claude-mem-lite 6.23.3 → 6.25.0

package/.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "6.23.3",
+      "version": "6.25.0",
       "source": "./",
       "homepage": "https://github.com/sdsrss/claude-mem-lite",
       "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. FTS5 BM25 keyword search with episode batching. Single SQLite DB, no external services. A lighter, lower-cost alternative to claude-mem (episode batching + a smaller model; cost savings are an internal estimate, not a measured benchmark)."

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "6.23.3",
+  "version": "6.25.0",
   "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. FTS5 BM25 keyword search with episode batching. Single SQLite DB, no external services. A lighter, lower-cost alternative to claude-mem (episode batching + a smaller model; cost savings are an internal estimate, not a measured benchmark).",
   "author": {
     "name": "sdsrss"

package/README.md

@@ -237,6 +237,44 @@ rm -rf ~/claude-mem-lite/   # pre-v0.5 unhidden (if not auto-moved)
     repos/               # Shallow-cloned source repos
 ```
 
+## Upgrading to 6.25.0
+
+**Every session gets the same first line on a file-recall block again.** No schema-version
+change: 6.24.0 still opens the database after this release has.
+
+- **What changes.** Since 6.16.0 half of sessions got a plain statement of where the notes come
+  from instead of "system-injected context, continue your planned action", so the two could be
+  compared by how often the agent cited a lesson. On the maintainer's own sessions the comparison
+  found no difference it could detect, and a scan of the agent's replies found neither line
+  surfaced to the user, so the split is closed: unset now means the older line in every session.
+  `CLAUDE_MEM_RECALL_FRAMING=factual` keeps the plain statement; `=ab` restores the split.
+- **Also in this release:** when the data directory cannot be read, `doctor` says "not checked"
+  for what it could not look at and `status` names the directory, instead of reporting a missing
+  or empty store. `maintain execute --ops dedup` names the merge pairs it skips. Details in
+  CHANGELOG.md.
+
+## Upgrading to 6.24.0
+
+**If you set `CLAUDE_CONFIG_DIR`, the installer, updates and hooks now use that directory, as
+Claude Code does.** No schema-version change: 6.23.3 still opens the database after this
+release has. Without `CLAUDE_CONFIG_DIR` nothing in this item changes.
+
+- **Who is affected.** Claude Code keeps `settings.json`, `plugins/` and `.claude.json` in
+  `CLAUDE_CONFIG_DIR` when it is set. The plugin kept using `~/.claude` and `~/.claude.json`:
+  `install` wrote its hooks where Claude Code never reads them, and with two profiles the
+  session-start setup and `self-update` pruned the other profile's plugin cache and edited its
+  `.claude.json`.
+- **What to do.** Installed with npm or npx: run `claude-mem-lite install` once more with the
+  variable set, so the hooks land in your config directory. Plugin installs need nothing. Hook
+  entries an older version wrote into `~/.claude/settings.json` stay there; if no other profile
+  uses `~/.claude`, `env -u CLAUDE_CONFIG_DIR claude-mem-lite cleanup-hooks` removes them.
+- **Also in this release:** `npm install -g claude-mem-lite` and `npx` no longer install the
+  development toolchain. As a dependency install it went from 298 packages / 541 MB to 96 / 57 MB;
+  after the release, `-g` and `npx` measured the same 96 packages. A confirmed
+  `maintain execute --ops purge_stale` now deletes every row its preview counted in one run,
+  behind one snapshot. `doctor` warns about a failing update check and about an
+  `ANTHROPIC_BASE_URL` that ends in `/v1`. Details in CHANGELOG.md.
+
 ## Upgrading to 6.23.1
 
 **If you set `ANTHROPIC_BASE_URL` and `ANTHROPIC_API_KEY`, background LLM calls now go to that
@@ -568,9 +606,13 @@ depends on the Claude Code version. Allow the server once in the `settings.json`
 { "permissions": { "allow": ["mcp__plugin_claude-mem-lite_mem-lite"] } }
 ```
 
-That is the server's name under a plugin install. An npx or git-clone install registers it as
+For a single run, `claude -p … --allowedTools "mcp__plugin_claude-mem-lite_mem-lite"` does the
+same (both forms checked on 2.1.292, 2026-10-07: no refusal, and `mem_search` answered). That is
+the server's name under a plugin install. An npx or git-clone install registers it as
 `mem-lite`, so the rule there is `mcp__mem-lite`. An interactive session asks instead; answering
-"always allow" there records a rule of its own.
