claude-mem-lite 2.34.1 → 2.34.3

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "2.34.1",
+      "version": "2.34.3",
       "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": "2.34.1",
+  "version": "2.34.3",
   "description": "Lightweight persistent memory system for Claude Code — FTS5 search, episode batching, error-triggered recall",
   "author": {
     "name": "sdsrss"

package/mem-cli.mjs

@@ -742,8 +742,11 @@ function cmdTimeline(db, args) {
     return;
   }
 
-  const projectFilter = project ? 'AND project = ?' : '';
-  const baseParams = project ? [project] : [];
+  // Auto-scope to anchor's project when --project not explicitly given: users asking
+  // "what happened around #N" expect same-project context, not cross-project time-bleed.
+  const effectiveProject = project || anchorRow.project;
+  const projectFilter = effectiveProject ? 'AND project = ?' : '';
+  const baseParams = effectiveProject ? [effectiveProject] : [];
 
   // Before anchor
   const beforeRows = db.prepare(`
@@ -1721,17 +1724,26 @@ function cmdMaintain(db, args) {
     }
 
     if (ops.includes('dedup') && flags['merge-ids']) {
-      // Parse merge-ids: "keepId:removeId1:removeId2,keepId2:removeId3" format
+      // Parse merge-ids: "keepId:removeId1:removeId2,keepId2:removeId3" format.
+      // Surface malformed segments (non-numeric tokens, single-element pairs) instead of
+      // silently dropping them, so typos like "abc:def" don't hide behind "Merged 0".
       let totalMerged = 0;
+      const invalidSegments = [];
       const mergeStmt = db.prepare('UPDATE observations SET compressed_into = ? WHERE id = ? AND COALESCE(compressed_into, 0) = 0');
-      const groups = flags['merge-ids'].split(',').map(g => g.trim().split(':').map(Number).filter(n => !isNaN(n)));
-      for (const group of groups) {
-        if (group.length < 2) continue;
-        const [keepId, ...removeIds] = group;
+      const rawSegments = flags['merge-ids'].split(',').map(s => s.trim()).filter(Boolean);
+      for (const seg of rawSegments) {
+        const parts = seg.split(':').map(s => s.trim());
+        const nums = parts.map(p => Number(p));
+        const badToken = parts.length < 2 || nums.some(n => !Number.isFinite(n) || n <= 0);
+        if (badToken) { invalidSegments.push(seg); continue; }
+        const [keepId, ...removeIds] = nums;
         for (const removeId of removeIds) {
           totalMerged += mergeStmt.run(keepId, removeId).changes;
         }
       }
+      if (invalidSegments.length) {
+        results.push(`Warning: ignored ${invalidSegments.length} malformed --merge-ids segment(s): ${invalidSegments.join(', ')} (expected keepId:removeId[:removeId...] with positive integers)`);
+      }
       results.push(`Merged ${totalMerged} duplicate observations`);
     }
 
@@ -2070,6 +2082,16 @@ Commands:
     --project P         Filter by project
     --retain-days N     For purge_stale: keep last N days (default 30)
 
+  optimize              LLM-powered memory optimization (preview by default)
+    --run               Execute (default: preview gates)
+    --run-all           Execute bypassing gates
+    --task T            Comma-separated: re-enrich,normalize,cluster-merge,smart-compress
+    --max N             Max items per task (1-100, default 15)
+    --scope S           re-enrich scope: narrow (default) or wide
+
+  doctor                Environment diagnostics and benchmarks
+    --benchmark         Run perf benchmark and emit JSON
+
   fts-check <check|rebuild>  FTS5 index check or rebuild
 
   stats                 Show memory statistics

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "2.34.1",
+  "version": "2.34.3",
   "description": "Lightweight persistent memory system for Claude Code",
   "type": "module",
   "engines": {

package/scripts/pre-tool-recall.js

@@ -188,15 +188,19 @@ try {
     const lines = [];
     if (allRows.length > 0) {
       lines.push(`[mem] Lessons for ${fname}:`);
+      // R3-UX: raised from 120 → 240 after measuring 97% of lessons exceed 120 chars
+      // (p50=218, avg=247). Previous limit truncated the actionable "Fix:" tail in 80%
+      // of lessons containing it. 3 × 240 ≈ 180 tokens/Edit — negligible context cost.
+      const LESSON_MAX = 240;
       for (const r of allRows) {
         if (r.lesson_learned) {
-          const lesson = r.lesson_learned.length > 120
-            ? r.lesson_learned.slice(0, 117) + '...'
+          const lesson = r.lesson_learned.length > LESSON_MAX
+            ? r.lesson_learned.slice(0, LESSON_MAX - 3) + '...'
             : r.lesson_learned;
           lines.push(`  #${r.id} [${r.type}] ${lesson}`);
         } else {
-          const title = (r.title || '').length > 120
-            ? r.title.slice(0, 117) + '...'
+          const title = (r.title || '').length > LESSON_MAX
+            ? r.title.slice(0, LESSON_MAX - 3) + '...'
             : (r.title || '');
           lines.push(`  #${r.id} [${r.type}] ${title}`);
         }

package/scripts/user-prompt-search.js

@@ -16,18 +16,18 @@ const INJECTED_IDS_FILE = join(DB_DIR, 'runtime', `.claude-mem-injected-${inferP
 const MAX_RESULTS = 5;
 const LOOKBACK_MS = 60 * 86400000; // 60 days
 
