openclaw-memory-alibaba-local 1.0.7 → 1.0.8

package/index.ts

@@ -1298,12 +1298,27 @@ async function captureUserMemoryFromInboundTexts(
       embeddingResults.push({ item: { ...item, text: truncated }, vectors });
     }
   
-    // 2. Recall top-3 similar existing world_facts per item, merge & dedup (cap ~10)
-    const allVectors = embeddingResults.flatMap((r) => r.vectors);
+    // 2. Per-item recall: for each event item recall top-3 similar existing world_facts, then dedup
     const recallMinScore = Math.max(0.5, cfg.similarityThresholdUserMemory - 0.35);
-    const existingCandidates = allVectors.length > 0
-      ? await db.searchMerged(agentId, allVectors, 3, recallMinScore, [WORLD_FACT])
-      : [];
+    const PER_ITEM_RECALL = 3;
+    const candidateMap = new Map<string, MemorySearchResult>();
+    for (const er of embeddingResults) {
+      const perItemHits = er.vectors.length > 0
+        ? await db.searchMerged(agentId, er.vectors, PER_ITEM_RECALL, recallMinScore, [WORLD_FACT])
+        : [];
+      for (const h of perItemHits) {
+        const key = `${String(h.entry.category)}\0${h.entry.text}`;
+        const prev = candidateMap.get(key);
+        if (!prev || h.score > prev.score) {
+          candidateMap.set(key, h);
+        }
+      }
+    }
+    const existingCandidates = [...candidateMap.values()].sort((a, b) => b.score - a.score);
+    console.log(`[openclaw-memory-alibaba-local] worldImageExtraction recall: ${embeddingResults.length} items × ${PER_ITEM_RECALL} = raw ${embeddingResults.length * PER_ITEM_RECALL}, deduped ${existingCandidates.length}, minScore=${recallMinScore}`);
+    if (existingCandidates.length > 0) {
+      console.log(`[openclaw-memory-alibaba-local] worldImageExtraction recall found ${existingCandidates.length} candidates: ${existingCandidates.map((c) => `[${c.score.toFixed(3)}] ${c.entry.text.slice(0, 60)}`).join(" | ")}`);
+    }
   
     // 3. LLM CRUD decision
     console.log(`[openclaw-memory-alibaba-local] worldImageExtraction input: ${eventItems.length} event items, ${existingCandidates.length} existing candidates`);
@@ -1400,9 +1415,13 @@ async function captureUserMemoryFromInboundTexts(
     // 2. Recall top-10 similar existing memories for ALL new extractions (agentId global, USER_MEMORY scope)
     const allVectors = embeddingResults.flatMap((r) => r.vectors);
     const recallMinScore = Math.max(0.5, cfg.similarityThresholdUserMemory - 0.35);
+    console.log(`[openclaw-memory-alibaba-local] userImageExtraction recall: ${allVectors.length} query vectors, minScore=${recallMinScore}`);
     const existingCandidates = allVectors.length > 0
       ? await db.searchMerged(agentId, allVectors, 10, recallMinScore, [...USER_MEMORY_CATEGORIES])
       : [];
+    if (existingCandidates.length > 0) {
+      console.log(`[openclaw-memory-alibaba-local] userImageExtraction recall found ${existingCandidates.length} candidates: ${existingCandidates.map((c) => `[${c.score.toFixed(3)}] ${c.entry.text.slice(0, 60)}`).join(" | ")}`);
+    }
 
     // 3. Call user image extraction LLM
     console.log(`[openclaw-memory-alibaba-local] userImageExtraction input: ${userItems.length} user items (${eventItems.length} event items bypassed), ${existingCandidates.length} existing candidates`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openclaw-memory-alibaba-local",
-  "version": "1.0.7",
+  "version": "1.0.8",
   "description": "OpenClaw memory plugin: local LanceDB + DashScope-compatible embeddings",
   "type": "module",
   "engines": {