@memtensor/memos-local-openclaw-plugin 1.0.5 → 1.0.6-beta.10

package/src/ingest/providers/index.ts

@@ -329,6 +329,27 @@ export class Summarizer {
     return this.strongCfg;
   }
 
+  // ─── OpenClaw Prompts ───
+
+  static readonly OPENCLAW_TOPIC_JUDGE_PROMPT = `You are a conversation topic change detector.
+Given a CURRENT CONVERSATION SUMMARY and a NEW USER MESSAGE, decide: has the user started a COMPLETELY NEW topic that is unrelated to the current conversation?
+Reply with a single word: "NEW" if topic changed, "SAME" if it continues.`;
+
+  static readonly OPENCLAW_FILTER_RELEVANT_PROMPT = `You are a memory relevance judge.
+Given a QUERY and CANDIDATE memories, decide: does each candidate help answer the query?
+RULES:
+1. Include candidates whose content provides useful facts/context for the query.
+2. Exclude candidates that merely share a topic but contain no useful information.
+3. DEDUPLICATION: When multiple candidates convey the same or very similar information, keep ONLY the most complete one and exclude the rest.
+4. If none help, return {"relevant":[],"sufficient":false}.
+OUTPUT — JSON only: {"relevant":[1,3],"sufficient":true}`;
+
+  static readonly OPENCLAW_DEDUP_JUDGE_PROMPT = `You are a memory deduplication system.
+Given a NEW memory summary and EXISTING candidates, decide if the new memory duplicates any existing one.
+Reply with JSON: {"action":"MERGE","mergeTarget":2,"reason":"..."} or {"action":"NEW","reason":"..."}`;
+
+  static readonly OPENCLAW_TASK_SUMMARY_PROMPT = `Summarize the following task conversation into a structured report. Preserve key decisions, code, commands, and outcomes. Use the same language as the input.`;
+
   // ─── OpenClaw API Implementation ───
 
   private requireOpenClawAPI(): void {
@@ -360,7 +381,7 @@ export class Summarizer {
   private async summarizeTaskOpenClaw(text: string): Promise<string> {
     this.requireOpenClawAPI();
     const prompt = [
-      OPENCLAW_TASK_SUMMARY_PROMPT,
+      Summarizer.OPENCLAW_TASK_SUMMARY_PROMPT,
       ``,
       text,
     ].join("\n");
@@ -378,7 +399,7 @@ export class Summarizer {
   private async judgeNewTopicOpenClaw(currentContext: string, newMessage: string): Promise<boolean> {
     this.requireOpenClawAPI();
     const prompt = [
-      OPENCLAW_TOPIC_JUDGE_PROMPT,
+      Summarizer.OPENCLAW_TOPIC_JUDGE_PROMPT,
       ``,
       `CURRENT CONVERSATION SUMMARY:`,
       currentContext,
@@ -409,7 +430,7 @@ export class Summarizer {
       .join("\n");
 
     const prompt = [
-      OPENCLAW_FILTER_RELEVANT_PROMPT,
+      Summarizer.OPENCLAW_FILTER_RELEVANT_PROMPT,
       ``,
       `QUERY: ${query}`,
       ``,
@@ -437,7 +458,7 @@ export class Summarizer {
       .join("\n");
 