+"always allow" there records a rule of its own. The plugin does not tell the model to fall back
+to the CLI when a call is refused: the refusal is Claude Code's permission check for a tool you
+have not allowed, and routing the same action through Bash would sidestep it.
 
 ### Skill Commands (in Claude Code chat)
 
@@ -1209,7 +1251,7 @@ and names can change between releases.
 |----------|-------------|---------|
 | `CLAUDE_MEM_TASK_IMPERATIVE` | `on`/`1` injects the single most relevant lesson at prompt position under an imperative template. | _(off)_ |
 | `CLAUDE_MEM_SUBAGENT_INJECT` | Dispatch-time memory injection for subagents. | _(off)_ |
-| `CLAUDE_MEM_RECALL_FRAMING` | First line of a PreToolUse / PostToolUse recall block. `ab` gives each session one of two wordings, the older "system-injected context, continue your planned action" or a plain statement of source, so their cite-rates can be compared in one run (`benchmark/citation-live-replay.mjs --by-framing`); `legacy` / `factual` pin one. | `ab` |
+| `CLAUDE_MEM_RECALL_FRAMING` | First line of a PreToolUse / PostToolUse recall block. Unset or `legacy` = "system-injected context, continue your planned action"; `factual` = a plain statement of where the notes come from; `ab` = each session gets one of the two, for `benchmark/citation-live-replay.mjs --by-framing`. Any other value is `legacy`. | `legacy` |
 | `CLAUDE_MEM_SALIENCE` | Selects how the pre-edit lesson line asks for a response: unset = name a lesson's `#NN` only where it changed the edit; `verdict` = the pre-6.17 per-lesson `applied` / `n/a` verdict (the adoption row in CLAUDE.md and the detail doc keep the new wording, and since 6.17.1 the directive says it overrides them); `bind` / `bridge` = comprehension-bridge arms (`bridge` keeps the pre-6.17 verdict wording as its fallback; neither says it overrides the adopted text, so their measured wording stays fixed); `legacy` = no directive. | _(unset)_ |
 | `CLAUDE_MEM_EDGE_DECAY` | Enables decay of file↔observation edges. | _(off)_ |
 | `CLAUDE_MEM_EDGE_DECAY_K` | Edge-decay threshold when the flag above is on (clamped to ≥1). | `3` |

package/README.zh-CN.md

@@ -199,6 +199,35 @@ rm -rf ~/claude-mem-lite/   # v0.5 前的非隐藏目录（如未自动迁移）
     repos/               # 浅克隆的源代码仓库
 ```
 
+## 升级到 6.25.0
+
+**文件召回块的第一行，所有会话重新用同一种写法。** 没有 schema 版本变更：本版本打开过的数据库，6.24.0 仍能打开。
+
+- **变了什么。** 从 6.16.0 起，一半会话的第一行是说明笔记来源的事实陈述，另一半保留 "system-injected context,
+  continue your planned action"，用来比较 agent 引用教训的频率。在维护者自己的会话上，两种写法看不出差别，
+  按回复文本扫描，两组里 agent 也都没有把这一行当作可疑内容告诉用户，所以结束分组：
+  不设变量时，所有会话都用旧的那一行。
+  `CLAUDE_MEM_RECALL_FRAMING=factual` 保留事实陈述的写法，`=ab` 恢复分组。
+- **本版本还有：** 数据目录读不了时，`doctor` 对看不到的项目显示“not checked”，`status` 直接指出这个目录
