claude-mem-lite 3.40.0 → 3.41.0

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "3.40.0",
+      "version": "3.41.0",
       "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": "3.40.0",
+  "version": "3.41.0",
   "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-context.mjs

@@ -305,6 +305,7 @@ export function buildSessionContextLines(db, project, now = new Date(), currentC
       FROM observations
       WHERE COALESCE(compressed_into, 0) = 0
         AND superseded_at IS NULL
+        AND ${notLowSignalTitleClause('')}
         AND (
           (created_at_epoch > ? AND importance >= 1)
           OR (created_at_epoch > ? AND importance >= 2)

package/hook-handoff.mjs

@@ -271,13 +271,28 @@ export function detectContinuationIntent(db, promptText, project, currentCcSessi
   try {
     const currentSha = gitStateModule.readGitState({ cwd: process.cwd() }).headSha;
     if (currentSha) {
-      const anchor = db.prepare(`
-        SELECT created_at_epoch FROM session_handoffs
-        WHERE project = ? AND git_sha_at_handoff = ?
-        ORDER BY created_at_epoch DESC LIMIT 1
-      `).get(project, currentSha);
+      // Scope like Stage 2: an 'exit' anchor is cross-session (resume after /exit), but a 'clear'
+      // anchor is same-session only — else a parallel same-project session at the same commit
+      // would hijack (and then delete) another session's clear handoff.
+      const anchor = currentCcSessionId
+        ? db.prepare(`
+            SELECT created_at_epoch, match_keywords FROM session_handoffs
+            WHERE project = ? AND git_sha_at_handoff = ? AND (type = 'exit' OR session_id = ?)
+            ORDER BY created_at_epoch DESC LIMIT 1
+          `).get(project, currentSha, currentCcSessionId)
+        : db.prepare(`
+            SELECT created_at_epoch, match_keywords FROM session_handoffs
+            WHERE project = ? AND git_sha_at_handoff = ?
+            ORDER BY created_at_epoch DESC LIMIT 1
+          `).get(project, currentSha);
       if (anchor && (Date.now() - anchor.created_at_epoch <= HANDOFF_ANCHOR_MAX_AGE)) {
-        return true;
+        // Unmoved HEAD is a strong resume signal, but must not hijack a NEW task typed at the
+        // same commit: gate long prompts on keyword overlap (mirror Stage 0). Short prompts
+        // (resume nudges) auto-continue.
+        if (promptText.length < 40) return true;
+        const hTokens = anchor.match_keywords ? new Set(tokenizeHandoff(anchor.match_keywords)) : null;
+        if (!hTokens || tokenizeHandoff(promptText).some(t => hTokens.has(t))) return true;
+        // long prompt with zero keyword overlap → fall through to Stage 0/1/2
       }
     }
   } catch { /* git/DB failure must not break the rest of the pipeline */ }

package/hook-llm.mjs

@@ -11,7 +11,7 @@ import {
 } from './utils.mjs';
 import { acquireLLMSlot, releaseLLMSlot } from './hook-semaphore.mjs';
 import { scrubRecord } from './lib/scrub-record.mjs';
-import { getVocabulary, computeVector } from './tfidf.mjs';
+import { getVocabulary, computeVector, vecTextForRow } from './tfidf.mjs';
 import { insertObservationRow, insertObservationFiles, insertObservationVector } from './lib/observation-write.mjs';
 import { DEDUP_JACCARD_THRESHOLD, AUTO_MERGE_THRESHOLD } from './lib/dedup-constants.mjs';
 import {
@@ -124,10 +124,12 @@ function buildFtsTextField(obs) {
 // weight) and search_aliases (finding #8: previously omitted, so even with vectors
 // enabled the paraphrase-bridge alias terms were invisible to cosine similarity).
 export function buildVecText(obs) {
-  const conceptsText = Array.isArray(obs.concepts) ? obs.concepts.join(' ') : (obs.concepts || '');
-  const aliasesText = obs.searchAliases || '';
-  return [obs.title || '', obs.narrative || '', conceptsText, obs.lessonLearned || '', aliasesText]
-    .filter(Boolean).join(' ');
+  // Single source (V-F1): map the camelCase obs onto vecTextForRow's row shape so save and
+  // every rebuild path encode the identical field set (title/narrative/concepts/lesson/aliases).
+  return vecTextForRow({
+    title: obs.title, narrative: obs.narrative, concepts: obs.concepts,
+    lesson_learned: obs.lessonLearned, search_aliases: obs.searchAliases,
+  });
 }
 