     const prompt = [
-      OPENCLAW_DEDUP_JUDGE_PROMPT,
+      Summarizer.OPENCLAW_DEDUP_JUDGE_PROMPT,
       ``,
       `NEW MEMORY:`,
       newSummary,
@@ -466,6 +487,8 @@ function callSummarize(cfg: SummarizerConfig, text: string, log: Logger): Promis
     case "azure_openai":
     case "zhipu":
     case "siliconflow":
+    case "deepseek":
+    case "moonshot":
     case "bailian":
     case "cohere":
     case "mistral":
@@ -489,6 +512,8 @@ function callSummarizeTask(cfg: SummarizerConfig, text: string, log: Logger): Pr
     case "azure_openai":
     case "zhipu":
     case "siliconflow":
+    case "deepseek":
+    case "moonshot":
     case "bailian":
     case "cohere":
     case "mistral":
@@ -512,6 +537,8 @@ function callGenerateTaskTitle(cfg: SummarizerConfig, text: string, log: Logger)
     case "azure_openai":
     case "zhipu":
     case "siliconflow":
+    case "deepseek":
+    case "moonshot":
     case "bailian":
     case "cohere":
     case "mistral":
@@ -535,6 +562,8 @@ function callTopicJudge(cfg: SummarizerConfig, currentContext: string, newMessag
     case "azure_openai":
     case "zhipu":
     case "siliconflow":
+    case "deepseek":
+    case "moonshot":
     case "bailian":
     case "cohere":
     case "mistral":
@@ -558,6 +587,8 @@ function callFilterRelevant(cfg: SummarizerConfig, query: string, candidates: Ar
     case "azure_openai":
     case "zhipu":
     case "siliconflow":
+    case "deepseek":
+    case "moonshot":
     case "bailian":
     case "cohere":
     case "mistral":
@@ -581,6 +612,8 @@ function callJudgeDedup(cfg: SummarizerConfig, newSummary: string, candidates: A
     case "azure_openai":
     case "zhipu":
     case "siliconflow":
+    case "deepseek":
+    case "moonshot":
     case "bailian":
     case "cohere":
     case "mistral":
@@ -629,91 +662,3 @@ function wordCount(text: string): number {
   if (noCjk) count += noCjk.split(/\s+/).filter(Boolean).length;
   return count;
 }
-
-// ─── OpenClaw Prompt Templates ───
-
-const OPENCLAW_TASK_SUMMARY_PROMPT = `You create a DETAILED task summary from a multi-turn conversation. This summary will be the ONLY record of this conversation, so it must preserve ALL important information.
-
-CRITICAL LANGUAGE RULE: You MUST write in the SAME language as the user's messages. Chinese input → Chinese output. English input → English output. NEVER mix languages.
-
-Output EXACTLY this structure:
-
-📌 Title
-A short, descriptive title (10-30 characters). Like a chat group name.
-
-🎯 Goal
-One sentence: what the user wanted to accomplish.
-
-📋 Key Steps
-- Describe each meaningful step in detail
-- Include the ACTUAL content produced: code snippets, commands, config blocks, formulas, key paragraphs
-- For code: include the function signature and core logic (up to ~30 lines per block), use fenced code blocks
-- For configs: include the actual config values and structure
-- For lists/instructions: include the actual items, not just "provided a list"
-- Merge only truly trivial back-and-forth (like "ok" / "sure")
-- Do NOT over-summarize: "provided a function" is BAD; show the actual function
-
-✅ Result
-What was the final outcome? Include the final version of any code/config/content produced.
-
-💡 Key Details
-- Decisions made, trade-offs discussed, caveats noted, alternative approaches mentioned
-- Specific values: numbers, versions, thresholds, URLs, file paths, model names
-- Omit this section only if there truly are no noteworthy details
-
-RULES:
-- This summary is a KNOWLEDGE BASE ENTRY, not a brief note. Be thorough.
-- PRESERVE verbatim: code, commands, URLs, file paths, error messages, config values, version numbers, names, amounts
-- DISCARD only: greetings, filler, the assistant explaining what it will do before doing it
-- Replace secrets (API keys, tokens, passwords) with [REDACTED]
-- Target length: 30-50% of the original conversation length. Longer conversations need longer summaries.
-- Output summary only, no preamble.`;
-
-const OPENCLAW_TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given a summary of the CURRENT conversation and a NEW user message, determine if the new message starts a DIFFERENT topic/task.
-
-Answer ONLY "NEW" or "SAME".
-
-Rules:
-- "NEW" = the new message is about a completely different subject, project, or task
-- "SAME" = the new message continues, follows up on, or is closely related to the current topic
-- Follow-up questions, clarifications, refinements, bug fixes, or next steps on the same task = SAME
-- Greetings or meta-questions like "你好" or "谢谢" without new substance = SAME
-- A clearly unrelated request (e.g., current topic is deployment, new message asks about cooking) = NEW
-
-Output exactly one word: NEW or SAME`;
-
-const OPENCLAW_FILTER_RELEVANT_PROMPT = `You are a memory relevance judge. Given a user's QUERY and a list of CANDIDATE memory summaries, do two things:
-
-1. Select ALL candidates that could be useful for answering the query. When in doubt, INCLUDE the candidate.
-   - For questions about lists, history, or "what/where/who" across multiple items, include ALL matching items.
-   - For factual lookups, a single direct answer is enough.
-2. Judge whether the selected memories are SUFFICIENT to fully answer the query WITHOUT fetching additional context.
-
-IMPORTANT for "sufficient" judgment:
-- sufficient=true ONLY when the memories contain a concrete ANSWER, fact, decision, or actionable information that directly addresses the query.
-- sufficient=false when the memories only repeat the question, show related topics but lack the specific detail, or contain partial information.
-
-Output a JSON object with exactly two fields:
-{"relevant":[1,3,5],"sufficient":true}
-
-- "relevant": array of candidate numbers that are useful. Empty array [] if none are relevant.
-- "sufficient": true ONLY if the memories contain a direct answer; false otherwise.
-
-Output ONLY the JSON object, nothing else.`;
-
-const OPENCLAW_DEDUP_JUDGE_PROMPT = `You are a memory deduplication system. Given a NEW memory summary and several EXISTING memory summaries, determine the relationship.
-
-For each EXISTING memory, the NEW memory is either:
-- "DUPLICATE": NEW is fully covered by an EXISTING memory — no new information at all
-- "UPDATE": NEW contains information that supplements or updates an EXISTING memory (new data, status change, additional detail)
-- "NEW": NEW is a different topic/event despite surface similarity
-
-Pick the BEST match among all candidates. If none match well, choose "NEW".
-
-Output a single JSON object:
-- If DUPLICATE: {"action":"DUPLICATE","targetIndex":2,"reason":"..."}
-- If UPDATE: {"action":"UPDATE","targetIndex":3,"reason":"...","mergedSummary":"a combined summary preserving all info from both old and new, same language as input"}
-- If NEW: {"action":"NEW","reason":"..."}
-
-CRITICAL: mergedSummary must use the SAME language as the input. Output ONLY the JSON object.`;
-