+  无法访问，不再报成数据库缺失或为空。`maintain execute --ops dedup` 会列出跳过的合并对。详见 CHANGELOG.md。
+
+## 升级到 6.24.0
+
+**如果你设置了 `CLAUDE_CONFIG_DIR`，安装程序、更新和 hook 现在都使用这个目录，和 Claude Code 一致。**
+没有 schema 版本变更：本版本打开过的数据库，6.23.3 仍能打开。没设 `CLAUDE_CONFIG_DIR` 时，这一条不带来任何变化。
+
+- **受影响的是谁。** 设了 `CLAUDE_CONFIG_DIR` 时，Claude Code 把 `settings.json`、`plugins/` 和 `.claude.json`
+  都放在那个目录里，插件却一直用 `~/.claude` 和 `~/.claude.json`：`install` 把 hook 写到了 Claude Code 不读的
+  地方；有两个配置目录时，会话启动的整理步骤和 `self-update` 会清掉另一个配置目录的插件缓存，并改动它的
+  `.claude.json`。
+- **要做什么。** 用 npm 或 npx 安装的：在设好这个变量的情况下再运行一次 `claude-mem-lite install`，让 hook 写进
+  你的配置目录。插件安装不用做什么。旧版本写进 `~/.claude/settings.json` 的 hook 条目会留在那里；如果没有别的
+  配置目录在用 `~/.claude`，可以用 `env -u CLAUDE_CONFIG_DIR claude-mem-lite cleanup-hooks` 清掉。
+- **本版本还有：** `npm install -g claude-mem-lite` 和 `npx` 不再安装开发工具链。按依赖方式安装，从 298 个包 /
+  541 MB 降到 96 个包 / 57 MB；发布后实测 `-g` 和 `npx` 也是同样的 96 个包。确认执行的
+  `maintain execute --ops purge_stale` 现在一次运行就删掉预览里统计的全部行，只做一次快照。`doctor` 会对失败的
+  更新检查、以及以 `/v1` 结尾的 `ANTHROPIC_BASE_URL` 发出警告。详见 CHANGELOG.md。
+
 ## 升级到 6.23.1
 
 **如果你设置了 `ANTHROPIC_BASE_URL` 和 `ANTHROPIC_API_KEY`，后台 LLM 调用现在会发往那个网关。**

package/bash-utils.mjs

@@ -632,6 +632,28 @@ function withoutDiffHunks(text) {
 // Paths excluded from observation capture (ephemeral / virtual filesystems) — applied
 // uniformly to both command-parsed paths and direct file_path/path/filePath fields.
 
+// A line a test runner prints for one test (`# Subtest:` heads a failing one too): it carries the
+// test's name, not a summary.
+const PASSED_TEST_LINE_RE = /^[ \t]*(?:✔|✓|ok[ \t]+\d+[ \t]+-|# Subtest:)/;
+
+/**
+ * Whether test-runner output carries a red summary or failure banner (detectBashSignificance's
+ * veto on a green "0 fail" marker; D#179).
+ * @param {string} scan
+ * @returns {boolean}
+ */
+function hasRedTestSummary(scan) {
+  const lines = scan
+    .split('\n')
+    .filter((l) => !PASSED_TEST_LINE_RE.test(l))
+    .join('\n');
+  return (
+    /\b[1-9]\d*[ \t]+(?:fail|failed|failures|failing|errors?)\b(?=[ \t]*(?:$|[,;|(]|\.(?:[ \t]|$)|in[ \t]+\S))|^[ \t]*(?:ℹ|#)[ \t]*(?:fail|cancelled)[ \t]+[1-9]\d*[ \t]*$/im.test(
+      lines,
+    ) || /^[ \t]*(?:--- )?FAIL(?:ED)?(?::|[ \t]|$)/m.test(lines)
+  );
+}
+
 /**
  * Detect significance signals in a Bash command and its response.
  * Checks for errors, test runs, builds, git operations, and deployments.
@@ -675,12 +697,23 @@ export function detectBashSignificance(input, response) {
   // suites prints one summary each, and a node:test run whose failures were all timeouts says
   // "fail 0" next to "cancelled 1" (pre-tag defect review, item 7 F1). Red = a nonzero fail /
   // cancelled count in either summary form.
