n8n-nodes-tembory 1.0.12 → 1.0.13

package/README.md

@@ -2,7 +2,7 @@
 
 Node de memoria operacional da Tembory para agentes de IA no n8n.
 
-Versao atual: `1.0.12`.
+Versao atual: `1.0.13`.
 
 O Tembory entrega contexto rico para o AI Agent sem depender apenas do historico textual da conversa. Ele combina memoria semantica, working memory, decision state, fatos estaveis do lead, historico de tools, estado operacional, action ledger, timeline de entidades, compressao de memoria, grafo, mensagens recentes e diagnosticos.
 
@@ -10,6 +10,8 @@ Nos presets de producao, o contexto e organizado como um pacote acionavel: secoe
 
 Quando um modelo barato esta conectado ao Tembory, os presets de producao usam esse SLM para gerar um resumo ativo do contexto. Em modo `auto`, ele organiza tanto o estado atual (`working_memory`, `tool_history`, `operational_state`, `decision_state`) quanto memorias vetoriais recuperadas. Assim o primeiro turno ja pode receber highlights uteis mesmo antes de haver resultados do banco vetorial.
 
+O resumo do SLM usa limites conservadores e cache curto por input para evitar chamadas duplicadas quando o AI Agent carrega a memoria mais de uma vez no mesmo turno.
+
 ## Smoke tecnico
 
 Antes de publicar uma versao, rode:

package/dist/nodes/Mem0/Mem0Memory.node.js

@@ -862,13 +862,15 @@ const getMemoryStore = (ctx) => {
         data.tembory.workingMemory = data.tembory.workingMemory || {};
         data.tembory.decisionState = data.tembory.decisionState || {};
         data.tembory.memoryCompression = data.tembory.memoryCompression || {};
+        data.tembory.connectedModelSummaryCache = data.tembory.connectedModelSummaryCache || {};
         return data.tembory;
     }
     catch {
-        global.__temboryMemory = global.__temboryMemory || { toolHistory: {}, recentMessages: {}, profileFacts: {}, workingMemory: {}, decisionState: {}, memoryCompression: {} };
+        global.__temboryMemory = global.__temboryMemory || { toolHistory: {}, recentMessages: {}, profileFacts: {}, workingMemory: {}, decisionState: {}, memoryCompression: {}, connectedModelSummaryCache: {} };
         global.__temboryMemory.workingMemory = global.__temboryMemory.workingMemory || {};
         global.__temboryMemory.decisionState = global.__temboryMemory.decisionState || {};
         global.__temboryMemory.memoryCompression = global.__temboryMemory.memoryCompression || {};
+        global.__temboryMemory.connectedModelSummaryCache = global.__temboryMemory.connectedModelSummaryCache || {};
         return global.__temboryMemory;
     }
 };
@@ -998,8 +1000,8 @@ const applyOperationalPreset = (advanced = {}) => {
             recentMessagesLastN: 2,
             vectorMemoryMaxChars: 220,
             contextMaxChars: 6000,
-            connectedModelSummaryMaxChars: 700,
-            connectedModelSummaryInputMaxChars: 2200,
+            connectedModelSummaryMaxChars: 450,
+            connectedModelSummaryInputMaxChars: 1500,
         },
         audit: {
             summarySource: 'auto',
@@ -1608,7 +1610,7 @@ const buildConnectedModelSummaryInput = ({ query, profileFacts, workingMemory, d
     const hasVectorSignal = includeVectors && Array.isArray(payload.vector_memories) && payload.vector_memories.length > 0;
     if (!hasActiveSignal && !hasVectorSignal)
         return '';
-    return truncate(safeStringify(payload), Number(adv.connectedModelSummaryInputMaxChars || 3000));
+    return truncate(safeStringify(payload), Number(adv.connectedModelSummaryInputMaxChars || 2000));
 };
 const invokeConnectedModelSummary = async (connectedLanguageModel, summaryInput, adv = {}) => {
     if (!connectedLanguageModel || typeof connectedLanguageModel.invoke !== 'function' || !summaryInput)
@@ -1616,10 +1618,10 @@ const invokeConnectedModelSummary = async (connectedLanguageModel, summaryInput,
     const response = await connectedLanguageModel.invoke([
         toBaseMessage({
             role: 'user',
-            content: `Organize the Tembory memory context for the next agent turn. Return concise Portuguese bullets only. Preserve IDs, dates, tool names, next action, do-not-repeat instructions, and contradictions. Do not invent facts.\n\nContext:\n${summaryInput}`,
+                            content: `Organize the Tembory memory context for the next agent turn. Return only 3 to 6 short Portuguese bullets. No JSON, no markdown table, no field-by-field dump. Preserve important IDs, dates, tool names, next action, do-not-repeat instructions, and contradictions. Do not invent facts.\n\nContext:\n${summaryInput}`,
         }),
     ]);
-    return cleanModelSummaryText(response, Number(adv.connectedModelSummaryMaxChars || 900));
+    return cleanModelSummaryText(response, Number(adv.connectedModelSummaryMaxChars || 700));
 };
 const contextSizeOfMessages = (messages = []) => {
     const perMessage = (messages || []).map((message, index) => {
@@ -2784,7 +2786,26 @@ class Mem0Memory {
                     highlights,
                     adv,
                 });
-                connectedModelSummary = await invokeConnectedModelSummary(connectedLanguageModel, summaryInput, adv);
+                const cacheKey = summaryInput ? stableHash({
+                    key,
+                    source: adv.summarySource || 'auto',
+                    input: summaryInput,
+                    max: adv.connectedModelSummaryMaxChars || 700,
+                }) : '';
+                const cached = cacheKey ? store.connectedModelSummaryCache[cacheKey] : null;
+                if (cached && cached.summary && Date.now() - Number(cached.at || 0) < Number(adv.connectedModelSummaryCacheTTLSeconds || 60) * 1000) {
+                    connectedModelSummary = cached.summary;
+                    connectedAi.languageModelSummaryCached = true;
+                }
+                else {
+                    connectedModelSummary = await invokeConnectedModelSummary(connectedLanguageModel, summaryInput, adv);
+                    if (cacheKey && connectedModelSummary) {
+                        store.connectedModelSummaryCache[cacheKey] = { summary: connectedModelSummary, at: Date.now() };
+                        const keys = Object.keys(store.connectedModelSummaryCache);
+                        for (const oldKey of keys.slice(0, Math.max(0, keys.length - Number(adv.connectedModelSummaryCacheMaxItems || 50))))
+                            delete store.connectedModelSummaryCache[oldKey];
+                    }
+                }
                 connectedAi.languageModelSummary = Boolean(connectedModelSummary);
             }
             catch (error) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "n8n-nodes-tembory",
-  "version": "1.0.12",
+  "version": "1.0.13",
   "description": "Tembory node for n8n AI Agents with profile, tools, timeline, graph and semantic memory",
   "license": "MIT",
   "homepage": "https://tembory.com",