package/src/ingest/providers/openai.ts

@@ -70,7 +70,7 @@ export async function summarizeTaskOpenAI(
   const resp = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({
+    body: JSON.stringify(buildRequestBody(cfg, {
       model,
       temperature: cfg.temperature ?? 0.1,
       max_tokens: 4096,
@@ -78,7 +78,7 @@ export async function summarizeTaskOpenAI(
         { role: "system", content: TASK_SUMMARY_PROMPT },
         { role: "user", content: text },
       ],
-    }),
+    })),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
   });
 
@@ -119,7 +119,7 @@ export async function generateTaskTitleOpenAI(
   const resp = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({
+    body: JSON.stringify(buildRequestBody(cfg, {
       model,
       temperature: 0,
       max_tokens: 100,
@@ -127,7 +127,7 @@ export async function generateTaskTitleOpenAI(
         { role: "system", content: TASK_TITLE_PROMPT },
         { role: "user", content: text },
       ],
-    }),
+    })),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
   });
 
@@ -156,14 +156,14 @@ export async function summarizeOpenAI(
   const resp = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({
+    body: JSON.stringify(buildRequestBody(cfg, {
       model,
       temperature: cfg.temperature ?? 0,
       messages: [
         { role: "system", content: SYSTEM_PROMPT },
         { role: "user", content: `[TEXT TO SUMMARIZE]\n${text}\n[/TEXT TO SUMMARIZE]` },
       ],
-    }),
+    })),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 30_000),
   });
 
@@ -223,7 +223,7 @@ export async function judgeNewTopicOpenAI(
   const resp = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({
+    body: JSON.stringify(buildRequestBody(cfg, {
       model,
       temperature: 0,
       max_tokens: 10,
@@ -231,7 +231,7 @@ export async function judgeNewTopicOpenAI(
         { role: "system", content: TOPIC_JUDGE_PROMPT },
         { role: "user", content: userContent },
       ],
-    }),
+    })),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
   });
 
@@ -258,10 +258,11 @@ RULES:
 1. A candidate is relevant if its content provides facts, context, or data that directly supports answering the query.
 2. A candidate that merely shares the same broad topic/domain but contains NO useful information for answering is NOT relevant.
 3. If NO candidate can help answer the query, return {"relevant":[],"sufficient":false} — do NOT force-pick the "least irrelevant" one.
+4. DEDUPLICATION: When multiple candidates convey the same or very similar information, keep ONLY the most complete/detailed one and exclude the rest. Do NOT return near-duplicate snippets.
 