 /**
@@ -996,7 +998,8 @@ ${actionList}`;
           search_aliases: obs.searchAliases || null,
         });
         db.prepare(`
-          UPDATE observations SET type=?, title=?, subtitle=?, narrative=?, concepts=?, facts=?,
+          UPDATE observations SET type=?, title=?, subtitle=?,
+            narrative=COALESCE(NULLIF(?, ''), narrative), concepts=?, facts=?,
             text=?, importance=?, files_read=?, minhash_sig=?, lesson_learned=?, search_aliases=?
           WHERE id = ?
         `).run(
@@ -1018,7 +1021,7 @@ ${actionList}`;
         try {
           const vocab = getVocabulary(db);
           if (vocab) {
-            const vecText = [obs.title || '', obs.narrative || '', conceptsText].filter(Boolean).join(' ');
+            const vecText = vecTextForRow({ title: obs.title, narrative: obs.narrative, concepts: conceptsText, lesson_learned: safe.lesson_learned, search_aliases: safe.search_aliases });
             const vec = computeVector(vecText, vocab);
             if (vec) {
               db.prepare('INSERT OR REPLACE INTO observation_vectors (observation_id, vector, vocab_version, created_at_epoch) VALUES (?, ?, ?, ?)')

package/hook-optimize.mjs

@@ -12,7 +12,7 @@ import {
 import { callModelJSON } from './haiku-client.mjs';
 import { acquireLLMSlot, releaseLLMSlot } from './hook-semaphore.mjs';
 import { scrubRecord } from './lib/scrub-record.mjs';
-import { getVocabulary, computeVector, cosineSimilarity } from './tfidf.mjs';
+import { getVocabulary, computeVector, cosineSimilarity, vecTextForRow } from './tfidf.mjs';
 import { MERGE_JACCARD_LOW, AUTO_MERGE_THRESHOLD } from './lib/dedup-constants.mjs';
 import { DB_DIR } from './schema.mjs';
 
