openclaw-memory-alibaba-local 1.0.5 → 1.0.6

package/index.ts

@@ -63,6 +63,7 @@ import type { MemoryEntry, MemorySearchResult } from "./db.js";
 import {
   buildMemoryExtractionPrompt,
   buildUserImageExtractionPrompt,
+  buildWorldImageExtractionPrompt,
   SELF_IMPROVING_EXTRACTION_INSTRUCTIONS,
 } from "./prompts.js";
 import { extractUserQueryForRecall, stripForLogicalMemoryExtraction } from "./prompt-strip.js";
@@ -782,6 +783,119 @@ async function extractUserImageWithLLM(
   }
 }
 
+// ---------------------------------------------------------------------------
+// World image extraction: recall + LLM CRUD for world facts
+// ---------------------------------------------------------------------------
+
+type WorldImageAction =
+  | { action: "insert"; text: string; importance: number }
+  | { action: "update"; memoryId: string; text: string; importance: number }
+  | { action: "delete"; memoryId: string }
+  | { action: "skip" };
+
+/**
+ * World image extraction: given N new world-fact extractions + M existing
+ * similar world-fact memories, ask LLM to decide insert / update / skip / delete for each.
+ */
+async function extractWorldImageWithLLM(
+  llmConfig: LLMConfig,
+  newItems: LLMExtractionItem[],
+  existingCandidates: MemorySearchResult[],
+): Promise<WorldImageAction[]> {
+  if (newItems.length === 0) return [];
+
+  const newForPrompt = newItems.map((item, i) => ({
+    index: i,
+    text: item.text,
+    importance: item.importance,
+  }));
+  const existingForPrompt = existingCandidates.map((r) => ({
+    id: r.entry.id,
+    text: r.entry.text,
+  }));
+
+  const prompt = buildWorldImageExtractionPrompt(newForPrompt, existingForPrompt);
+  logLlmCall("worldImageExtraction", prompt.length);
+
+  const openai = new OpenAI({
+    apiKey: llmConfig.apiKey,
+    baseURL: llmConfig.baseUrl,
+  });
+
+  const completion = await openai.chat.completions.create({
+    model: llmConfig.model,
+    messages: [{ role: "user", content: prompt }],
+    temperature: 0,
+    max_tokens: 8192,
+  });
+  const raw = completion.choices[0]?.message?.content?.trim() ?? "";
+
+  const existingIdSet = new Set(existingCandidates.map((r) => r.entry.id));
+
+  try {
+    const parsed = JSON.parse(stripMarkdownJsonFence(raw)) as {
+      actions?: Array<{
+        index?: number;
+        action?: string;
+        text?: string;
+        importance?: unknown;
+        memoryId?: string;
+      }>;
+    };
+    const list = Array.isArray(parsed.actions) ? parsed.actions : [];
+    const resultMap = new Map<number, WorldImageAction[]>();
+    // Default: insert for each index
+    for (let i = 0; i < newItems.length; i++) {
+      resultMap.set(i, [{
+        action: "insert" as const,
+        text: newItems[i]!.text,
+        importance: newItems[i]!.importance,
+      }]);
+    }
+
+    for (const a of list) {
+      const idx = typeof a.index === "number" ? a.index : -1;
+      if (idx < 0 || idx >= newItems.length) continue;
+
+      if (a.action === "skip") {
+        resultMap.set(idx, [{ action: "skip" }]);
+      } else if (a.action === "delete" && typeof a.memoryId === "string" && existingIdSet.has(a.memoryId)) {
+        const existing = resultMap.get(idx) ?? [];
+        resultMap.set(idx, [{ action: "delete", memoryId: a.memoryId }, ...existing.filter(x => x.action !== "skip")]);
+      } else if (a.action === "update" && typeof a.memoryId === "string" && existingIdSet.has(a.memoryId)) {
+        const text = typeof a.text === "string" ? a.text.trim() : "";
+        if (text.length < 10) continue;
+        resultMap.set(idx, [{
+          action: "update",
+          memoryId: a.memoryId,
+          text,
+          importance: clampImportance(a.importance ?? newItems[idx]!.importance),
+        }]);
+      } else if (a.action === "insert") {
+        const text = typeof a.text === "string" ? a.text.trim() : "";
+        const finalText = text.length >= 10 ? text : newItems[idx]!.text;
+        resultMap.set(idx, [{
+          action: "insert",
+          text: finalText,
+          importance: clampImportance(a.importance ?? newItems[idx]!.importance),
+        }]);
+      }
+    }
+    const result: WorldImageAction[] = [];
+    for (let i = 0; i < newItems.length; i++) {
+      result.push(...(resultMap.get(i) ?? []));
+    }
+    return result;
+  } catch (err: unknown) {
+    console.warn(`[openclaw-memory-alibaba-local] worldImageExtraction JSON parse failed, fallback insert all: ${err}`);
+    return newItems.map((item) => ({
+      action: "insert" as const,
+      text: item.text,
+      importance: item.importance,
+    }));
+  }
+}
+
 function shouldCapture(text: string, options?: { maxChars?: number }): boolean {
   const maxChars = options?.maxChars ?? DEFAULT_CAPTURE_MAX_CHARS;
   if (text.length < 10 || text.length > maxChars) return false;
@@ -1158,50 +1272,103 @@ async function captureUserMemoryFromInboundTexts(
   // ---- Parallel: event-item pipeline & user-item pipeline ----
   const eventPipeline = async () => {
     if (eventItems.length === 0) return;
-    if (cfg.memory_duplication_conflict_process) {
-      type EmbeddedItem = { item: LLMExtractionItem; vectors: number[][]; primary: number[] };
-      const embedded: EmbeddedItem[] = [];
-      for (const item of eventItems) {
-        const truncated = truncateForCapture(item.text, cfg.captureMaxChars);