 OUTPUT — JSON only:
 {"relevant":[1,3],"sufficient":true}
-- "relevant": candidate numbers whose content helps answer the query. [] if none can help.
+- "relevant": candidate numbers whose content helps answer the query. [] if none can help. Duplicates removed — only unique information.
 - "sufficient": true only if the selected memories fully answer the query.`;
 
 export interface FilterResult {
@@ -293,7 +294,7 @@ export async function filterRelevantOpenAI(
   const resp = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({
+    body: JSON.stringify(buildRequestBody(cfg, {
       model,
       temperature: 0,
       max_tokens: 200,
@@ -301,7 +302,7 @@ export async function filterRelevantOpenAI(
         { role: "system", content: FILTER_RELEVANT_PROMPT },
         { role: "user", content: `QUERY: ${query}\n\nCANDIDATES:\n${candidateText}` },
       ],
-    }),
+    })),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
   });
 
@@ -385,7 +386,7 @@ export async function judgeDedupOpenAI(
   const resp = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({
+    body: JSON.stringify(buildRequestBody(cfg, {
       model,
       temperature: 0,
       max_tokens: 300,
@@ -393,7 +394,7 @@ export async function judgeDedupOpenAI(
         { role: "system", content: DEDUP_JUDGE_PROMPT },
         { role: "user", content: `NEW MEMORY:\n${newSummary}\n\nEXISTING MEMORIES:\n${candidateText}` },
       ],
-    }),
+    })),
     signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
   });
 
@@ -432,3 +433,20 @@ function normalizeChatEndpoint(url: string): string {
   if (stripped.endsWith("/completions")) return stripped;
   return `${stripped}/chat/completions`;
 }
+
+/**
+ * Zhipu AI (glm-4.7, glm-5, etc.) uses reasoning tokens that consume max_tokens budget,
+ * leaving no room for actual output. This helper injects {"thinking":{"type":"disabled"}}
+ * for zhipu endpoints to disable the built-in reasoning mode.
+ */
+function isZhipuEndpoint(endpoint: string): boolean {
+  return /bigmodel\.cn|zhipuai/.test(endpoint);
+}
+
+function buildRequestBody(cfg: SummarizerConfig, body: Record<string, unknown>): Record<string, unknown> {
+  const endpoint = cfg.endpoint ?? "";
+  if (isZhipuEndpoint(endpoint)) {
+    body.thinking = { type: "disabled" };
+  }
+  return body;
+}

package/src/recall/engine.ts

@@ -62,10 +62,20 @@ export class RecallEngine {
 
     // Step 1b: Pattern search (LIKE-based) as fallback for short terms that
     // trigram FTS cannot match (trigram requires >= 3 chars).
-    const shortTerms = query
-      .replace(/[."""(){}[\]*:^~!@#$%&\\/<>,;'`?？。，！、：""''（）【】《》]/g, " ")
-      .split(/\s+/)
-      .filter((t) => t.length === 2);
+    // For CJK text without spaces, extract bigrams (2-char sliding windows)
+    // so that queries like "唐波是谁" produce ["唐波", "波是", "是谁"].
+    const cleaned = query.replace(/[."""(){}[\]*:^~!@#$%&\\/<>,;'`?？。，！、：""''（）【】《》]/g, " ");
+    const spaceSplit = cleaned.split(/\s+/).filter((t) => t.length === 2);
+    const cjkBigrams: string[] = [];
+    const cjkRuns = cleaned.match(/[\u4e00-\u9fff\u3400-\u4dbf\uF900-\uFAFF]{2,}/g);
+    if (cjkRuns) {
+      for (const run of cjkRuns) {
+        for (let i = 0; i <= run.length - 2; i++) {
+          cjkBigrams.push(run.slice(i, i + 2));
+        }
+      }
+    }
+    const shortTerms = [...new Set([...spaceSplit, ...cjkBigrams])];
     const patternHits = shortTerms.length > 0
       ? this.store.patternSearch(shortTerms, { limit: candidatePool })
       : [];
@@ -74,49 +84,41 @@ export class RecallEngine {
       score: 1 / (i + 1),
     }));
 