@@ -25,9 +25,14 @@ export function distributeBudget(total = 15) {
   const reenrich = Math.max(1, Math.floor(total * 0.4));
   const clusterMerge = Math.max(1, Math.floor(total * 0.3));
   const smartCompress = Math.max(1, total - reenrich - normalize - clusterMerge);
-  // Clamp: if total is too small for 4 tasks, cap each so sum ≤ total
+  // Clamp: if total is too small for all 4 tasks, allocate 1 each by priority until `total`
+  // is exhausted so the returned sum is ≤ total (the old fallback returned {1,1,1,1}=4 for
+  // total≤3, over-running `optimize --max N`).
   if (reenrich + normalize + clusterMerge + smartCompress > total) {
-    return { reenrich: Math.max(1, total - 3), normalize: 1, clusterMerge: 1, smartCompress: 1 };
+    const alloc = { reenrich: 0, clusterMerge: 0, smartCompress: 0, normalize: 0 };
+    const order = ['reenrich', 'clusterMerge', 'smartCompress', 'normalize'];
+    for (let i = 0; i < total && i < order.length; i++) alloc[order[i]] = 1;
+    return alloc;
   }
   return { reenrich, normalize, clusterMerge, smartCompress };
 }
@@ -39,11 +44,16 @@ export function distributeBudget(total = 15) {
  * Exported for testing; also kept as the single source of vector-rebuild logic
  * for the optimize / re-enrich path to avoid drift with the hook-llm write path.
  */
-export function rebuildVector(db, obsId, textParts) {
+export function rebuildVector(db, obsId, textPartsOrRow) {
   try {
     const vocab = getVocabulary(db);
     if (!vocab) return;
-    const vec = computeVector(textParts.filter(Boolean).join(' '), vocab);
+    // Accept a legacy [parts] array OR an observation row (preferred — single-source field set
+    // incl. lesson_learned/search_aliases via vecTextForRow, so rebuilds match the save path).
+    const text = Array.isArray(textPartsOrRow)
+      ? textPartsOrRow.filter(Boolean).join(' ')
+      : vecTextForRow(textPartsOrRow);
+    const vec = computeVector(text, vocab);
     if (vec) {
       // Bug #1 fix: column is `created_at_epoch`, not `computed_at`. Every other
       // INSERT callsite (server.mjs, hook-llm.mjs, mem-cli.mjs) uses the correct
@@ -87,7 +97,7 @@ export function findReenrichCandidates(db, limit = 10, { scope = 'narrow', proje
     // deliberately NOT gated on optimized_at, so a lesson-less row can still be
     // picked up by wide scope for lesson enrichment afterward.
     const stmt = db.prepare(`
-      SELECT id, title, narrative, type, subtitle, concepts, facts, text, search_aliases, project
+      SELECT id, title, narrative, type, subtitle, concepts, facts, text, search_aliases, importance, project
       FROM observations
       WHERE COALESCE(compressed_into, 0) = 0
         AND superseded_at IS NULL
@@ -102,7 +112,7 @@ export function findReenrichCandidates(db, limit = 10, { scope = 'narrow', proje
   }
   if (scope === 'wide') {
     const stmt = db.prepare(`
-      SELECT id, title, narrative, type, subtitle, concepts, facts, search_aliases, project
+      SELECT id, title, narrative, type, subtitle, concepts, facts, search_aliases, importance, project
       FROM observations
       WHERE COALESCE(compressed_into, 0) = 0
         AND superseded_at IS NULL
@@ -120,7 +130,7 @@ export function findReenrichCandidates(db, limit = 10, { scope = 'narrow', proje
     return project ? stmt.all(project, limit) : stmt.all(limit);
   }
   const stmt = db.prepare(`
-    SELECT id, title, narrative, type, subtitle, project
+    SELECT id, title, narrative, type, subtitle, importance, project
     FROM observations
     WHERE COALESCE(compressed_into, 0) = 0
       AND (concepts IS NULL OR concepts = '')
@@ -205,7 +215,10 @@ search_aliases: 2-6 alternative search terms (include CJK if applicable).`;
         continue;
       }
 
-      const type = validTypes.has(parsed.type) ? parsed.type : cand.type || 'change';
+      // 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.
+      let type = validTypes.has(parsed.type) ? parsed.type : (cand.type || 'change');
+      if (type === 'change' && cand.type && cand.type !== 'change') type = cand.type;
       const concepts = Array.isArray(parsed.concepts) ? parsed.concepts.slice(0, 10) : [];
       const facts = Array.isArray(parsed.facts) ? parsed.facts.slice(0, 10) : [];
       // Preserve-on-empty: wide-scope candidates can already carry concepts/facts/aliases
@@ -227,7 +240,10 @@ search_aliases: 2-6 alternative search terms (include CJK if applicable).`;
         ? parsed.search_aliases.slice(0, 6).join(' ') : (cand.search_aliases || null);
       const title = truncate(scrubSecrets(parsed.title || ''), 120);
       const narrative = truncate(scrubSecrets(parsed.narrative || cand.narrative || ''), 500);
-      const importance = clampImportance(parsed.importance);
+      // 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.
+      const importance = Math.max(clampImportance(parsed.importance), cand.importance || 1);
 
       const bigramText = cjkBigrams((title || '') + ' ' + (narrative || ''));
       const textField = [conceptsText, factsText, searchAliases || '', bigramText].filter(Boolean).join(' ');
@@ -250,7 +266,7 @@ search_aliases: 2-6 alternative search terms (include CJK if applicable).`;
       `).run(type, safe.title, safe.narrative, safe.concepts, safe.facts, safe.text,
         importance, safe.lesson_learned, safe.search_aliases, minhashSig, Date.now(), cand.id);
 