+  // D#179: a count-first red must sit where a summary puts it — end of line, or before `,` `;`
+  // `|` `(` `. ` or `in …` — on ONE line, and never on a line a runner prints for a PASSED test
+  // (node:test `✔ name (1ms)`, vitest / bun `✓ name`, TAP `ok 1 - name` and its `# Subtest:`),
+  // which carries the test's NAME. Matched anywhere and across newlines, a passing run whose test
+  // name or log said "3 failed" / "2 failed attempts" read red, and "want 2\nFAIL" read as
+  // "2 FAIL". `errors?` is red too (pytest "1 passed, 1 error", tsc "Found 1 error in …",
+  // eslint "(2 errors, 0 warnings)"). go / jest / vitest / shell-suite failures carry a
+  // case-sensitive FAIL / FAILED banner at line start ("--- FAIL:", "FAILED: 3 case(s)", pytest
+  // "FAILED test_x.py::…"), read on its own. The green marker still
+  // spans lines on purpose: made one-line, 17 of this machine's 77,819 recorded results flipped
+  // to errors, and 15 of them were passing batteries, file reads and reports that only ever
+  // matched "0\nFAIL" by accident; 2 were real failures (measured 2026-10-07, recounted by the
+  // pre-tag delta review).
   const hasGreenTestSummary =
     looksLikeError &&
     /\b0\s+(fail|failed|failures)\b|^[ \t]*(?:ℹ|#)[ \t]*fail[ \t]+0[ \t]*$/im.test(scan) &&
-    !/\b[1-9]\d*\s+(fail|failed|failures|failing)\b|^[ \t]*(?:ℹ|#)[ \t]*(?:fail|cancelled)[ \t]+[1-9]\d*[ \t]*$/im.test(
-      scan,
-    );
+    !hasRedTestSummary(scan);
   // NOTE: do not add `\bFAIL\s` here — with /i flag it would re-match the
   // very `0 fail\n` token green-summary is trying to exempt. A real test
   // failure produces "N fail" (N≥1) which never triggers hasGreenTestSummary,

package/cli/activity.mjs

@@ -188,7 +188,16 @@ export async function cmdActivity(db, args) {
       return;
     }
 
-    const result = db.prepare(`DELETE FROM events WHERE id IN (${placeholders})`).run(...ids);
+    // The opt-in search telemetry keeps a title per returned result; an explicit delete drops the
+    // deleted events' titles, as lib/delete-core.mjs does for observations (D#245 N6).
+    const result = db.transaction(() => {
+      if (db.prepare("SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = 'search_results'").get()) {
+        db.prepare(
+          `UPDATE search_results SET snapshot_label = NULL WHERE source = 'event' AND result_id IN (${placeholders})`,
+        ).run(...ids);
+      }
+      return db.prepare(`DELETE FROM events WHERE id IN (${placeholders})`).run(...ids);
+    })();
     out(`[mem] Deleted ${result.changes} event(s).`);
     return;
   }

package/haiku-client.mjs

@@ -20,6 +20,7 @@ import { randomUUID } from 'crypto';
 import { debugLog, debugCatch, parseJsonFromLLM } from './utils.mjs';
 import { DB_DIR } from './schema.mjs';
 import { resolveRuntimeDir } from './lib/resolve-data-dir.mjs';
+import { workerDirGone } from './lib/worker-data-dir.mjs';
 import { httpConnectProxyFor, postViaConnectProxy, postDirectHttp } from './lib/proxy-fetch.mjs';
 import { resolveAnthropicBaseUrl } from './lib/anthropic-base-url.mjs';
 
@@ -40,6 +41,9 @@ import { resolveAnthropicBaseUrl } from './lib/anthropic-base-url.mjs';
  */
 function cliSpawnCwd() {
   const dir = join(resolveRuntimeDir(DB_DIR), 'cli-cwd');
+  // A background worker whose data dir was removed during its API attempt must not re-create
+  // it here (D#265); the spawn then fails on the missing cwd, and the worker gives up.