-        const { vectors } = await backend.encodeForStorage(truncated);
-        embedded.push({ item: { ...item, text: truncated }, vectors, primary: vectors[0] });
+  
+    // --- No LLM fallback: simple dedup ---
+    if (!cfg.llm) {
+      for (const e of eventItems) {
+        const text = truncateForCapture(e.text, cfg.captureMaxChars);
+        if (await db.existsSemanticDuplicate(agentId, sessionKey, e.category, text)) {
+          continue;
+        }
+        await storeOneCaptureItem(agentId, { category: e.category, text, importance: e.importance }, cfg, db, backend, {
+          userId,
+          sessionId: sessionKey,
+        });
       }
-
-      const clusters = greedyCluster(embedded, (e) => e.primary, (e) => extractDatePrefix(e.item.text), 0.85, 3);
-
-      for (const cluster of clusters) {
-        const text = concatDedupeDate(cluster.map((e) => e.item.text));
-        const importance = Math.max(...cluster.map((e) => e.item.importance));
-        const { vectors } = await backend.encodeForStorage(text);
+      return;
+    }
+  
+    // --- LLM path: recall + CRUD ---
+  
+    // 1. Batch embed all event items
+    const embeddingResults: { item: LLMExtractionItem; vectors: number[][] }[] = [];
+    for (const item of eventItems) {
+      const truncated = truncateForCapture(item.text, cfg.captureMaxChars);
+      const { vectors } = await backend.encodeForStorage(truncated);
+      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);
+    const recallMinScore = Math.max(0.5, cfg.similarityThresholdUserMemory - 0.35);
+    const existingCandidates = allVectors.length > 0
+      ? await db.searchMerged(agentId, allVectors, 3, recallMinScore, [WORLD_FACT])
+      : [];
+  
+    // 3. LLM CRUD decision
+    console.log(`[openclaw-memory-alibaba-local] worldImageExtraction input: ${eventItems.length} event items, ${existingCandidates.length} existing candidates`);
+    const worldActions = await extractWorldImageWithLLM(
+      cfg.llm,
+      embeddingResults.map((r) => r.item),
+      existingCandidates,
+    ).catch((err: unknown) => {
+      console.warn(`[openclaw-memory-alibaba-local] worldImageExtraction LLM failed, fallback insert all: ${err}`);
+      return embeddingResults.map((r): WorldImageAction => ({
+        action: "insert" as const,
+        text: r.item.text,
+        importance: r.item.importance,
+      }));
+    });
+  
+    // 4. Execute actions
+    let insertCount = 0;
+    for (const action of worldActions) {
+      if (action.action === "skip") continue;
+  
+      if (action.action === "delete") {
+        const hit = existingCandidates.find((c) => c.entry.id === action.memoryId);
+        if (hit) {
+          await deleteSimilarLogicalMemory(db, agentId, sessionKey, hit);
+        }
+        continue;
+      }
+  
+      if (action.action === "update") {
+        const hit = existingCandidates.find((c) => c.entry.id === action.memoryId);
+        if (hit) {
+          await deleteSimilarLogicalMemory(db, agentId, sessionKey, hit);
+        }
+        const { vectors } = await backend.encodeForStorage(action.text);
         const rows = buildChunkRows(
-          { category: WORLD_FACT as MemoryCategory, text, importance },
+          { category: WORLD_FACT as MemoryCategory, text: action.text, importance: action.importance },
           vectors,
           { userId, sessionId: sessionKey },
         );
         await db.storeMany(agentId, rows);
+        insertCount++;
+      } else {
+        // insert
+        const { vectors } = await backend.encodeForStorage(action.text);
+        const rows = buildChunkRows(
+          { category: WORLD_FACT as MemoryCategory, text: action.text, importance: action.importance },
+          vectors,
+          { userId, sessionId: sessionKey },
+        );
+        await db.storeMany(agentId, rows);
+        insertCount++;
       }
-      console.log(`[openclaw-memory-alibaba-local] clustered ${eventItems.length} event items \u2192 ${clusters.length} entries (max 3 per cluster, threshold 0.85)`);
-
-      // World fact LRU GC: every 10 insertions
-      worldFactGcCounter++;
+    }
+    console.log(`[openclaw-memory-alibaba-local] worldImageExtraction done: ${worldActions.length} actions, ${insertCount} stored`);
+  
+    // 5. World fact LRU GC: every 10 insertions
+    if (insertCount > 0) {
+      worldFactGcCounter += insertCount;
       if (worldFactGcCounter >= 10) {
         worldFactGcCounter = 0;
         db.gcWorldFact(agentId, 10_000, 30 * 24 * 60 * 60 * 1000, 25_000).catch((err) =>
           console.warn(`[openclaw-memory-alibaba-local] gcWorldFact failed: ${err}`),
         );
       }
-    } else {
-      // Simple dedup path for event items
-      for (const e of eventItems) {
-        const text = truncateForCapture(e.text, cfg.captureMaxChars);
-        if (await db.existsSemanticDuplicate(agentId, sessionKey, e.category, text)) {
-          continue;
-        }
-        await storeOneCaptureItem(agentId, { category: e.category, text, importance: e.importance }, cfg, db, backend, {
-          userId,
-          sessionId: sessionKey,
-        });
-      }
     }
   };
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openclaw-memory-alibaba-local",
-  "version": "1.0.5",
+  "version": "1.0.6",
   "description": "OpenClaw memory plugin: local LanceDB + DashScope-compatible embeddings",
   "type": "module",
   "engines": {