-      rebuildVector(db, cand.id, [title, narrative, conceptsText]);
+      rebuildVector(db, cand.id, { title, narrative, concepts: conceptsText, lesson_learned: safe.lesson_learned, search_aliases: safe.search_aliases });
 
       processed++;
     } catch (e) {
@@ -362,7 +378,7 @@ export function applyNormalization(db, groups, { project = null } = {}) {
   // contamination the --project flag was added to prevent. NULL → all projects (legacy
   // unscoped run), matching the search-engine `(? IS NULL OR project = ?)` idiom.
   const rows = db.prepare(`
-    SELECT id, concepts, search_aliases FROM observations
+    SELECT id, title, narrative, concepts, search_aliases, lesson_learned FROM observations
     WHERE COALESCE(compressed_into, 0) = 0
       AND concepts IS NOT NULL AND concepts != ''
       AND (? IS NULL OR project = ?)
@@ -395,6 +411,9 @@ export function applyNormalization(db, groups, { project = null } = {}) {
         search_aliases: newAliases,
       });
       updateStmt.run(safe.concepts, safe.search_aliases, Date.now(), row.id);
+      // V-F3: normalize mutated concepts + search_aliases (both vector fields) — rebuild the
+      // vector so it reflects the canonicalized terms (no-op when the vector arm is disabled).
+      rebuildVector(db, row.id, { title: row.title, narrative: row.narrative, concepts: safe.concepts, search_aliases: safe.search_aliases, lesson_learned: row.lesson_learned });
       updated++;
     }
   }
@@ -433,7 +452,7 @@ export function findMergeCandidates(db, maxClusters = 5, { project } = {}) {
   const cutoff = Date.now() - MERGE_TIME_WINDOW_MS;
   const projectClause = project ? 'AND project = ?' : '';
   const stmt = db.prepare(`
-    SELECT id, title, narrative, project, type, access_count, importance, created_at_epoch, minhash_sig, lesson_learned
+    SELECT id, title, narrative, project, type, access_count, importance, created_at_epoch, minhash_sig, lesson_learned, concepts, facts
     FROM observations
     WHERE COALESCE(compressed_into, 0) = 0
       AND superseded_at IS NULL
@@ -520,12 +539,15 @@ Return ONLY valid JSON:
 
     const concepts = Array.isArray(parsed.merged_concepts) ? parsed.merged_concepts.slice(0, 10) : [];
     const facts = Array.isArray(parsed.merged_facts) ? parsed.merged_facts.slice(0, 10) : [];
-    const conceptsText = concepts.join(' ');
-    const factsText = facts.join(' ');
+    // Preserve-on-empty (mirror the merged_lesson guard below + the re-enrich path): the merge
+    // overwrites the keeper in place, so a partial LLM response that omits these must fall back to
+    // the keeper's own values, not blank its live concepts/facts (findMergeCandidates now selects them).
+    const conceptsText = concepts.length ? concepts.join(' ') : (keeper.concepts || '');
+    const factsText = facts.length ? facts.join(' ') : (keeper.facts || '');
     // Scrub BEFORE truncate (see re-enrich note): keep the boundary cut on
     // already-scrubbed text so a straddling secret can't leak a sub-floor head.
     const title = truncate(scrubSecrets(parsed.merged_title || ''), 120);
-    const narrative = truncate(scrubSecrets(parsed.merged_narrative || ''), 800);
+    const narrative = truncate(scrubSecrets(parsed.merged_narrative || keeper.narrative || ''), 800);
     // Preserve-on-empty. The merge overwrites the keeper in place and hides every non-keeper
     // member (compressed_into=keeper.id), so if the LLM returns merged_lesson:null (the prompt
     // at :429 explicitly permits it) every cluster lesson would leave all live surfaces at once
@@ -589,7 +611,7 @@ Return ONLY valid JSON:
         .run(keeper.id, ...otherIds);
     })();
 
-    rebuildVector(db, keeper.id, [title, narrative, conceptsText]);
+    rebuildVector(db, keeper.id, { title, narrative, concepts: conceptsText, lesson_learned: lessonLearned, search_aliases: keeper.search_aliases });
 
     debugLog('DEBUG', 'llm-optimize', `merged ${cluster.length} observations into #${keeper.id}`);
     return { merged: true, keeperId: keeper.id, mergedCount: others.length };
@@ -793,7 +815,7 @@ JSON: {"title":"descriptive summary ≤120 chars","narrative":"comprehensive sum
       return sId;
     })();
 
