claude-mem-lite 2.43.0 → 2.45.0

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "2.43.0",
+      "version": "2.45.0",
       "source": "./",
       "description": "Lightweight persistent memory system for Claude Code — FTS5 search, episode batching, error-triggered recall"
     }

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "2.43.0",
+  "version": "2.45.0",
   "description": "Lightweight persistent memory system for Claude Code — FTS5 search, episode batching, error-triggered recall",
   "author": {
     "name": "sdsrss"

package/hook-llm.mjs

@@ -431,15 +431,30 @@ export function buildImmediateObservation(episode) {
   }
 
   const ruleImportance = computeRuleImportance(episode);
-  // Low-signal degraded titles ("Error in...", "Modified...") should not inflate importance.
-  // Cap at 1 unless rule-based signals indicate genuine importance (error-in-test → 3, config → 2).
+  // Low-signal degraded titles ("Modified X", "Worked on X", "Reviewed N files")
+  // should not inflate importance. computeRuleImportance's file-name heuristics
+  // (schema.*, migration, auth.*, .env, .pem) fire on any matching file in the
+  // episode, so a 5-file review that incidentally reads one schema.js triggers
+  // imp=3 even though schema.js was one of 5 scanned — not the focus. Combined
+  // with a LOW_SIGNAL title (Haiku couldn't extract meaning), we can't justify
+  // imp=3; cap at 2 so rule says "notable" but not "critical".
+  //
+  // Production baseline (2026-04-23, projects--mem): 34/100 discovery/imp=3
+  // obs were LOW_SIGNAL titles; 7 change/imp=3 same. Prior cap `rule<=2 → 1`
+  // only fired when rule was weak, letting rule=3 leak through. New cap:
+  //   isReviewPattern → 2 (was Math.max(2, rule) → rule=3 leaked as 3)
+  //   isLowSignal & !review:
+  //     rule=3 → 2 (was 3)             — the fix
+  //     rule<=2 → 1 (unchanged)        — original cap preserved
   const LOW_SIGNAL = LOW_SIGNAL_TITLE;
   const isLowSignal = LOW_SIGNAL.test(title);
   let importance;
   if (isReviewPattern) {
-    importance = Math.max(2, ruleImportance);
-  } else if (isLowSignal && ruleImportance <= 2) {
-    importance = 1; // Degraded titles stay low unless rule signals critical (imp=3)
+    // Review titles are auto-generated from file count — can't distinguish
+    // "critical file was primary focus" from "one of N files read". Cap at 2.
+    importance = 2;
+  } else if (isLowSignal) {
+    importance = ruleImportance === 3 ? 2 : 1;
   } else {
     importance = ruleImportance;
   }

package/install.mjs

@@ -1205,7 +1205,7 @@ async function doctor() {
     if (r.drift) {
       const names = r.details.join(', ');
       const suffix = r.plainCount > r.details.length ? ` +${r.plainCount - r.details.length} more` : '';
-      warn(`Dev drift: ${r.plainCount} non-symlink file(s) in dev install: ${names}${suffix} (re-run: node install.mjs install --dev)`);
+      warn(`Dev drift: ${r.plainCount} non-symlink file(s) in dev install: ${names}${suffix} (re-run: node ${join(PROJECT_DIR, 'install.mjs')} install --dev)`);
       issues++;
     } else if (r.devMode) {
       ok(`Dev drift: clean (${r.symlinkCount} symlinks, 0 plain)`);

package/mem-cli.mjs

@@ -484,23 +484,26 @@ function cmdRecall(db, args) {
   const { positional, flags } = parseArgs(args);
   const file = positional.join(' ');
   if (!file) {
-    fail('[mem] Usage: mem recall <file>');
+    fail('[mem] Usage: mem recall <file> [--limit N] [--include-noise]');
     return;
   }
 
   const filename = basename(file);
   const rawLimit = flags.limit !== undefined ? parseInt(flags.limit, 10) : NaN;
   const limit = Number.isInteger(rawLimit) ? Math.max(1, rawLimit) : 10;
+  const includeNoise = flags['include-noise'] === true || flags['include-noise'] === 'true';
 