+  if (workerDirGone(DB_DIR)) return dir;
   try {
     mkdirSync(dir, { recursive: true, mode: 0o700 });
   } catch {
@@ -555,8 +559,8 @@ async function callModelAPI(prompt, model, { timeout, maxTokens, temperature = D
     };
     // Plain http is loopback-only (resolveAnthropicBaseUrl), and loopback must stay on this
     // machine: native fetch would hand it, x-api-key in cleartext, to HTTP_PROXY under
-    // NODE_USE_ENV_PROXY=1. (D#250) Parsed, not a prefix test: the resolver accepts the scheme
-    // in any case (`HTTP://127.0.0.1`) and returns the raw text.
+    // NODE_USE_ENV_PROXY=1. (D#250) Parsed, not a prefix test: a prefix test let `HTTP://127.0.0.1`
+    // through to the proxy when the resolver still returned the raw text (b59e0cec, D#259).
     const plainHttp = new URL(apiUrl).protocol === 'http:';
     const send = (payload) => {
       const json = JSON.stringify(payload);

package/hook-optimize.mjs

@@ -484,8 +484,8 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
       // clampImportance floor it to 1 (kept visible, low-ranked) instead of hiding.
       // A compression keeper is never hidden: hiding it hides every member compressed into it
       // (the same rule NOT_COMPRESSION_KEEPER_SQL enforces on the maintenance writers). It falls
-      // through to the normal update instead: floored to importance 1 and stamped, so it is not
-      // re-sent. In narrow scope that update still replaces its title and narrative.
+      // through instead: in narrow scope to the stamp-only write below (floored to importance 1
+      // and stamped, so it is not re-sent), in wide scope to the normal update.
       const isKeeper = !!db
         .prepare('SELECT 1 FROM observations WHERE compressed_into = ? LIMIT 1')
         .get(cand.id);
@@ -494,12 +494,8 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
       const humanSet =
         (db.prepare('SELECT importance_set_at FROM observations WHERE id = ?').get(cand.id)
           ?.importance_set_at ?? null) !== null;
-      if (
-        (parsed.importance === 0 || parsed.importance === '0') &&
-        scope !== 'wide' &&
-        !isKeeper &&
-        !humanSet
-      ) {
+      const scoredZero = parsed.importance === 0 || parsed.importance === '0';
+      if (scoredZero && scope !== 'wide' && !isKeeper && !humanSet) {
         // D#12, and this one is not a stale-write guard — it is a POINTER guard.
         // `compressed_into` is the child -> keeper link, and COMPRESSED_AUTO is -1. If a
         // concurrent cluster-merge or smart-compress adopts this row during the 45 s Haiku
@@ -511,7 +507,8 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
         const res = db
           .prepare(
             `UPDATE observations SET compressed_into = ${COMPRESSED_AUTO}, optimized_at = ?
-             WHERE id = ? AND ${liveObsFilterSql('')} AND optimized_at IS NULL AND ${NOT_COMPRESSION_KEEPER_SQL}`,
+             WHERE id = ? AND ${liveObsFilterSql('')} AND optimized_at IS NULL AND ${NOT_COMPRESSION_KEEPER_SQL}
+               AND importance_set_at IS NULL`,
           )
           .run(Date.now(), cand.id);
         if (res.changes === 0) {
@@ -521,6 +518,31 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
         processed++;
         continue;
       }
+      // A narrow 0 reply reaches this point only for a row that may not be hidden (a compression
+      // keeper, or an importance a person set). It judged the row worthless, so it is no basis for
+      // any of the row's text either: the stored title and narrative stay (D#207: 'Weekly summary:
+      // auth refactor' became 'Weekly summary', narrative 'x'), and so do type, lesson, concepts,
+      // facts, aliases and scope (D#268). Written instead: what hiding would have written bar the
+      // hide, i.e. the stamp, so the row is not re-sent, plus the importance floor below. Wide
+      // scope stays on the normal path: its rows are substantive by construction, so a 0 there
+      // reads as a misjudgment of importance, not of the row.