-// T3 (v2.31): BM25 magnitude threshold. OBS_BM25 (in scoring-sql.mjs) returns the
-// raw bm25() value, which in SQLite FTS5 is always negative — lower = better match.
-// The `relevance` column multiplies that negative bm25 by positive decay / type /
-// importance weights, keeping the sign negative. "Stronger match" therefore means
-// larger magnitude, so we compare against `Math.abs(relevance)`.
+// T3 (v2.31): per-row BM25 magnitude floor. OBS_BM25 (in scoring-sql.mjs)
+// returns the raw bm25() value — negative, smaller = better. Multiplied by
+// decay × type-quality × (0.5+0.5·importance), sign stays negative. We
+// compare against Math.abs(relevance).
 //
-// Empirically (see Task 3 probe in docs/plans/2026-04-14-mem-v2.31-mvp.md):
-//   - OR-fallback single-stem match: |rel| ~ 3e-6
-//   - Multi-term AND match w/ importance+type boost: |rel| ~ 2e-5 .. 5e-5
-// The plan's hinted default (3.5) was a guess that's six orders of magnitude too
-// high for this codebase's scoring expression. 1e-5 suppresses OR-fallback noise
-// while preserving real hits. Env-overridable for tuning without a redeploy.
+// v2.34.3 note: the historic comment claimed |rel| falls in 3e-6..5e-5 range.
+// Re-measured against real data (see v2.34.3 CHANGELOG probe), actual scores
+// span ~6..133 across SIGNAL / META / NOISE prompts — the scoring expression
+// was revised in later versions and this constant was never retuned. 1e-5 now
+// acts as a NULL-rel guard, not a real noise filter. The primary noise gate
+// is TOP_REL_FLOOR below, which drops the whole FTS set when the best match
+// is weak.
 const BM25_MIN_SCORE = Number(process.env.CLAUDE_MEM_UPS_BM25_MIN || 1e-5);
 // Raw-character minimum length for the prompt. Additional to the CJK-weighted
 // `shouldSkip()` effective-length gate; catches medium-short Latin prompts that
@@ -41,6 +41,27 @@ const PROMPT_MIN_LENGTH = 15;
 const FOLLOWUP_PROMPT_MIN_LENGTH = 8;
 const FOLLOWUP_BM25_MIN_SCORE = Number(process.env.CLAUDE_MEM_UPS_BM25_MIN_FOLLOWUP || 5e-6);
 
+// v2.34.3: top-|rel| sanity gate. BM25_MIN_SCORE filters per-row; this floor
+// gates the entire FTS set. Noise prompts ("today's date", "current time")
+// produce OR-fallback leakage where every hit shares one tangential stem and
+// per-row filtering leaves all of them through. When the best match scores
+// below this floor, the whole FTS result set is dropped.
+//
+// Empirical distribution (v2.34.3 probe, 12 prompts):
+//   SIGNAL top-|rel|   60..133
+//   NOISE  top-|rel|   25..48
+//   WEAK-META          6.86..33
+// Default 50 sits in the clean 48→60 gap. Env override for project tuning.
+// Error-signature hits (sigRows) and file-recall (fileRows) bypass this gate —
+// both are precision passes with independent relevance signal.
+//
+// Note: no follow-up halving (unlike PROMPT_MIN_LENGTH / BM25_MIN_SCORE).
+// Those lower the length/per-row bar to let short context-dependent prompts
+// through, but the top-|rel| gap is an absolute distribution separator —
+// lowering it in follow-up mode re-admits the 37..48 noise band that the
+// gate exists to drop.
+const TOP_REL_FLOOR = Number(process.env.CLAUDE_MEM_UPS_TOP_MIN || 50);
+
 function isFollowUpSession() {
   try {
     const raw = readFileSync(INJECTED_IDS_FILE, 'utf8');
@@ -323,6 +344,16 @@ async function main() {
         typeof r.relevance === 'number' && Math.abs(r.relevance) >= bm25Floor
       );
 
+      // 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
+      // whole FTS set — noise prompts should produce no FTS injection.
+      // Query orders by `relevance` ASC; negative values → ftsRows[0] has the
+      // largest magnitude (strongest match) in this scoring expression.
+      if (ftsRows.length > 0 && Math.abs(ftsRows[0].relevance) < TOP_REL_FLOOR) {
+        ftsRows = [];
+      }
+
       // Merge: FTS results first, then file results, deduplicated
       const seen = new Set(ftsRows.map(r => r.id));
       rows = [...ftsRows];

package/server.mjs

@@ -789,8 +789,11 @@ server.registerTool(
       db.prepare('UPDATE observations SET access_count = COALESCE(access_count, 0) + 1, last_accessed_at = ? WHERE id = ?').run(Date.now(), anchorId);
     } catch { /* non-critical: FTS5 trigger may fail on corrupted index */ }
 
-    const projectFilter = args.project ? 'AND project = ?' : '';
-    const baseParams = args.project ? [args.project] : [];
+    // Auto-scope to anchor's project when caller didn't pass one: "timeline around #N"
+    // means same-project context by default; cross-project bleed breaks user mental model.
+    const effectiveProject = args.project || anchorRow.project;
+    const projectFilter = effectiveProject ? 'AND project = ?' : '';
+    const baseParams = effectiveProject ? [effectiveProject] : [];
 
     // Before anchor
     const beforeRows = db.prepare(`