-    // Step 1c: Hub memories search — only in Hub mode where local DB owns the
-    // hub_memories data and embeddings were generated by the same Embedder.
-    // Client mode must use remote API (hubSearchMemories) to avoid cross-model
-    // embedding mismatch.
+    // Step 1c: Hub memories — FTS + pattern + cached embeddings (same strategy as chunks/skills).
     let hubMemFtsRanked: Array<{ id: string; score: number }> = [];
     let hubMemVecRanked: Array<{ id: string; score: number }> = [];
     let hubMemPatternRanked: Array<{ id: string; score: number }> = [];
     if (query && this.ctx.config.sharing?.enabled && this.ctx.config.sharing.role === "hub") {
       try {
         const hubFtsHits = this.store.searchHubMemories(query, { maxResults: candidatePool });
-        hubMemFtsRanked = hubFtsHits.map(({ hit }, i) => ({
-          id: `hubmem:${hit.id}`, score: 1 / (i + 1),
-        }));
+        hubMemFtsRanked = hubFtsHits.map(({ hit }, i) => ({ id: `hubmem:${hit.id}`, score: 1 / (i + 1) }));
       } catch { /* hub_memories table may not exist */ }
       if (shortTerms.length > 0) {
         try {
           const hubPatternHits = this.store.hubMemoryPatternSearch(shortTerms, { limit: candidatePool });
-          hubMemPatternRanked = hubPatternHits.map((h, i) => ({
-            id: `hubmem:${h.memoryId}`, score: 1 / (i + 1),
-          }));
+          hubMemPatternRanked = hubPatternHits.map((h, i) => ({ id: `hubmem:${h.memoryId}`, score: 1 / (i + 1) }));
         } catch { /* best-effort */ }
       }
+
       try {
-        const hubMemEmbs = this.store.getVisibleHubMemoryEmbeddings("");
-        if (hubMemEmbs.length > 0) {
-          const qv = await this.embedder.embedQuery(query).catch(() => null);
-          if (qv) {
-            const scored: Array<{ id: string; score: number }> = [];
-            for (const e of hubMemEmbs) {
-              let dot = 0, nA = 0, nB = 0;
-              for (let i = 0; i < qv.length && i < e.vector.length; i++) {
-                dot += qv[i] * e.vector[i]; nA += qv[i] * qv[i]; nB += e.vector[i] * e.vector[i];
-              }
-              const sim = nA > 0 && nB > 0 ? dot / (Math.sqrt(nA) * Math.sqrt(nB)) : 0;
-              if (sim > 0.3) {
-                scored.push({ id: `hubmem:${e.memoryId}`, score: sim });
-              }
+        const qv = await this.embedder.embedQuery(query).catch(() => null);
+        if (qv) {
+          const memEmbs = this.store.getVisibleHubMemoryEmbeddings("__hub__");
+          const scored: Array<{ id: string; score: number }> = [];
+          for (const e of memEmbs) {
+            let dot = 0, nA = 0, nB = 0;
+            const len = Math.min(qv.length, e.vector.length);
+            for (let i = 0; i < len; i++) {
+              dot += qv[i] * e.vector[i]; nA += qv[i] * qv[i]; nB += e.vector[i] * e.vector[i];
             }
-            scored.sort((a, b) => b.score - a.score);
-            hubMemVecRanked = scored.slice(0, candidatePool);
+            const sim = nA > 0 && nB > 0 ? dot / (Math.sqrt(nA) * Math.sqrt(nB)) : 0;
+            if (sim > 0.3) scored.push({ id: `hubmem:${e.memoryId}`, score: sim });
           }
+          scored.sort((a, b) => b.score - a.score);
+          hubMemVecRanked = scored.slice(0, candidatePool);
         }
       } catch { /* best-effort */ }
+
       const hubTotal = hubMemFtsRanked.length + hubMemVecRanked.length + hubMemPatternRanked.length;
       if (hubTotal > 0) {
         this.ctx.log.debug(`recall: hub_memories candidates: fts=${hubMemFtsRanked.length}, vec=${hubMemVecRanked.length}, pattern=${hubMemPatternRanked.length}`);