+      if (scoredZero && scope !== 'wide') {
+        const res = db
+          .prepare(
+            `UPDATE observations SET
+               importance = CASE WHEN importance_set_at IS NOT NULL THEN importance
+                                 ELSE MAX(?, COALESCE(NULLIF(importance, 0), 1)) END,
+               optimized_at = ?
+             WHERE id = ? AND ${liveObsFilterSql('')} AND optimized_at IS NULL`,
+          )
+          .run(clampImportance(parsed.importance), Date.now(), cand.id);
+        if (res.changes === 0) {
+          skipped++;
+          continue;
+        }
+        processed++;
+        continue;
+      }
 
       // Enrichment ("add a lesson") must not reclassify a specific type down to the generic
       // 'change' (lower TYPE_QUALITY + faster decay); keep the stored type on that downgrade.
@@ -562,8 +584,9 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
       // fallback — truncating the row's own stored narrative to buy nothing was the same
       // bug in miniature.
       const isWide = scope === 'wide';
-      const title = isWide ? cand.title : truncate(scrubSecrets(parsed.title || ''), 120);
-      const narrative = isWide
+      const keepStoredText = isWide;
+      const title = keepStoredText ? cand.title : truncate(scrubSecrets(parsed.title || ''), 120);
+      const narrative = keepStoredText
         ? cand.narrative
         : parsed.narrative
           ? truncate(scrubSecrets(parsed.narrative), 500)
@@ -571,9 +594,10 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
       // Floor at the stored importance: re-enrich adds a lesson, it must never silently downgrade
       // a user-set/promoted importance (the UPDATE also sets optimized_at → the loss is permanent).
       // Upgrades are still honored — except over an importance a person set (D10), which stays.
-      const importance = humanSet
-        ? cand.importance
-        : Math.max(clampImportance(parsed.importance), cand.importance || 1);
+      // Decided in the UPDATE itself, on the row as it is then: the candidate's importance was read
+      // before the model call, and writing it back undid a person's change made during the call
+      // (D#206).
+      const modelImportance = clampImportance(parsed.importance);
 
       const bigramText = cjkBigrams((title || '') + ' ' + (narrative || ''));
       const textField = [conceptsText, factsText, searchAliases || '', bigramText].filter(Boolean).join(' ');
@@ -607,10 +631,10 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
       // Narrow replaces title and narrative with model text, so an explicit save it rewrites (one
       // whose save-time enrich failed) moves to the re-enrich writer's id, as a cluster-merge
       // keeper does (D#146, D#138). Wide keeps the stored title and narrative (it fills the lesson
-      // and the side fields), so the row stays an explicit save. The writer's session row is
-      // best-effort.
+      // and the side fields), so the row stays an explicit save; a narrow 0 reply never gets here.
+      // The writer's session row is best-effort.
       let rewriteSessionId = null;
-      if (!isWide) {
+      if (!keepStoredText) {
         const cur = db.prepare('SELECT memory_session_id FROM observations WHERE id = ?').get(cand.id);
         if (cur?.memory_session_id?.startsWith(MANUAL_SESSION_ID_PREFIX)) {
           const enrichSessionId = writerSessionId('enrich-', cand.project);
@@ -629,7 +653,10 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
         .prepare(
           `
         UPDATE observations SET type=?, title=?, narrative=?, concepts=?, facts=?,
-          text=?, importance=?, lesson_learned=?, search_aliases=?, minhash_sig=?, optimized_at=?,
+          text=?,
+          importance = CASE WHEN importance_set_at IS NOT NULL THEN importance
+                            ELSE MAX(?, COALESCE(NULLIF(importance, 0), 1)) END,
+          lesson_learned=?, search_aliases=?, minhash_sig=?, optimized_at=?,
           scope=COALESCE(?, scope), memory_session_id = COALESCE(?, memory_session_id)
         WHERE id = ? AND ${liveObsFilterSql('')} AND optimized_at IS NULL
       `,
@@ -641,7 +668,7 @@ scope: ${SCOPE_PROMPT_LEGEND}`;
           safe.concepts,
           safe.facts,
           safe.text,
-          importance,
+          modelImportance,
           safe.lesson_learned,
           safe.search_aliases,
           minhashSig,
@@ -1300,6 +1327,8 @@ Return ONLY valid JSON:
     const bigramText = cjkBigrams((title || '') + ' ' + (narrative || ''));
     const textField = [conceptsText, factsText, bigramText].filter(Boolean).join(' ');
     const minhashSig = computeMinHash((title || '') + ' ' + (narrative || ''));
+    // An importance a person set on the keeper stays (D10); the UPDATE decides that on the row as
+    // it is then, since a person can set it while this call waits on the model (D#198).
