openclaw-memory-alibaba-local 1.0.13 → 1.0.14

package/index.ts

@@ -181,13 +181,6 @@ function formatRelevantMemoriesContext(
   ].join("\n");
 }
 
-function getThresholdForCategory(cfg: MemoryConfig, category: MemoryCategory): number {
-  if (isUserMemoryCategory(category) || isFullContextSourceCategory(category) || category === FULL_CONTEXT_MEMORY) {
-    return cfg.similarityThresholdUserMemory;
-  }
-  return cfg.similarityThresholdSelfImproving;
-}
-
 /** 精简日志：仅记录 tag + prompt 字符数，不贴原文。 */
 function logLlmCall(tag: string, promptChars: number): void {
   console.debug(`[openclaw-memory-alibaba-local] llm ${tag} prompt (${promptChars} chars)`);
@@ -1182,6 +1175,16 @@ async function runAgentEndCapture(
 
   console.debug(`[openclaw-memory-alibaba-local] agentEndCapture fullRows=${fullRows.length} userTexts=${userRawTexts.length} uaLines=${uaLines.length}`);
 
+  // Save cursor BEFORE extraction so that a failure in extraction/storage
+  // does not leave the cursor un-advanced, which would cause duplicate
+  // full_context rows on retry.
+  map[key] = {
+    version: 2,
+    roleCounts: { ...running },
+    lastMessagesLength: messages.length,
+  };
+  saveAgentEndCursorMap(lancedbDir, map);
+
   if (fullRows.length > 0) {
     await db.storeMany(
       agentId,
@@ -1205,13 +1208,6 @@ async function runAgentEndCapture(
     captureUserMemoryFromInboundTexts(cfg, db, backend, agentId, sid, userId, userRawTexts),
     captureSelfImprovingFromLines(cfg, db, backend, agentId, sid, userId, uaLines),
   ]);
-
-  map[key] = {
-    version: 2,
-    roleCounts: { ...running },
-    lastMessagesLength: messages.length,
-  };
-  saveAgentEndCursorMap(lancedbDir, map);
 }
 
 /** User memory from raw user message texts (agent_end user delta). */
@@ -1274,8 +1270,10 @@ async function captureUserMemoryFromInboundTexts(
   // ---- Always split: User-related items vs event items ----
   const userItems: LLMExtractionItem[] = [];
   const eventItems: LLMExtractionItem[] = [];
+  // Match both English "User" and Chinese "用户" to correctly route user preferences
+  const USER_SUBJECT_RE = /\bUser\b|用户/;
   for (const item of extractions) {
-    if (/\bUser\b/.test(item.text)) {
+    if (USER_SUBJECT_RE.test(item.text)) {
       userItems.push(item);
     } else {
       eventItems.push(item);
@@ -1646,21 +1644,16 @@ async function storeOneCaptureItem(
   if (vectors.length === 0) {
     throw new Error("openclaw-memory-alibaba-local: encodeForStorage returned no vectors");
   }
-  const threshold = getThresholdForCategory(cfg, item.category);
   const dedupCategories = getDedupCategories(item.category);
   const rows = buildChunkRows(item, vectors, options);
 
