFF1B\u5173\u6389\u53EA\u5269\u542F\u52A8\u544A\u8B66 + vectors --fix\n#   self_heal:                        # \u8BED\u4E49\u4F9D\u8D56\u81EA\u6108\uFF08\u9ED8\u8BA4\u5173\uFF09\uFF1A\u5141\u8BB8\u542F\u52A8\u671F\u5B89\u88C5\u7F3A\u5931\u7684 @huggingface/transformers\n#     enabled: false                  # \u6253\u5F00\u540E\u8D70\u5E95\u5EA7 runtime-deps \u5171\u4EAB\u6839\uFF08base \u2265 0.7.0 / interface v6\uFF09\uFF1B\u5931\u8D25\u53EA\u964D\u7EA7\u5230 FTS+\u5B9E\u4F53\n#                                     # \u88C5\u54EA\u4E2A\u5305 / \u4ECE\u54EA\u88C5 / \u63D0\u793A\u8BF4\u4EC0\u4E48\uFF1A\u89C1\u63D2\u4EF6\u81EA\u5E26\u7684 res/runtime-deps.yaml\uFF08\u968F\u5305\u53D1\u5E03\uFF09\n#   native_self_heal:                 # \u539F\u751F\u53EF\u9009\u4F9D\u8D56\u81EA\u6108\uFF1Anodejieba / hnswlib-node\uFF08\u9ED8\u8BA4\u5173\uFF09\n#     enabled: false                  # \u88C5\u54EA\u4E2A\u5305\u3001\u4ECE\u54EA\u88C5\u3001\u63D0\u793A\u8BF4\u4EC0\u4E48\uFF0C\u90FD\u7531\u63D2\u4EF6\u81EA\u5E26\u7684 res/runtime-deps.yaml \u51B3\u5B9A\n#   \u26A0\uFE0F cache_dir / mirror \u7531\u5BB6\u65CF\u5E95\u5EA7 @avantf/dsh-plugin-base \u7BA1\u7406\uFF1A\u5199\u5728\u8FD9\u91CC\u4F1A\u88AB\u5FFD\u7565\u5E76\u544A\u8B66\u3002\n#      \u7F13\u5B58\u76EE\u5F55\u7528\u73AF\u5883\u53D8\u91CF AVANTF_MEM_MODEL_CACHE\uFF0C\u955C\u50CF\u7528 AVANTF_MEM_MODEL_MIRROR\uFF08\u6216 HF_ENDPOINT\uFF09\u3002\n\n# vectorStore:\n#   backend: auto                     # auto = \u8D8A\u8FC7\u9608\u503C\u81EA\u52A8\u4ECE local_numpy \u8FC1\u5230 hnswlib\n#   auto_thresholds:\n#     hnswlib: 2000\n#   hnswlib_ef_search: 256            # ANN \u7684\u53EC\u56DE/\u901F\u5EA6\u65CB\u94AE\uFF0C\u8BED\u6599\u53D8\u5927\u5148\u8C03\u5B83\n\n# retriever:                          # \u4E09\u6761\u68C0\u7D22\u817F\u5728\u878D\u5408\u91CC\u7684\u6743\u91CD\u4E0E\"\u76F8\u5173\u6027\u95E8\u69DB\"\n#   weight_semantic: 0.55\n#   weight_fts: 0.30\n#   weight_jaccard: 0.15\n#   # \u817F\u95E8\u69DB\u662F\u7EDD\u5BF9\u95E8\u69DB\uFF0C\u6253\u5728\u6BCF\u6761\u817F\u7684\u539F\u59CB\u5206\u4E0A\u3001fuse() \u4E4B\u524D\uFF1B0 = \u5173\u95ED\u8BE5\u95E8\u69DB\uFF0C\u5206\u6570\u7B49\u4E8E\u95E8\u69DB\u4FDD\u7559\u3002\n#   # \u8BED\u4E49\u540E\u7AEF\u4E0D\u53EF\u7528\u65F6 min_fts_terms \u7684\u751F\u6548\u503C\u653E\u5BBD\u5230 1\u3002\n#   min_semantic_similarity: 0.5      # \u8BED\u4E49\u817F\u4F59\u5F26\uFF1B0.5 \u53EA\u5BF9 bge-base-zh-v1.5/768\uFF08mean pooling\uFF09\u6807\u5B9A\u8FC7\uFF0C\u6362\u6A21\u578B\u8981\u91CD\u6807\n#   min_fts_terms: 2                  # FTS \u817F\uFF1A\u8FD9\u4E00\u884C\u547D\u4E2D\u51E0\u4E2A\u4E0D\u540C\u7684\u67E5\u8BE2\u8BCD\u5143\uFF08\u62C9\u4E01\u8BCD\u22655\u5B57\u7B26 + CJK 3-gram\uFF09\n#   min_jaccard: 0.2                  # \u5B9E\u4F53\u817F\uFF1A\u951A\u70B9 Jaccard \u7684**\u4E0B\u754C**\uFF1B\u5B9E\u9645\u8BC1\u636E\u95E8\u69DB\u518D\u6309\u951A\u70B9\u7A00\u6709\u5EA6\u52A0\u6743\uFF08\u00A720.19 \u7EED\uFF09\n#   # min_score \u662F\u552F\u4E00\u7684**\u663E\u793A\u5206**\u95E8\u69DB\uFF0C\u4E5F\u662F UI \u663E\u793A\u7684\u90A3\u4E2A score\uFF1A\u5408\u5E76\u6C60\u6309\u6700\u5927\u878D\u5408\u5206\u5F52\u4E00\u5316\u540E\u7684\u503C\uFF0C\n#   # \u662F**\u76F8\u5BF9\u91CF**\u2014\u2014\u540C\u4E00\u4E8B\u5B9E\u5728\u4E0D\u540C\u67E5\u8BE2\u91CC\u53EF\u80FD\u901A\u8FC7\u3001\u4E5F\u53EF\u80FD\u4E0D\u901A\u8FC7\uFF08\u53D6\u51B3\u4E8E\u540C\u4E00\u67E5\u8BE2\u91CC\u6709\u6CA1\u6709\u66F4\u5F3A\u7684\u5019\u9009\uFF09\u3002\n#   # \u5B83\u53EA\u5E94\u7528\u4E00\u6B21\uFF1A\u5728\u5408\u5E76\u540E\u7684\u5019\u9009\u6C60\u4E0A\u3001\u6309 kind/source \u8FC7\u6EE4\u4E0E slice \u4E4B\u524D\uFF08mem_recall \u4E0E UI\n#   # \u56E0\u6B64\u662F\u540C\u4E00\u4E2A\u523B\u5EA6\uFF09\uFF1B0 = \u5173\u95ED\uFF0C\u5206\u6570\u7B49\u4E8E\u95E8\u69DB\u4FDD\u7559\uFF1B\u653E\u5BBD\u6863\u53EA\u964D\u817F\u95E8\u69DB\uFF0C\u4E0D\u964D\u5B83\u3002\n#   # \u26A0\uFE0F \u88AB\u5B83\u4E22\u5F03\u7684\u6761\u76EE\u4E0D\u8FDB `dropped_by_floor`\uFF0C\u56E0\u6B64 UI \u7684\u300C\u4E25\u683C\u95E8\u69DB\u4E22\u5F03\u4E86 N \u6761\u300D**\u4E0D\u542B**\u8FD9\u4E00\u7C7B\n#   #    \uFF08\u4FE1\u5C01\u91CC\u5355\u72EC\u6709 dropped_by_score\uFF1B\u9762\u677F\u8FD9\u4E00\u7248\u8FD8\u6CA1\u6E32\u67D3\u5B83\uFF09\u3002\n#   min_score: 0.5\n#   # P-11 \u4E2D\u6587\u65F6\u95F4\u7A97\u817F\uFF08\u9ED8\u8BA4\u5173\uFF09\uFF1A\u5F00\u542F\u540E\uFF0C\u5E26\u4E2D\u6587\u65F6\u95F4\u8868\u8FBE\u7684\u67E5\u8BE2\uFF08\u4E0A\u4E2A\u6708 / \u6700\u8FD1\u4E09\u5929 / 2026\u5E749\u6708 \u2026\uFF09\n#   # \u4F1A\u989D\u5916\u628A\"\u4E8B\u4EF6\u65F6\u95F4 valid_from \u843D\u5728\u8BE5\u7A97\u53E3\"\u7684\u6D3B\u8DC3\u4E8B\u5B9E\u4F5C\u4E3A\u4E00\u6761\u817F\u53C2\u4E0E\u878D\u5408\u3002\u5B83\u662F**\u52A0\u4E00\u6761\u817F\u3001\u4E0D\u662F\u8FC7\u6EE4\u5668**\uFF1A\n#   # \u53EA\u589E\u52A0\u5019\u9009\uFF0C\u7EDD\u4E0D\u5220\u6389\u8BED\u4E49 / FTS / \u5B9E\u4F53\u817F\u5DF2\u7ECF\u627E\u5230\u7684\u4E8B\u5B9E\uFF1B\u6CA1\u6709\u4E8B\u4EF6\u65F6\u95F4\u7684\u4E8B\u5B9E\u4E0D\u4F5C\u4E3A\u8BE5\u817F\u5019\u9009\uFF0C\n#   # \u4E5F**\u4E0D\u56DE\u9000\u5230\u5199\u5165\u65F6\u95F4 created_at**\u3002\u5B83\u5171\u7528 jaccard \u6743\u91CD\u4F4D\uFF08\u5BF9\u5916\u7684\u4E09\u952E\u6743\u91CD\u4E0D\u53D8\uFF09\u3002\n#   # \u4EC0\u4E48\u65F6\u5019\u8BE5\u5F00\uFF1A\u5148\u7528 mem_admin stats \u770B validity.coverage\uFF08\u6709\u591A\u5C11\u6D3B\u8DC3\u4E8B\u5B9E\u5E26\u4E8B\u4EF6\u65F6\u95F4\uFF09\uFF0C\n#   # \u518D\u7528\u56DE\u5F52\u7F51\u786E\u8BA4\u5B83\u786E\u5B9E\u6539\u5584\u65F6\u95F4\u7C7B\u67E5\u8BE2\u2014\u2014\u8986\u76D6\u7387\u22480 \u65F6\u5B83\u7B49\u4E8E\u6CA1\u5F00\u3002\u5199\u5165\u4E8B\u4EF6\u65F6\u95F4\u7528 mem_remember \u7684 event_date\u3002\n#   time_window: false\n\n# lifecycle:\n#   purge_after_archived_days: 365    # \u5F52\u6863\u4E8B\u5B9E\u7684\u7269\u7406\u6E05\u7406\u7A97\u53E3\uFF08\u6D3B\u8DC3\u65E5\uFF09\n#   contradiction_threshold: 0.6\n\n# trust:                              # \u4FE1\u4EFB\u4E0E\u9057\u5FD8\uFF1A\u65CB\u94AE\u662F\u300C\u5929\u6570 + \u5F62\u72B6\u300D\uFF0C\u4E0D\u662F\u659C\u7387\n#   base_lifetime_days: 90            # \u951A\u70B9\uFF1A0.5 \u8D77\u70B9 \u2192 0 \u7684\u6D3B\u8DC3\u65E5\u6570\uFF08\u65B0\u4E8B\u5B9E\u4E0D\u88AB\u53EC\u56DE\u65F6\u7684\u5BFF\u547D\uFF09\n#   max_retention_days: 174           # \u53EC\u56DE\u52A0\u5F3A\u6700\u591A\u80FD\u6362\u6765\u7684\u6D3B\u8DC3\u65E5\u6570\uFF1Bceiling \u7531\u5B83\u53CD\u89E3\uFF08accelerating \u4E0B \u2248 0.7492\uFF09\n#   decay_shape: accelerating         # accelerating\uFF08\u9ED8\u8BA4\uFF09| linear\uFF08\u56DE\u6EDA\u6863\uFF1A\u914D 153 \u65F6 ceiling \u56DE\u5230 0.85\uFF09\n#   decay_curvature: 0.05             # accelerating \u7684\u4E0B\u9650\u9879 a\uFF0C\u5408\u6CD5\u533A\u95F4 (0, 0.5]\uFF1Blinear \u4E0B\u4E0D\u53C2\u4E0E\n#   recall_gain_below_base: 0.1       # \u4F4E\u4E8E\u57FA\u51C6 0.5 \u65F6\u6BCF\u6B21\u53EC\u56DE\u52A0\u591A\u5C11\uFF08\u5C01\u9876\u57FA\u51C6\uFF1A0.05 \u8D77 5 \u6B21\u5230 0.5\uFF09\n#   recall_gain_above_base: 0.01      # \u9AD8\u4E8E\u57FA\u51C6 0.5 \u65F6\u6BCF\u6B21\u53EC\u56DE\u52A0\u591A\u5C11\uFF08\u5C01\u9876\u4E0A\u9762\u53CD\u89E3\u51FA\u7684 ceiling\uFF09\n#   presence:\n#     heartbeat_minutes: 60           # \u5E38\u9A7B\u8FDB\u7A0B\u7684\u5FC3\u8DF3\uFF08\u6D3B\u8DC3\u65E5\u65F6\u949F\u4E0E\u751F\u547D\u5468\u671F\u626B\u63CF\uFF09\n\n# tools:                              # \u5916\u90E8\u4E8C\u8FDB\u5236\uFF08pandoc\uFF09\u7684\u53D7\u7BA1\u76EE\u5F55\n#   dir: ''                           # \u7559\u7A7A = \u5BB6\u65CF\u6839 tools\uFF08~/.avantf/env/tools\uFF09\uFF1BAVANTF_TOOLS_DIR \u53EF\u8986\u76D6\n#   auto_install: true\n";
 /**
  * Ensure the common config exists with content, and hand back the path either way.
  *