   // Search via observation_files junction table for indexed filename lookups
   const escaped = filename.replace(/%/g, '\\%').replace(/_/g, '\\_');
   const likePattern = `%${escaped}`;
+  const noiseClause = includeNoise ? '' : `AND ${notLowSignalTitleClause('o')}`;
   const rows = db.prepare(`
     SELECT DISTINCT o.id, o.type, o.title, o.lesson_learned, o.created_at, o.project
     FROM observations o
     JOIN observation_files of2 ON of2.obs_id = o.id
     WHERE COALESCE(o.compressed_into, 0) = 0
       AND (of2.filename = ? OR of2.filename LIKE ? ESCAPE '\\')
+      ${noiseClause}
     ORDER BY o.created_at_epoch DESC
     LIMIT ?
   `).all(filename, likePattern, limit);
@@ -1972,6 +1975,7 @@ Commands:
 
   recall <file>         Show observations related to a file
     --limit N           Max results (default 10)
+    --include-noise     Include hook-llm fallback titles ("Modified X", raw error logs)
 
   get <id1,id2,...>     Get full details by ID
     IDs accept search-output prefixes: #123 (obs), P#123 (prompt), S#123 (session).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "2.43.0",
+  "version": "2.45.0",
   "description": "Lightweight persistent memory system for Claude Code",
   "type": "module",
   "engines": {

package/scripts/prompt-search-utils.mjs

@@ -9,6 +9,19 @@ const CONFIRM_RE = /^(y(es)?|no?|ok|done|go|sure|lgtm|thanks?|ty|继续|确认|
 const SLASH_CMD_RE = /^\//;
 const PURE_OP_RE = /^(git\s+(commit|push|merge)|npm\s+(publish|deploy))\b/i;
 
+// v2.43.x: pure continuation directives — "keep going on what you were doing"
+// with no new topic. Long enough to evade CONFIRM_RE / length gate but
+// semantically empty for memory-recall purposes; injecting [mem] context
+// here reads like a turn boundary and can prematurely end the model's
+// in-flight tool chain. Conservative match: must be SOLELY the directive,
+// not directive + new instruction (those keep getting injection).
+const CONTINUATION_RE = /^(继续|接着|继续做|接着做|继续干|继续做下一步|接着做下一步|别停|不要停|next|continue|go\s*on|keep\s+going|carry\s+on|proceed|more(?:\s+please)?)\s*[?？!！。.，,]*\s*$/i;
+
+// v2.43.x: meta-pause questions — user is asking the model to reflect on
+// its own pause/stop, then continue. No new topic = no useful memory hit;
+// injection just adds reminder noise on top of an already-reflective turn.
+const META_PAUSE_RE = /(怎么停|为什么停|为何停|你怎么停|工作停下来|刚才停|why\s+(?:did\s+you\s+)?(?:stop|pause|halt))/i;
+
 /**
  * CJK-weighted effective length. CJK characters (CJK Unified Ideographs
  * main + extension A) carry ~3x the semantic token density of Latin
@@ -30,6 +43,8 @@ export function shouldSkip(text) {
   if (CONFIRM_RE.test(trimmed)) return true;
   if (SLASH_CMD_RE.test(trimmed)) return true;
   if (PURE_OP_RE.test(trimmed)) return true;
+  if (CONTINUATION_RE.test(trimmed)) return true;
+  if (META_PAUSE_RE.test(trimmed)) return true;
   return false;
 }
 

package/scripts/user-prompt-search.js

@@ -66,6 +66,34 @@ const FOLLOWUP_BM25_MIN_SCORE = Number(process.env.CLAUDE_MEM_UPS_BM25_MIN_FOLLO
 // gate exists to drop.
 const TOP_REL_FLOOR = Number(process.env.CLAUDE_MEM_UPS_TOP_MIN || 50);
 