-    rebuildVector(db, summaryId, [title, narrative, conceptsText]);
+    rebuildVector(db, summaryId, { title, narrative, concepts: conceptsText });
 
     debugLog('DEBUG', 'llm-optimize', `smart-compressed ${observations.length} observations into #${summaryId}`);
     return { compressed: true, summaryId, count: observations.length };

package/hook-shared.mjs

@@ -98,7 +98,9 @@ export function sweepOrphanEpisodeFiles(runtimeDir, { ageMs = ORPHAN_EPISODE_AGE
   const readsCutoff = now - readsAgeMs;
   let count = 0;
   for (const f of entries) {
-    const isEpisode = f.startsWith('ep-flush-') || f.startsWith('pending-');
+    // `.claim-` = handleStop's lock-contended fallback claim file (ep-<project>.json.claim-<pid>-<ts>),
+    // which leaks if the process dies between rename and unlink; sweep it on the same 1h cutoff.
+    const isEpisode = f.startsWith('ep-flush-') || f.startsWith('pending-') || f.includes('.claim-');
     const isReads = f.startsWith('reads-') && f.endsWith('.txt');
     if (!isEpisode && !isReads) continue;
     const full = join(runtimeDir, f);

package/hook.mjs

@@ -484,6 +484,10 @@ async function handleStop() {
           for (const sub of planEpisodeFlush(episode)) {
             if (!sub.sessionId) sub.sessionId = sessionId;
             if (!sub.project) sub.project = project;
+            // Per-sub significance gate — parity with flushEpisodeGroup. The whole-episode check
+            // above can pass while an interleaved concurrent-session sub is pure noise (e.g. a lone
+            // Read); without this the fallback persists that noise sub AND spawns an LLM worker for it.
+            if (!episodeHasSignificantContent(sub)) continue;
             try {
               const obs = buildImmediateObservation(sub);
               const id = saveObservation(obs, sub.project, sub.sessionId);
@@ -1412,7 +1416,14 @@ async function handleUserPrompt() {
     );
 
     // Cross-session handoff injection (first 3 prompts window, before semantic memory).
-    if (promptNumber <= 3) {
+    // prompt_counter is project-scoped (shared across concurrent same-project CC sessions), so a
+    // parallel session would start past the window and never get its handoff injected. Count THIS
+    // cc_session's own prompts instead (the current one is already inserted above); legacy null cc
+    // id falls back to the shared counter.
+    const windowPos = ccSessionId
+      ? (db.prepare('SELECT COUNT(*) c FROM user_prompts WHERE cc_session_id = ?').get(ccSessionId)?.c || promptNumber)
+      : promptNumber;
+    if (windowPos <= 3) {
       try {
         if (detectContinuationIntent(db, promptText, project, ccSessionId)) {
           const picked = pickHandoffToInject(db, project, ccSessionId);

package/install.mjs

@@ -1688,12 +1688,18 @@ async function doctor() {
 
 // ─── Settings helpers ───────────────────────────────────────────────────────
 
-function isMemHook(cfg) {
+// Identify a settings hook as one of OURS (to replace on install / strip on uninstall).
+// Must be tight: the old `hook.mjs` + event-word test matched a user's OWN generic hook
+// (`node ~/.config/hook.mjs session-start`) and install/uninstall silently deleted it.
+// The launcher marker (`hook-launcher.mjs`, which every Node hook routes through since v2.84)
+// replaces that clause; the product-name substring (our install-dir / legacy-direct hooks)
+// and the bash prefilter round out the real markers.
+export function isMemHook(cfg) {
   if (!cfg.hooks) return false;
   return cfg.hooks.some(h => {
     const cmd = h.command || '';
     return cmd.includes('claude-mem-lite') ||
-      (cmd.includes('hook.mjs') && /\b(session-start|stop|user-prompt|pre-tool-use)\b/.test(cmd)) ||
+      cmd.includes('hook-launcher.mjs') ||
       cmd.includes('scripts/post-tool-use.sh');
   });
 }

