@memtensor/memos-local-openclaw-plugin 1.0.4 → 1.0.6-beta.1

package/src/ingest/providers/index.ts

@@ -1,7 +1,20 @@
 import * as fs from "fs";
 import * as path from "path";
 import type { SummarizerConfig, SummaryProvider, Logger, OpenClawAPI } from "../../types";
-import { summarizeOpenAI, summarizeTaskOpenAI, generateTaskTitleOpenAI, judgeNewTopicOpenAI, filterRelevantOpenAI, judgeDedupOpenAI, parseFilterResult, parseDedupResult } from "./openai";
+import {
+  summarizeOpenAI,
+  summarizeTaskOpenAI,
+  generateTaskTitleOpenAI,
+  judgeNewTopicOpenAI,
+  filterRelevantOpenAI,
+  judgeDedupOpenAI,
+  parseFilterResult,
+  parseDedupResult,
+  TASK_SUMMARY_PROMPT,
+  TOPIC_JUDGE_PROMPT,
+  FILTER_RELEVANT_PROMPT,
+  DEDUP_JUDGE_PROMPT,
+} from "./openai";
 import type { FilterResult, DedupResult } from "./openai";
 export type { FilterResult, DedupResult } from "./openai";
 import { summarizeAnthropic, summarizeTaskAnthropic, generateTaskTitleAnthropic, judgeNewTopicAnthropic, filterRelevantAnthropic, judgeDedupAnthropic } from "./anthropic";
@@ -360,7 +373,7 @@ export class Summarizer {
   private async summarizeTaskOpenClaw(text: string): Promise<string> {
     this.requireOpenClawAPI();
     const prompt = [
-      OPENCLAW_TASK_SUMMARY_PROMPT,
+      TASK_SUMMARY_PROMPT,
       ``,
       text,
     ].join("\n");
@@ -378,7 +391,7 @@ export class Summarizer {
   private async judgeNewTopicOpenClaw(currentContext: string, newMessage: string): Promise<boolean> {
     this.requireOpenClawAPI();
     const prompt = [
-      OPENCLAW_TOPIC_JUDGE_PROMPT,
+      TOPIC_JUDGE_PROMPT,
       ``,
       `CURRENT CONVERSATION SUMMARY:`,
       currentContext,
@@ -409,7 +422,7 @@ export class Summarizer {
       .join("\n");
 
     const prompt = [
-      OPENCLAW_FILTER_RELEVANT_PROMPT,
+      FILTER_RELEVANT_PROMPT,
       ``,
       `QUERY: ${query}`,
       ``,
@@ -437,7 +450,7 @@ export class Summarizer {
       .join("\n");
 
     const prompt = [
-      OPENCLAW_DEDUP_JUDGE_PROMPT,
+      DEDUP_JUDGE_PROMPT,
       ``,
       `NEW MEMORY:`,
       newSummary,
@@ -466,6 +479,8 @@ 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 +504,8 @@ 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 +529,8 @@ 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 +554,8 @@ 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 +579,8 @@ 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 +604,8 @@ 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 +654,3 @@ 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

@@ -16,7 +16,7 @@ Requirements:
 - Non-Chinese: 5-15 words (aim for 8-12)
 - Output title only`;
 
-const 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.
+export const 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.
 
 ## LANGUAGE RULE (HIGHEST PRIORITY)
 Detect the PRIMARY language of the user's messages. If most user messages are Chinese, ALL output (title, goal, steps, result, details) MUST be in Chinese. If English, output in English. NEVER mix. This rule overrides everything below.
@@ -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),
   });
 
@@ -178,7 +178,7 @@ export async function summarizeOpenAI(
   return json.choices[0]?.message?.content?.trim() ?? "";
 }
 
-const TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given the CURRENT task context and a NEW user message, decide if the new message belongs to the SAME task or starts a NEW one.
+export const TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given the CURRENT task context and a NEW user message, decide if the new message belongs to the SAME task or starts a NEW one.
 
 Answer ONLY "NEW" or "SAME".
 
@@ -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),
   });
 
@@ -246,7 +246,7 @@ export async function judgeNewTopicOpenAI(
   return answer.startsWith("NEW");
 }
 
-const FILTER_RELEVANT_PROMPT = `You are a memory relevance judge.
+export const FILTER_RELEVANT_PROMPT = `You are a memory relevance judge.
 
 Given a QUERY and CANDIDATE memories, decide: does each candidate's content contain information that would HELP ANSWER the query?
 
@@ -293,7 +293,7 @@ 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 +301,7 @@ 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 +385,7 @@ 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 +393,7 @@ 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 +432,20 @@ 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

@@ -74,49 +74,58 @@ 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 — two-stage retrieval (no cached embeddings).
+    // Stage 1: FTS + pattern to get candidates.
+    // Stage 2: embed candidates on-the-fly + cosine rerank.
     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") {
+      // Stage 1: cheap text retrieval
+      const hubCandidateTexts = new Map<string, string>();
       try {
         const hubFtsHits = this.store.searchHubMemories(query, { maxResults: candidatePool });
-        hubMemFtsRanked = hubFtsHits.map(({ hit }, i) => ({
-          id: `hubmem:${hit.id}`, score: 1 / (i + 1),
-        }));
+        hubMemFtsRanked = hubFtsHits.map(({ hit }, i) => {
+          hubCandidateTexts.set(hit.id, (hit.summary || hit.content || "").slice(0, 500));
+          return { id: `hubmem:${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:${h.memoryId}`, score: 1 / (i + 1),
-          }));
+          hubMemPatternRanked = hubPatternHits.map((h, i) => {
+            hubCandidateTexts.set(h.memoryId, (h.content || "").slice(0, 500));
+            return { id: `hubmem:${h.memoryId}`, score: 1 / (i + 1) };
+          });
         } catch { /* best-effort */ }
       }
-      try {
-        const hubMemEmbs = this.store.getVisibleHubMemoryEmbeddings("");
-        if (hubMemEmbs.length > 0) {
+
+      // Stage 2: embed candidates on-the-fly and cosine rerank
+      if (hubCandidateTexts.size > 0) {
+        try {
           const qv = await this.embedder.embedQuery(query).catch(() => null);
           if (qv) {
+            const ids = [...hubCandidateTexts.keys()];
+            const texts = ids.map(id => hubCandidateTexts.get(id)!);
+            const vecs = await this.embedder.embed(texts);
             const scored: Array<{ id: string; score: number }> = [];
-            for (const e of hubMemEmbs) {
+            for (let j = 0; j < ids.length; j++) {
+              if (!vecs[j]) continue;
+              const v = vecs[j];
               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];
+              for (let i = 0; i < qv.length && i < v.length; i++) {
+                dot += qv[i] * v[i]; nA += qv[i] * qv[i]; nB += v[i] * v[i];
               }
               const sim = nA > 0 && nB > 0 ? dot / (Math.sqrt(nA) * Math.sqrt(nB)) : 0;
               if (sim > 0.3) {
-                scored.push({ id: `hubmem:${e.memoryId}`, score: sim });
+                scored.push({ id: `hubmem:${ids[j]}`, score: sim });
               }
             }
             scored.sort((a, b) => b.score - a.score);
             hubMemVecRanked = scored.slice(0, candidatePool);
           }
-        }
-      } catch { /* best-effort */ }
+        } 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}`);