+// v2.43.x: OR-fallback raw BM25 magnitude floor. The composite TOP_REL_FLOOR
+// above gates on `bm25 × importance × type_quality × decay × noise_penalty`.
+// For importance=3 bugfix obs, those multipliers compound to ~6×, so a modest
+// BM25 of -17..-22 can clear a composite floor of 50 via inflation alone.
+// When the FTS query relaxes to OR (AND returned 0), a single strongly-
+// matching stem on a big multi-topic prompt leaks through — observed
+// failure mode: broad Chinese prompts surfacing unrelated importance=3
+// bugfix obs whose concepts share exactly one stem with the prompt.
+//
+// Empirical OR-mode distribution (11-prompt probe, 2026-04-23):
+//   real signal      top-|bm25_raw| ≥ 41
+//   broad/meta noise top-|bm25_raw| ≤ 22
+//   below threshold  top-|bm25_raw| < 12
+// Default 30 sits in the clean 22→41 gap. AND mode bypasses this gate —
+// AND's all-stems-must-match constraint is already a precision signal,
+// and there are legitimate AND hits (GOOD-narrow probe: bm25_raw=19.3,
+// rel=81) that we must not drop.
+//
+// CLAUDE_MEM_UPS_TOP_MIN=0 disables this too: on small test corpora (1–2
+// seeded obs) absolute BM25 magnitudes collapse to near-zero (observed
+// |bm25|≈4e-6) because FTS5 IDF normalization needs a real document
+// distribution. The existing TOP_REL_FLOOR knob already encodes the
+// "seed-mode: kill absolute floors" semantic for integration tests, so
+// we piggy-back on it rather than introducing a second override env.
+const OR_TOP_BM25_FLOOR = TOP_REL_FLOOR === 0
+  ? 0
+  : Number(process.env.CLAUDE_MEM_UPS_OR_BM25_MIN || 30);
+
 function isFollowUpSession() {
   try {
     const raw = readFileSync(INJECTED_IDS_FILE, 'utf8');
@@ -77,9 +105,15 @@ function isFollowUpSession() {
 
 // ─── DB Query Functions ─────────────────────────────────────────────────────
 
+// Returns { rows, mode } where mode is 'AND' (initial pass), 'OR' (fallback
+// after AND returned 0), or null (no FTS query / sanitize rejected). Callers
+// use `mode` to apply OR-specific gates — see OR_TOP_BM25_FLOOR rationale.
+// Each row includes `bm25_raw` (pre-multiplier bm25 magnitude) alongside the
+// composite `relevance`, so callers can distinguish raw-match strength from
+// importance/type/decay inflation.
 function searchByFts(db, queryText, project, limit, typeFilter) {
   const ftsQuery = sanitizeFtsQuery(queryText);
-  if (!ftsQuery) return [];
+  if (!ftsQuery) return { rows: [], mode: null };
 
   const cutoff = Date.now() - LOOKBACK_MS;
 
@@ -92,6 +126,7 @@ function searchByFts(db, queryText, project, limit, typeFilter) {
   // docs/p0-injection-noise-baseline.txt.
   const sql = `
     SELECT o.id, o.type, o.title, o.lesson_learned,
+           ${OBS_BM25} as bm25_raw,
            ${OBS_BM25}
              * (1.0 + EXP(-0.693 * (? - o.created_at_epoch) / ${TYPE_DECAY_CASE}))
              * ${TYPE_QUALITY_CASE}
@@ -115,6 +150,7 @@ function searchByFts(db, queryText, project, limit, typeFilter) {
   params.push(limit);
 
   let rows = db.prepare(sql).all(...params);
+  let mode = 'AND';
 
   // OR fallback if AND query returned nothing
   if (rows.length === 0) {
@@ -122,10 +158,11 @@ function searchByFts(db, queryText, project, limit, typeFilter) {
     if (orQuery) {
       params[1] = orQuery;
       rows = db.prepare(sql).all(...params);
+      mode = 'OR';
     }
   }
 
-  return rows;
+  return { rows, mode };
 }
 
 function searchByFile(db, files, project, limit) {
@@ -256,7 +293,7 @@ const QUIET_HOOKS = process.env.MEM_QUIET_HOOKS === '1';
 function formatResults(rows) {
   if (!rows || rows.length === 0) return null;
 
