claude-mem-lite 3.13.0 → 3.14.0

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "3.13.0",
+      "version": "3.14.0",
       "source": "./",
       "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF 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": "3.13.0",
+  "version": "3.14.0",
   "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF 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/bash-utils.mjs

@@ -76,7 +76,14 @@ export function detectBashSignificance(input, response) {
   // Allow intervening global git options (`-C <path>`, `-c k=v`, `--no-pager`, …) between
   // `git` and the subcommand — `git -C /repo push` is the standard multi-repo/scripted form.
   const isGit = /\bgit\s+(?:(?:-[cC]\s+\S+|--?[\w-]+(?:=\S+)?)\s+)*(commit|merge|rebase|cherry-pick|push)\b/i.test(cmd);
-  const isDeploy = /\b(deploy|docker|kubectl|terraform)\b/i.test(cmd);
+  // Deploy + publish/release: the actual "ship". Package publish and GitHub
+  // release are rare, high-value events; without them a release session records
+  // the git push but not the publish that defines it. `npm run publish-*` is
+  // excluded (custom script, ambiguous) — only the direct publish verb counts.
+  const isDeploy = /\b(deploy|docker|kubectl|terraform)\b/i.test(cmd)
+    || /\b(?:npm|pnpm|yarn|bun|cargo)\s+publish\b/i.test(cmd)
+    || /\bgh\s+release\s+(?:create|edit|upload|delete)\b/i.test(cmd)
+    || /\btwine\s+upload\b/i.test(cmd);
   return {
     isError, isTest, isBuild, isGit, isDeploy,
     isSignificant: isError || isTest || isBuild || isGit || isDeploy,

package/hook-memory.mjs

@@ -131,7 +131,23 @@ function hasFilePaths(filesModified) {
  * @returns {object[]} Top memories (max 3) with {id, type, title, lesson_learned}
  */
 export function searchRelevantMemories(db, userPrompt, project, excludeIds = []) {
-  if (!db || !userPrompt || userPrompt.length < 5) return [];
+  // Min-length guard is English-centric: 5 chars ≈ one short English word. A CJK
+  // query is meaningful at 2 chars (状态/架构) and most real Chinese queries are
+  // 2-4 chars (状态管理, 召回率, 熔断降级) — the bare `.length < 5` silently
+  // rejected ALL of them, so a Chinese-primary user got zero memory injection.
+  // Apply the 5-char floor only to non-CJK queries; CJK needs ≥2.
+  if (!db || !userPrompt) return [];
+  const queryHasCjk = /[一-鿿㐀-䶿]/.test(userPrompt);
+  if (userPrompt.length < (queryHasCjk ? 2 : 5)) return [];
+  // CJK-DOMINANT (not merely CJK-containing) gates the OR-fallback bypass below.
+  // A substring test would let one incidental CJK char — an IME-leaked particle,
+  // a 中文 noun in an otherwise-English prompt — flip OR-fallback on and inject
+  // off-topic noise (a real precision regression for bilingual users). Require
+  // CJK chars to be at least as many as ASCII letters so only genuinely-Chinese
+  // queries (优化召回率) get the bigram-inflation rescue, not "fix the bug 啊".
+  const _cjkChars = (userPrompt.match(/[一-鿿㐀-䶿]/g) || []).length;
+  const _asciiLetters = (userPrompt.match(/[A-Za-z]/g) || []).length;
+  const queryIsCjkDominant = _cjkChars > 0 && _cjkChars >= _asciiLetters;
 
   // v2.41 metrics: record timing + candidate/filter/return counts per call.
   // Gated by CLAUDE_MEM_METRICS=1 — no-op when disabled (zero hot-path cost).
@@ -192,9 +208,16 @@ export function searchRelevantMemories(db, userPrompt, project, excludeIds = [])
     const queryTokenCount = ftsQuery.includes(' AND ')
       ? ftsQuery.split(' AND ').length
       : ftsQuery.split(/\s+/).filter(t => t && !t.startsWith('(') || !t.endsWith(')')).length;
+    // CJK-dominant queries bypass the token-count gate: a single CJK word becomes
+    // 2-N overlapping bigrams (优化召回率 → 优化/召回/回率), inflating
+    // queryTokenCount past the gate, so the AND-too-strict query never gets the OR
+    // rescue that CJK retrieval relies on. The CLI/hybrid path relaxes CJK to OR
+    // unconditionally; mirror that here. Noise is contained downstream (0.4x OR
+    // penalty + BM25 threshold + term-coverage filter). queryIsCjkDominant (not
+    // mere CJK presence) is the gate — see its definition at the function top.
     if (rows.length === 0) {
       const orQuery = relaxFtsQueryToOr(ftsQuery);
-      if (orQuery && queryTokenCount <= OR_FALLBACK_MAX_TOKENS) {
+      if (orQuery && (queryIsCjkDominant || queryTokenCount <= OR_FALLBACK_MAX_TOKENS)) {
         try { rows = selectStmt.all(orQuery, project, cutoff); usedOrFallback = true; } catch {}
       }
     }
@@ -225,7 +248,7 @@ export function searchRelevantMemories(db, userPrompt, project, excludeIds = [])
       crossRows = crossStmt.all(ftsQuery, project, cutoff);
       if (crossRows.length === 0) {
         const orQuery = relaxFtsQueryToOr(ftsQuery);
-        if (orQuery && queryTokenCount <= OR_FALLBACK_MAX_TOKENS) {
+        if (orQuery && (queryIsCjkDominant || queryTokenCount <= OR_FALLBACK_MAX_TOKENS)) {
           try { crossRows = crossStmt.all(orQuery, project, cutoff); crossUsedOr = true; } catch {}
         }
       }

