claude-mem-lite 3.24.0 → 3.26.0

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "3.24.0",
+      "version": "3.26.0",
       "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": "3.24.0",
+  "version": "3.26.0",
   "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/hook.mjs

@@ -1424,7 +1424,10 @@ async function handleUserPrompt() {
         process.stdout.write(lines.join('\n') + '\n');
       }
       if (imperativePick) {
-        process.stdout.write(formatTaskImperative(imperativePick.lesson_learned, imperativePick.id) + '\n');
+        // Guard the write on a non-empty return — formatTaskImperative yields '' for a
+        // lesson that strips to empty (e.g. "."), which would otherwise emit a bare line.
+        const imperativeLine = formatTaskImperative(imperativePick.lesson_learned, imperativePick.id);
+        if (imperativeLine) process.stdout.write(imperativeLine + '\n');
       }
     } catch (e) { debugCatch(e, 'handleUserPrompt-memory'); }
   } finally {

package/mem-cli.mjs

@@ -2867,7 +2867,10 @@ async function cmdOptimize(db, args) {
   // Sibling-command flag footgun: optimize executes with --run (compress uses
   // --execute, maintain uses positional `execute`). A bare --execute previously fell
   // through to a silent preview, so the user thought they ran a mutation but didn't.
-  if (args.includes('--execute') && !run && !runAll) {
+  // Catch both `--execute` and the `--execute=value` form (the latter slipped a bare
+  // `args.includes('--execute')` and fell through to a silent preview — review minor).
+  const hasExecuteFlag = args.some(a => a === '--execute' || a.startsWith('--execute='));
+  if (hasExecuteFlag && !run && !runAll) {
     fail("[mem] optimize executes with --run, not --execute (--execute is compress's flag). Re-run: claude-mem-lite optimize --run");
     return;
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "3.24.0",
+  "version": "3.26.0",
   "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/scripts/pre-tool-recall.js

@@ -116,7 +116,7 @@ function inferProject() {
   const parent = basename(join(dir, '..'));
   const raw = (parent && parent !== '.' && parent !== '/')
     ? `${parent}--${base}` : base;
-  return raw.replace(/[^a-zA-Z0-9_.-]/g, '-').slice(0, 100) || 'unknown';
+  return raw.replace(/[^a-zA-Z0-9_.-]/g, '-').slice(0, 100);
 }
 
 function readCooldown(cooldownPath) {

package/scripts/user-prompt-search.js

@@ -172,6 +172,20 @@ function isFollowUpSession() {
 // catches the prompt via those channels rather than the identifier itself.
 const TECH_IDENTIFIER_RE = /\b(?:[a-z][a-z0-9]*_[a-z0-9_]+|[A-Z][A-Z0-9]*_[A-Z0-9_]+|[A-Z]{2,}[0-9][A-Z0-9_]*|[a-z]{2,}[A-Z][a-zA-Z0-9]+|[a-z]+(?:-[a-z]+){2,})\b/;
 
+// Reviewer #2 (v3.25.0): the kebab (≥3-seg) and camelCase arms above structurally
+// match a handful of ordinary English phrases / product names that are NOT code
+// identifiers — `up-to-date`, `state-of-the-art`, `macOS`, etc. Left unfiltered they
+// (a) widen the default-ON hasExplicitSignal gate on prose-only prompts and (b) let
+// such a token drive the opt-in identifier bypass. Stop-list them (lowercased). Real
+// 3-segment kebab identifiers (`pre-tool-use`, `user-prompt-search`) are deliberately
+// NOT here — only attested non-identifier prose.
+const IDENTIFIER_STOPWORDS = new Set([
+  'up-to-date', 'out-of-date', 'up-to-speed', 'out-of-the-box', 'state-of-the-art',
+  'end-to-end', 'off-by-one', 'easy-to-use', 'day-to-day', 'step-by-step',
+  'face-to-face', 'one-to-one', 'one-on-one', 'back-to-back', 'side-by-side',
+  'apples-to-apples', 'macos',
+]);
+
 // CJK presence channel (Important #2): bilingual users (project memory
 // `feedback_*` calls this out explicitly) ask CJK questions that may carry
 // genuine debug intent without containing an English identifier. CJK is
@@ -192,34 +206,37 @@ export function hasExplicitSignal(text, { errSig, files, intent } = {}) {
   if (errSig === undefined && extractErrorSignature(text)) return true;
   if (files === undefined && extractFiles(text).length > 0) return true;
   if (intent === undefined && detectIntent(text)) return true;
-  if (TECH_IDENTIFIER_RE.test(text)) return true;
+  if (extractTechIdentifiers(text).length > 0) return true;
   if (CJK_CHAR_RE.test(text) && computeEffectiveLen(text) >= CJK_MIN_EFFECTIVE_LEN) return true;
   return false;
 }
 