-  if (!cfg.memory_duplication_conflict_process) {
-    const similar = await db.searchMerged(agentId, vectors, 1, threshold, [...dedupCategories]);
-    if (similar.length > 0) {
-      await deleteSimilarLogicalMemory(db, agentId, options?.sessionId, similar[0]!);
-    }
-    const stored = await db.storeMany(agentId, rows);
-    return { action: similar.length > 0 ? "updated" : "created", entry: stored[0]! };
-  }
-
-  // With conflict_process: simple similarity-based dedup (batch merge handles LLM dedup for user memory)
-  const recallMinScore = Math.max(0.5, threshold - 0.35);
+  // Guard against near-exact duplicates only (score >= 0.92).
+  // Lower thresholds (e.g. 0.65) would incorrectly treat similar-but-different
+  // items as duplicates (e.g. "likes apples" vs "likes fish").
+  // Semantic dedup for lower-similarity candidates is handled by the LLM CRUD
+  // pipeline (userPipeline / eventPipeline) when memory_duplication_conflict_process is on.
+  const NEAR_EXACT_THRESHOLD = 0.92;
+  const recallMinScore = Math.max(0.5, NEAR_EXACT_THRESHOLD - 0.15);
   const candidates = await db.searchMerged(
     agentId,
     vectors,
@@ -1668,11 +1661,11 @@ async function storeOneCaptureItem(
     recallMinScore,
     [...dedupCategories],
   );
-  if (candidates.length > 0 && candidates[0]!.score >= threshold) {
+  if (candidates.length > 0 && candidates[0]!.score >= NEAR_EXACT_THRESHOLD) {
     await deleteSimilarLogicalMemory(db, agentId, options?.sessionId, candidates[0]!);
   }
   const stored = await db.storeMany(agentId, rows);
-  return { action: candidates.length > 0 && candidates[0]!.score >= threshold ? "updated" : "created", entry: stored[0]! };
+  return { action: candidates.length > 0 && candidates[0]!.score >= NEAR_EXACT_THRESHOLD ? "updated" : "created", entry: stored[0]! };
 }
 
 // ---------------------------------------------------------------------------

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openclaw-memory-alibaba-local",
-  "version": "1.0.13",
+  "version": "1.0.14",
   "description": "OpenClaw memory plugin: local LanceDB + DashScope-compatible embeddings",
   "type": "module",
   "engines": {

package/prompts.ts

@@ -205,8 +205,9 @@ Only INSERT or UPDATE information that reveals something lasting about the User:
 
 # Refinement Principles
 1. **Prefer the richer version**: When a batch item and a Store item describe the same topic, keep whichever has the most information. If the batch adds new details, UPDATE to include them.
-2. **High cohesion**: Only merge entries about the exact same topic. Entries about different topics stay separate.
-3. **Strip date prefixes**: Input text may contain [date] or [as of ...] prefixes — remove them from the output text. User profile memories are evergreen and should not carry temporal tags.
+2. **High cohesion**: Only merge entries about the exact same specific topic. Entries about different topics stay separate.
+3. **Multiple preferences coexist**: Different concrete items under the same category are NOT duplicates. For example, "User likes apples" and "User likes fish" are two separate preferences — INSERT both, do NOT UPDATE or DELETE one for the other. Only UPDATE/DELETE when the new item truly contradicts or refines the old one (e.g. "User no longer likes apples" replaces "User likes apples").
+4. **Strip date prefixes**: Input text may contain [date] or [as of ...] prefixes — remove them from the output text. User profile memories are evergreen and should not carry temporal tags.
 
 # Actions (one per batch index)
 - **INSERT**: New lasting personal info not in Store.
@@ -270,7 +271,8 @@ Only INSERT or UPDATE information that captures a concrete, verifiable fact or e
 1. **Prefer the richer version**: When a batch item and a Store item describe the same topic, keep whichever has the most information. If the batch is richer, DELETE the old Store item and INSERT the batch item. If they are roughly equal, UPDATE to merge details.
 2. **Preserve temporal markers**: Keep [as of ...] or [date] prefixes — world facts are time-sensitive.
 3. **High cohesion**: Only merge entries about the exact same event or fact. Different events stay separate even if related.
-4. **Contradiction = replace**: If a batch item directly contradicts a Store item (e.g. different outcome), DELETE the old item and INSERT the new one.
+4. **Multiple items coexist**: Different concrete items under the same category are NOT duplicates. For example, "likes apples" and "likes fish" are two separate facts — INSERT both. Only DELETE when the new item truly contradicts the old one (e.g. a corrected outcome).
+5. **Contradiction = replace**: If a batch item directly contradicts a Store item (e.g. different outcome), DELETE the old item and INSERT the new one.
 
 # Actions (one per batch index)
 - **INSERT**: New world fact not in Store.