package/prompts.ts

@@ -244,3 +244,60 @@ export function buildUserImageExtractionPrompt(
     `\nToday is ${today}.\n\nNew extractions:\n${newSection}\n\nExisting memories:\n${existingSection}\n`
   );
 }
+
+// ---------------------------------------------------------------------------
+// 4. WORLD IMAGE EXTRACTION (deduplicate & reconcile world facts)
+// ---------------------------------------------------------------------------
+
+export const WORLD_IMAGE_EXTRACTION_INSTRUCTIONS = `You are a World-Fact Organizer that keeps a concise, non-redundant knowledge base of real-world events, facts, and context mentioned in conversations.
+
+# Inputs
+- **Batch** (indexed 0..N-1): newly extracted world facts / events.
+- **Store** (with id): existing world-fact memories in the database. May be empty.
+
+# What to Keep
+Only INSERT or UPDATE information that captures a concrete, verifiable fact or event:
+1. Events — what happened, when, where, who was involved
+2. Factual statements — statistics, dates, locations, outcomes
+3. Contextual knowledge — project status, external conditions, third-party decisions
+
+# What to SKIP
+- Information already fully covered by a Store item (same meaning, same or less detail)
+- Vague or speculative statements with no concrete fact
+- Purely conversational filler with no informational content
+
+# Reconciliation 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.
+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.
+
+# Actions (one per batch index)
+- **INSERT**: New world fact not in Store.
+- **UPDATE** (memoryId): Store item covers the same event/fact; merge to produce the richer version.
+- **SKIP**: Already fully covered by Store, or not a concrete fact.
+- **DELETE** (memoryId): Batch item contradicts a Store item. Delete old; INSERT new.
+
+# Output
+Reply with ONLY a JSON object:
+{"actions":[
+  {"index":0,"action":"insert","text":"[2026-04-01] Project X launched v2.0","importance":0.5},
+  {"index":1,"action":"skip"},
+  {"index":2,"action":"update","memoryId":"uuid","text":"[2026-03-28] Company Y acquired Z for $2B, deal finalized","importance":0.5},
+  {"index":3,"action":"delete","memoryId":"uuid"}
+]}
+Every batch index must appear. A single index may produce BOTH delete + insert.
+`;
+
+export function buildWorldImageExtractionPrompt(
+  newItems: Array<{ index: number; text: string; importance: number }>,
+  existingMemories: Array<{ id: string; text: string }>,
+): string {
+  const today = new Date().toISOString().split("T")[0];
+  const newSection = JSON.stringify(newItems, null, 2);
+  const existingSection = JSON.stringify(existingMemories, null, 2);
+  return (
+    WORLD_IMAGE_EXTRACTION_INSTRUCTIONS +
+    `\nToday is ${today}.\n\nNew extractions:\n${newSection}\n\nExisting memories:\n${existingSection}\n`
+  );
+}