package/lib/maintain-core.mjs

@@ -13,7 +13,7 @@
 //   { projectFilter: 'AND project = ?' | '', baseParams: [project?] , staleAge, opCap }
 
 import { COMPRESSED_PENDING_PURGE, computeMinHash, estimateJaccardFromMinHash, jaccardSimilarity } from '../utils.mjs';
-import { rebuildVocabulary, computeVector, _resetVocabCache, vectorsEnabled } from '../tfidf.mjs';
+import { rebuildVocabulary, computeVector, _resetVocabCache, vectorsEnabled, vecTextForRow } from '../tfidf.mjs';
 import { DEDUP_JACCARD_THRESHOLD, MINHASH_PRE_THRESHOLD as MINHASH_PRE_THRESHOLD_SRC, FUZZY_DEDUP_THRESHOLD, FUZZY_BODY_THRESHOLD, MINHASH_PREFILTER } from './dedup-constants.mjs';
 
 export const STALE_AGE_MS = 30 * 86400000;
@@ -428,7 +428,7 @@ export function rebuildVectors(db) {
   const vocab = rebuildVocabulary(db);
   if (!vocab) return { ok: false, reason: 'no observations to build vocabulary from' };
   const allObs = db.prepare(`
-    SELECT id, title, narrative, concepts FROM observations
+    SELECT id, title, narrative, concepts, lesson_learned, search_aliases FROM observations
     WHERE COALESCE(compressed_into, 0) = 0 AND superseded_at IS NULL
   `).all();
   let updated = 0;
@@ -437,8 +437,7 @@ export function rebuildVectors(db) {
   db.transaction(() => {
     db.prepare('DELETE FROM observation_vectors').run();
     for (const obs of allObs) {
-      const text = [obs.title || '', obs.narrative || '', obs.concepts || ''].filter(Boolean).join(' ');
-      const vec = computeVector(text, vocab);
+      const vec = computeVector(vecTextForRow(obs), vocab);
       if (vec) {
         insertStmt.run(obs.id, Buffer.from(vec.buffer), vocab.version, now);
         updated++;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "3.40.0",
+  "version": "3.41.0",
   "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/setup.sh

@@ -39,20 +39,18 @@ fi
 mkdir -p "$DATA_DIR"
 log_ok "Data directory: $DATA_DIR"
 
-# 3. Migrate from old ~/.claude-mem/ if needed
+# 3. Legacy ~/.claude-mem/ DB is schema-v16 (no memory_session_id) with no migration bridge to
+#    the current schema — activating it FATALs on first launch ("no such column: memory_session_id")
+#    and the "! -f claude-mem-lite.db" guard would re-copy it every time the user deletes the broken
+#    DB (recovery loop). Mirror install.mjs migrateLegacyClaudeMemData: back it up (don't activate)
+#    and let a fresh DB be created. Source ~/.claude-mem/ is left intact.
 OLD_DIR="$HOME/.claude-mem"
 if [[ -f "$OLD_DIR/claude-mem.db" && ! -f "$DATA_DIR/claude-mem-lite.db" && ! -f "$DATA_DIR/claude-mem.db" ]]; then
-  log_info "Migrating data from ~/.claude-mem/ → ~/.claude-mem-lite/..."
-  if cp "$OLD_DIR/claude-mem.db" "$DATA_DIR/claude-mem-lite.db" 2>/dev/null; then
-    # Main DB copied successfully, WAL/SHM are optional
-    cp "$OLD_DIR/claude-mem.db-wal" "$DATA_DIR/claude-mem-lite.db-wal" 2>/dev/null || true
-    cp "$OLD_DIR/claude-mem.db-shm" "$DATA_DIR/claude-mem-lite.db-shm" 2>/dev/null || true
-    if [[ -d "$OLD_DIR/runtime" && ! -d "$DATA_DIR/runtime" ]]; then
-      cp -r "$OLD_DIR/runtime" "$DATA_DIR/runtime" 2>/dev/null || true
-    fi
-    log_ok "Data migrated (old ~/.claude-mem/ preserved)"
+  BACKUP="$DATA_DIR/claude-mem-lite.db.legacy-backup-$(date +%s)"
+  if cp "$OLD_DIR/claude-mem.db" "$BACKUP" 2>/dev/null; then
+    log_info "Legacy ~/.claude-mem/ DB is schema-incompatible; backed up to $(basename "$BACKUP") (a fresh DB will be created). Old ~/.claude-mem/ preserved."
   else
-    log_warn "Migration failed — using fresh database"
+    log_warn "Legacy DB backup failed — a fresh database will be created"
   fi
 fi
 