-  const lines = ['[mem] Related memories:'];
+  const lines = ['[mem] FYI — Related memories (continue your task):'];
   for (const r of rows) {
     const icon = typeIcon(r.type);
     const title = truncate(r.title || '', 70);
@@ -272,7 +309,7 @@ function formatResults(rows) {
 // chars (slightly longer than obs titles because prompts carry more context).
 function formatPromptResults(rows) {
   if (!rows || rows.length === 0) return null;
-  const lines = ['[mem] Past similar questions:'];
+  const lines = ['[mem] FYI — Past similar questions (continue your task):'];
   for (const r of rows) {
     const text = truncate((r.prompt_text || '').replace(/\s+/g, ' '), 80);
     lines.push(`P#${r.id} 💬 ${text}`);
@@ -375,7 +412,7 @@ async function main() {
     // take priority slots in the merged output.
     const errSig = extractErrorSignature(promptText);
     const sigRows = errSig
-      ? searchByFts(db, errSig.signature, project, 2, 'bugfix').filter(r =>
+      ? searchByFts(db, errSig.signature, project, 2, 'bugfix').rows.filter(r =>
           typeof r.relevance === 'number' && Math.abs(r.relevance) >= bm25Floor
         )
       : [];
@@ -386,11 +423,13 @@ async function main() {
     } else {
       // FTS search: use the prompt as query, optionally type-filtered
       const files = extractFiles(promptText);
-      let ftsRows = searchByFts(db, promptText, project, intent?.limit || MAX_RESULTS, intent?.type || null);
+      let ftsResult = searchByFts(db, promptText, project, intent?.limit || MAX_RESULTS, intent?.type || null);
       // Fallback: if typed search returned nothing, retry without type filter
-      if (ftsRows.length === 0 && intent?.type) {
-        ftsRows = searchByFts(db, promptText, project, intent.limit || MAX_RESULTS, null);
+      if (ftsResult.rows.length === 0 && intent?.type) {
+        ftsResult = searchByFts(db, promptText, project, intent.limit || MAX_RESULTS, null);
       }
+      let ftsRows = ftsResult.rows;
+      const ftsMode = ftsResult.mode;
       const fileRows = files.length > 0 ? searchByFile(db, files, project, 2) : [];
 
       // T3 (v2.31): BM25 magnitude threshold — drop FTS hits whose relevance
@@ -403,6 +442,19 @@ async function main() {
         typeof r.relevance === 'number' && Math.abs(r.relevance) >= bm25Floor
       );
 
+      // v2.43.x: OR-mode raw-BM25 floor. In OR-fallback mode the composite
+      // TOP_REL_FLOOR below is inflated by importance × type_quality × decay
+      // multipliers — a weak single-stem hit on an importance=3 bugfix obs
+      // can reach composite rel=66 while raw |bm25|=19. Gate on raw bm25
+      // magnitude for OR mode only; AND mode's all-stems-match constraint
+      // is a precision signal and routinely produces legitimate AND hits
+      // below raw |bm25|=20 that we do not want to drop (see GOOD-narrow
+      // probe). Skip gate when OR_TOP_BM25_FLOOR is set to 0 (test hook).
+      if (ftsMode === 'OR' && OR_TOP_BM25_FLOOR > 0 && ftsRows.length > 0) {
+        const topBm25 = Math.abs(ftsRows[0].bm25_raw || 0);
+        if (topBm25 < OR_TOP_BM25_FLOOR) ftsRows = [];
+      }
+
       // v2.34.3: top-|rel| sanity gate. Per-row filtering above leaves noise
       // prompts intact when many rows share a weak stem (all in 25..48 range).
       // If the best remaining FTS match is below the top floor, drop the

package/server.mjs

@@ -2176,15 +2176,18 @@ server.registerTool(
   safeHandler(async (args) => {
     const filename = basename(args.file);
     const limit = args.limit ?? 10;
+    const includeNoise = args.include_noise === true;
 
     const escaped = filename.replace(/%/g, '\\%').replace(/_/g, '\\_');
     const likePattern = `%${escaped}`;
+    const noiseClause = includeNoise ? '' : `AND ${notLowSignalTitleClause('o')}`;
     const rows = db.prepare(`
       SELECT DISTINCT o.id, o.type, o.title, o.lesson_learned, o.created_at, o.project
       FROM observations o
       JOIN observation_files of2 ON of2.obs_id = o.id
       WHERE COALESCE(o.compressed_into, 0) = 0
         AND (of2.filename = ? OR of2.filename LIKE ? ESCAPE '\\')
+        ${noiseClause}
       ORDER BY o.created_at_epoch DESC
       LIMIT ?
     `).all(filename, likePattern, limit);

package/tool-schemas.mjs

@@ -211,6 +211,7 @@ export const memExportSchema = {
 export const memRecallSchema = {
   file: z.string().min(1).describe('File path or filename to recall observations for'),
   limit: coerceInt.pipe(z.number().int().min(1).max(50)).optional().describe('Max results (default 10)'),
+  include_noise: z.boolean().optional().describe('Include hook-llm fallback titles ("Modified X", "Worked on X", raw error logs) — hidden by default for parity with mem_search'),
 };
 
 export const memFtsCheckSchema = {