package/hook.mjs

@@ -1188,8 +1188,13 @@ async function handleSessionStart() {
     // project that the `### Deferred Work` block now reads from the
     // deferred_work table (was: high-importance observations in v2.69.x).
     // Idempotent via marker file; subsequent SessionStarts are silent.
-    try { emitV270UpgradeBanner({ project, runtimeDir: RUNTIME_DIR }); }
-    catch (e) { debugCatch(e, 'session-start-v270-banner'); }
+    try {
+      // Gate on prior data: a brand-new install never had v2.69.x deferred-block
+      // semantics, so the migration notice is wrong noise (it fired for every
+      // fresh install since v2.70). Only genuine upgraders with observations see it.
+      const obsCount = db.prepare('SELECT COUNT(*) AS c FROM observations WHERE project = ?').get(project)?.c || 0;
+      emitV270UpgradeBanner({ project, runtimeDir: RUNTIME_DIR, hasPriorData: obsCount > 0 });
+    } catch (e) { debugCatch(e, 'session-start-v270-banner'); }
 
     // Pre-load TF-IDF vocabulary cache for this session (from DB, ~1ms)
     try { getVocabulary(db); } catch (e) { debugCatch(e, 'session-start-vocab'); }

package/lib/upgrade-banner.mjs

@@ -17,15 +17,26 @@ import { join } from 'path';
  * @param {object} args
  * @param {string} args.project Project name (used in banner + marker filename).
  * @param {string} args.runtimeDir RUNTIME_DIR (test override; production passes hook-shared.RUNTIME_DIR).
+ * @param {boolean} [args.hasPriorData=true] Whether the project has pre-existing
+ *   observations. A fresh install never had the v2.69.x deferred-block semantics,
+ *   so the migration notice (and its "pin to 2.69.x" advice, nonsensical 40+
+ *   versions on) is wrong noise — suppress it but still write the marker so it
+ *   stays suppressed once the new user starts saving memories. Defaults to true
+ *   for backward-compatibility with callers that don't probe.
  */
-export function emitV270UpgradeBanner({ project, runtimeDir }) {
+export function emitV270UpgradeBanner({ project, runtimeDir, hasPriorData = true }) {
   const marker = join(runtimeDir, `.deferred-block-migrated-${project}`);
   if (existsSync(marker)) return;
-  process.stderr.write(
-    `[mem] v2.70.0 upgrade notice (project "${project}"): Deferred Work block now ` +
-    `backed by deferred_work table. To keep an obs visible there, run ` +
-    `\`claude-mem-lite defer add "<title>" --priority 3\`. ` +
-    `Pin to 2.69.x to revert.\n`
-  );
+  // The banner only matters to someone who had high-importance observations
+  // under the old deferred-block semantics. No prior data → nothing migrated →
+  // suppress (but mark, so a brand-new user never sees it later either).
+  if (hasPriorData) {
+    process.stderr.write(
+      `[mem] v2.70.0 upgrade notice (project "${project}"): Deferred Work block now ` +
+      `backed by deferred_work table. To keep an obs visible there, run ` +
+      `\`claude-mem-lite defer add "<title>" --priority 3\`. ` +
+      `Pin to 2.69.x to revert.\n`
+    );
+  }
   try { writeFileSync(marker, String(Date.now())); } catch { /* best-effort marker */ }
 }