package/tfidf.mjs

@@ -207,6 +207,19 @@ let _vocabCache = null;
 /** Reset vocabulary cache (for testing). */
 export function _resetVocabCache() { _vocabCache = null; }
 
+/**
+ * Canonical TF-IDF vector text for an observation ROW (snake_case DB fields). Single source for
+ * EVERY (re)build path — save, enrich, rebuild_vectors, cluster-merge, compress, normalize — so
+ * they encode the identical field set and can't drift (V-F1). Mirrors the FTS-searchable columns
+ * incl. lesson_learned + search_aliases (finding #8). `concepts` may be an array or a string.
+ */
+export function vecTextForRow(row) {
+  if (!row) return '';
+  const concepts = Array.isArray(row.concepts) ? row.concepts.join(' ') : (row.concepts || '');
+  return [row.title || '', row.narrative || '', concepts, row.lesson_learned || '', row.search_aliases || '']
+    .filter(Boolean).join(' ');
+}
+
 /**
  * Build global vocabulary (IDF table) from all active observations.
  * @param {object} db - better-sqlite3 database
@@ -214,7 +227,7 @@ export function _resetVocabCache() { _vocabCache = null; }
  */
 export function buildVocabulary(db, { dim = VOCAB_DIM } = {}) {
   const rows = db.prepare(`
-    SELECT title, narrative, concepts FROM observations
+    SELECT title, narrative, concepts, lesson_learned, search_aliases FROM observations
     WHERE COALESCE(compressed_into, 0) = 0 AND superseded_at IS NULL
   `).all();
 
@@ -224,7 +237,10 @@ export function buildVocabulary(db, { dim = VOCAB_DIM } = {}) {
   // Count document frequency for each term
   const df = new Map();
   for (const row of rows) {
-    const text = [row.title || '', row.narrative || '', row.concepts || ''].join(' ');
+    // V-F2: include lesson_learned + search_aliases so terms living only there get a vocab
+    // dimension (else computeVector silently drops them — the exact paraphrase-bridge terms
+    // search_aliases exists to carry). Mirrors vecTextForRow's field set.
+    const text = [row.title || '', row.narrative || '', row.concepts || '', row.lesson_learned || '', row.search_aliases || ''].join(' ');
     const docTerms = new Set(tokenize(text));
     for (const term of docTerms) {
       df.set(term, (df.get(term) || 0) + 1);

package/utils.mjs

@@ -404,7 +404,13 @@ export function isMetaTriggerPrompt(text) {
     .replace(/上次(到哪了|说到哪了)?/g, '')
     .replace(/总结一下|复盘一下/g, '')
     .replace(/前面(的)?(工作|话题|讨论|内容)/g, '')
+    // R3 H-M2: control phrases named in lesson #8287 / common continuations. Phrase-anchored
+    // (or bare suffix/prefix) so a real subject that merely contains one keeps its other tokens.
+    .replace(/停了|停不下来/g, '')
+    .replace(/怎么(回事|了|样了|停的)?/g, '')
+    .replace(/再来(一次|一遍|一下)?/g, '')
     .replace(/\/?(clear|exit)\b/gi, '')
+    .replace(/\b(go on|go ahead|keep going|carry on|proceed|why(?:'?d| did) you stop)\b/gi, '')
     .replace(/\b(commit|continue|resume|push|save|restart|exit|next)\b/gi, '')
     .replace(/[，,。.!！?？:：;；()（）【】[\]\s/\\-]+/g, '')
     .trim();