-// ─── Identifier-exact-match precision bypass (default OFF) ───────────────────
+// ─── Identifier-exact-match precision bypass (default ON since v3.26.0) ──────
 //
-// CLAUDE_MEM_UPS_IDENTIFIER_BYPASS=1 enables it. Rationale: the score-floors below
+// Set CLAUDE_MEM_UPS_IDENTIFIER_BYPASS=0 to disable. Rationale: the score-floors below
 // (OR_TOP_BM25_FLOOR / TOP_REL_FLOOR) drop the WHOLE FTS set when the top row's
 // magnitude is weak. But a rare code identifier (camelCase / snake_case / CONST_CASE
 // / kebab≥3) match has high *semantic* precision even at modest BM25 — a df=1 term
 // like `sanitizeFtsQuery` scores only ~23 raw yet is an unambiguous hit. So an obs
 // whose title/lesson EXACT-matches an identifier the prompt names is a precision pass
 // — the same independent-signal rationale that already exempts sigRows (error
-// signatures) and fileRows (file names) from these floors. When enabled, such rows are
-// restored after the floors run.
+// signatures) and fileRows (file names) from these floors. Such rows are restored
+// after the floors run.
 //
-// Default OFF — opt-in. benchmark/ups-ab.mjs measured it TRADEOFF on the real corpus
-// (+3 recall recoveries / +4 injections over 12 positives / 8 hard-negatives); the
-// "cost" was eager surfacing of on-identifier obs on non-recall-framed prompts, not
-// off-topic noise. Kept opt-in (below the NET-POSITIVE/NEUTRAL ship bar). See #8858.
-export const IDENTIFIER_BYPASS = process.env.CLAUDE_MEM_UPS_IDENTIFIER_BYPASS === '1';
+// Default ON — flipped in v3.26.0 after a corrected benchmark/ups-ab.mjs re-measurement
+// (#8858 two-bucket: TRUE off-topic FP isolated from on-topic eagerness). Over 12
+// positives / 7 true-precision hard-negatives: recall up, TRUE off-topic FP = 0 (stable
+// ×3 runs) → VERDICT NET-POSITIVE. The only behavior delta is on-topic eagerness (naming
+// an identifier surfaces its obs) — the highest-precision injection trigger there is. The
+// prose stop-list (IDENTIFIER_STOPWORDS) keeps the extractor off ordinary English.
+export const IDENTIFIER_BYPASS = process.env.CLAUDE_MEM_UPS_IDENTIFIER_BYPASS !== '0';
 const TECH_IDENTIFIER_RE_G = new RegExp(TECH_IDENTIFIER_RE.source, 'g');
 
 // All tech-identifier tokens in `text`, lowercased + de-duped (for case-insensitive
 // row matching). Empty array when none — callers treat that as "no bypass candidates".
 export function extractTechIdentifiers(text) {
-  return [...new Set((String(text || '').match(TECH_IDENTIFIER_RE_G) || []).map(s => s.toLowerCase()))];
+  return [...new Set((String(text || '').match(TECH_IDENTIFIER_RE_G) || []).map(s => s.toLowerCase()))]
+    .filter(s => !IDENTIFIER_STOPWORDS.has(s));
 }
 
 // True when the obs row's title or lesson contains any of `idsLower` as a standalone
@@ -600,8 +617,8 @@ async function main() {
     const signalPresent = hasExplicitSignal(promptText, {
       errSig, files: filesForGate, intent,
     });
-    // Identifier tokens the prompt names (for the precision bypass below). Empty
-    // unless CLAUDE_MEM_UPS_IDENTIFIER_BYPASS=1, so this is a no-op when disabled.
+    // Identifier tokens the prompt names (for the precision bypass below). Empty only
+    // when CLAUDE_MEM_UPS_IDENTIFIER_BYPASS=0 (bypass is default-on), then it is a no-op.
     const promptIdentifiers = IDENTIFIER_BYPASS ? extractTechIdentifiers(promptText) : [];
 
     if (intent?.useRecent) {
@@ -634,7 +651,7 @@ async function main() {
         typeof r.relevance === 'number' && Math.abs(r.relevance) >= bm25Floor
       );
 
-      // Identifier-exact-match precision bypass (default off — see IDENTIFIER_BYPASS).
+      // Identifier-exact-match precision bypass (default on — see IDENTIFIER_BYPASS).
       // Capture rows that exact-match a prompt identifier BEFORE the set-floors below;
       // they carry independent precision signal (sigRows/fileRows rationale) and are
       // restored after the floors so a low top-score can't drop a named-identifier hit.

package/search-scoring.mjs

@@ -151,12 +151,14 @@ export function markSuperseded(results) {
   // Persistent superseding belongs in mem_maintain/mem_compress write paths.
   for (const [, obsForFile] of fileMap) {
     if (obsForFile.length < 2) continue;
-    // Sort newest-first by recency. The production search pipeline supplies
-    // created_at_epoch (numeric) + created_at (ISO) — NOT `date`; the prior
-    // `b.date` sort silently no-op'd there (every key undefined → localeCompare 0),
-    // so "newest" degraded to relevance order and a current top-importance obs got
-    // a false [SUPERSEDED] tag (#8821). Prefer numeric epoch, fall back to the ISO
-    // string (lexically chronological), then legacy `date` for older callers/tests.
+    // Sort newest-first by recency. Every production obs row carries `date`
+    // (= created_at ISO, set by ftsRowToResult); the FTS/type paths also carry
+    // created_at_epoch, the vector/type-list paths carry date-only. So the prior
+    // `b.date` sort already ordered correctly in production — this is defensive
+    // hardening, not a prod no-op fix (a post-release review corrected the original
+    // #8821 "missing-`date`" diagnosis). Prefer the numeric epoch, fall back to the
+    // ISO created_at (lexically chronological), then legacy `date`, so the key holds
+    // for any caller regardless of which recency fields it supplies.
     obsForFile.sort((a, b) => {
       const ka = a.created_at_epoch ?? a.created_at ?? a.date ?? '';
       const kb = b.created_at_epoch ?? b.created_at ?? b.date ?? '';