package/nlp.mjs

@@ -222,7 +222,19 @@ export function expandToken(token) {
 
 // ─── Stop Words ──────────────────────────────────────────────────────────────
 
-export const FTS_STOP_WORDS = new Set([...BASE_STOP_WORDS]);
+// Orphan stems left when apostrophe-splitting contractions ("I've"→"ve",
+// "wouldn't"→"wouldn"). Non-words only — they'd otherwise survive as required
+// AND terms and silently shrink recall. Ambiguous real words are intentionally
+// excluded so legitimate queries for them still work: don/won/haven/can, and
+// 're' ("re:"/regarding) — even though that leaves the they're/we're artifact,
+// dropping a real word is the worse failure.
+const CONTRACTION_FRAGMENTS = [
+  've', 'll',
+  'doesn', 'didn', 'isn', 'wasn', 'aren', 'weren',
+  'wouldn', 'couldn', 'shouldn', 'hasn', 'hadn', 'mustn', 'needn', 'mightn',
+];
+
+export const FTS_STOP_WORDS = new Set([...BASE_STOP_WORDS, ...CONTRACTION_FRAGMENTS]);
 
 // ─── FTS5 Query Sanitization ─────────────────────────────────────────────────
 
@@ -241,15 +253,30 @@ export function sanitizeFtsQuery(query) {
     // "never throws on MATCH" invariant. The metachar class below doesn't cover them.
     // eslint-disable-next-line no-control-regex -- intentional: stripping control chars IS the fix
     .replace(/[\x00-\x1f\x7f]/g, ' ')
+    // Apostrophe variants → space, matching FTS5 unicode61's own tokenization
+    // (it splits on apostrophe: "doesn't" is indexed as "doesn"+"t"). Without
+    // this, a possessive/contraction ("sister's", "What's") would phrase-quote
+    // to "sister s" (adjacency) and miss the bare-word mention ("my sister") in
+    // the doc, and contraction stems never reach the stopword filter. Straight
+    // ('), curly (' '), and modifier (ʼ) apostrophes all normalized.
+    .replace(/['‘’ʼ]/g, ' ')
     .replace(/[{}()[\]^~*:"\\]/g, ' ')
     .replace(/(^|\s)-/g, '$1')
     .trim();
   if (!cleaned) return null;
-  let tokens = cleaned.split(/\s+/).filter(t =>
-    t && !/^-+$/.test(t) && !FTS5_KEYWORDS.has(t.toUpperCase()) && !/^NEAR(\/\d*)?$/i.test(t)
-    // Skip single ASCII-letter tokens — too noisy for FTS5 (CJK single chars handled separately below)
-    && !(t.length === 1 && /^[a-zA-Z]$/.test(t))
-  );
+  let tokens = cleaned.split(/\s+/)
+    // Trim leading/trailing sentence punctuation (. , ? ! ; …) from each token.
+    // Natural-language queries end in "?" and clauses in ",": left on, the final
+    // (most salient) token rides as "bug?"/"month," which (a) misses the synonym
+    // map → no OR-expansion, and (b) gets phrase-quoted as a literal by
+    // expandToken. Edges only — internal dots/hyphens (cli.mjs, gardening-related)
+    // are preserved so filenames/compounds still phrase-match.
+    .map(t => t.replace(/^[.,;:!?]+|[.,;:!?]+$/g, ''))
+    .filter(t =>
+      t && !/^-+$/.test(t) && !FTS5_KEYWORDS.has(t.toUpperCase()) && !/^NEAR(\/\d*)?$/i.test(t)
+      // Skip single ASCII-letter tokens — too noisy for FTS5 (CJK single chars handled separately below)
+      && !(t.length === 1 && /^[a-zA-Z]$/.test(t))
+    );
   // Filter stop words (but keep all if filtering would empty the query)
   const filtered = tokens.filter(t => !FTS_STOP_WORDS.has(t.toLowerCase()));
   if (filtered.length > 0) tokens = filtered;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "3.13.0",
+  "version": "3.14.0",
   "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF 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).",
   "type": "module",
   "packageManager": "npm@10.9.2",

package/tier.mjs

@@ -110,7 +110,12 @@ export function tierSqlParams(ctx) {
  * @returns {string}
  */
 export function relativeTime(epoch, now) {
-  const diff = now - epoch;
+  // Clamp negative diffs: a future / clock-skewed epoch must not render as a
+  // negative duration ("-7200s ago"). The first branch below (diff < 60000)
+  // would otherwise fire on any negative diff and print Math.floor(diff/1000),
+  // i.e. "-7200s ago" for a 2h-future timestamp. (cli/common.mjs's sibling
+  // relativeTime handles this via an early `diff < 0` → "just now".)
+  const diff = Math.max(0, now - epoch);
   if (diff < 60000) return `${Math.floor(diff / 1000)}s ago`;
   if (diff < 3600000) return `${Math.floor(diff / 60000)}min ago`;
   if (diff < 86400000) return `${Math.floor(diff / 3600000)}h ago`;