     const importance = Math.max(clampImportance(parsed.importance || 2), maxClusterImportance);
 
     // Scrub LLM-output cluster-merge text fields at the UPDATE boundary.
@@ -1382,7 +1411,8 @@ Return ONLY valid JSON:
       db.prepare(
         `
         UPDATE observations SET title=?, narrative=?, concepts=?, facts=?, text=?,
-          importance=?, lesson_learned=?, minhash_sig=?, optimized_at=?,
+          importance = CASE WHEN importance_set_at IS NOT NULL THEN importance ELSE ? END,
+          lesson_learned=?, minhash_sig=?, optimized_at=?,
           memory_session_id = COALESCE(?, memory_session_id)
         WHERE id = ?
       `,

package/hook-shared.mjs

@@ -30,6 +30,7 @@ import {
 } from './lib/schema-skew.mjs';
 import { isDbUnusableError, DB_UNUSABLE_MARKER_PREFIX } from './lib/db-unusable.mjs';
 import { shouldRecordOnce } from './lib/record-once.mjs';
+import { backgroundWorkerEnv, workerDirGone } from './lib/worker-data-dir.mjs';
 import { hookSessionId } from './lib/provenance.mjs';
 import { PROJECT_REKEY_MARKER_PREFIX } from './lib/project-rekey.mjs';
 // Audit 2026-09-05 P1-2 (carried from 2026-09-02 P2-9): `callLLM`, the quiet/adoption
@@ -280,7 +281,8 @@ export const GC_PRESERVED_MARKER_PREFIXES = Object.freeze([
   // out. Both are version-keyed one-shot migration sentinels written by
   // scripts/setup.sh, and their gate is `! -f <marker>` — deleting one re-runs
   // its migration. `.mcp-dedup-v2.78` gates a block that removes
-  // mcpServers.mem / mcpServers["mem-lite"] from the user's ~/.claude.json with
+  // mcpServers.mem / mcpServers["mem-lite"] from the config home's .claude.json (one marker per
+  // .claude.json since D#270, the same prefix) with
   // a raw writeFileSync (no tmp+rename, no backup), which the repo's own test
   // documents as intentionally one-shot: "If a user later runs `claude mcp add
   // mem ...` themselves, the gate intentionally lets it stand." A 30-day sweep
@@ -378,9 +380,11 @@ export function sweepStaleProjectMarkers(
 // exists, so harden here too: create 0700, and chmod a pre-existing dir a prior version
 // created at the default umask. A 0700 dir blocks traversal to every file inside,
 // current and future, regardless of individual file mode (audit sec P3-2 2026-07-24).
+// A background worker whose data dir was removed while it ran does not re-create it (D#265).
 try {
-  if (!existsSync(RUNTIME_DIR)) mkdirSync(RUNTIME_DIR, { recursive: true, mode: 0o700 });
-  else chmodSync(RUNTIME_DIR, 0o700);
+  if (!existsSync(RUNTIME_DIR)) {
+    if (!workerDirGone(DB_DIR)) mkdirSync(RUNTIME_DIR, { recursive: true, mode: 0o700 });
+  } else chmodSync(RUNTIME_DIR, 0o700);
 } catch {}
 
 // ─── Session ID Management ───────────────────────────────────────────────────
@@ -604,7 +608,7 @@ export function spawnBackground(bgEvent, ...extraArgs) {
     const child = spawn(process.execPath, args, {
       detached: true,
       stdio: 'ignore',
-      env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1' },
+      env: backgroundWorkerEnv(),
     });
     child.on('error', (err) => {
       debugCatch(err, 'spawnBackground');