@vellumai/assistant 0.4.55 → 0.4.57

package/src/memory/conversation-starters-cadence.ts

@@ -0,0 +1,74 @@
+/**
+ * Cadence logic for conversation starters generation.
+ *
+ * Decides whether a new generation job should be enqueued based on how many
+ * active memory items have accumulated since the last generation.
+ */
+
+import { and, eq, inArray, like } from "drizzle-orm";
+
+import { getLogger } from "../util/logger.js";
+import { getDb } from "./db.js";
+import { enqueueMemoryJob } from "./jobs-store.js";
+import { rawGet } from "./raw-query.js";
+import { memoryCheckpoints, memoryJobs } from "./schema.js";
+
+const log = getLogger("conversation-starters-cadence");
+
+/**
+ * Check whether enough new memory items have accumulated to justify
+ * generating a fresh batch of conversation starters.
+ */
+export function maybeEnqueueConversationStartersJob(scopeId: string): void {
+  const db = getDb();
+
+  // Count total active memory items
+  const countRow = rawGet<{ c: number }>(
+    `SELECT COUNT(*) AS c FROM memory_items WHERE status = 'active' AND scope_id = ?`,
+    scopeId,
+  );
+  const totalActive = countRow?.c ?? 0;
+  if (totalActive === 0) return;
+
+  // Read checkpoint: item count at last generation (scoped so each scope tracks independently)
+  const checkpointKey = `conversation_starters:item_count_at_last_gen:${scopeId}`;
+  const checkpoint = db
+    .select({ value: memoryCheckpoints.value })
+    .from(memoryCheckpoints)
+    .where(eq(memoryCheckpoints.key, checkpointKey))
+    .get();
+  const lastCount = checkpoint ? parseInt(checkpoint.value, 10) : 0;
+
+  // Cadence formula
+  let threshold: number;
+  if (totalActive <= 10) {
+    threshold = 1;
+  } else if (totalActive <= 50) {
+    threshold = 5;
+  } else {
+    threshold = 10;
+  }
+
+  const delta = totalActive - lastCount;
+  if (delta < threshold) return;
+
+  // Dedup: don't enqueue if a pending/running job for this scope already exists
+  const existing = db
+    .select({ id: memoryJobs.id })
+    .from(memoryJobs)
+    .where(
+      and(
+        eq(memoryJobs.type, "generate_conversation_starters"),
+        inArray(memoryJobs.status, ["pending", "running"]),
+        like(memoryJobs.payload, `%"scopeId":"${scopeId}"%`),
+      ),
+    )
+    .get();
+  if (existing) return;
+
+  enqueueMemoryJob("generate_conversation_starters", { scopeId });
+  log.info(
+    { totalActive, lastCount, delta, threshold, scopeId },
+    "Enqueued conversation starters generation job",
+  );
+}

package/src/memory/conversation-title-service.ts

@@ -146,7 +146,6 @@ export async function generateAndPersistConversationTitle(
       signal: combinedSignal,
     },
   );
-
   const textBlock = response.content.find((b) => b.type === "text");
   if (textBlock && textBlock.type === "text") {
     let title = normalizeTitle(textBlock.text);
@@ -312,7 +311,7 @@ function buildTitlePrompt(
   assistantResponse?: string,
 ): string {
   const parts: string[] = [
-    "Generate a very short title for this conversation. Rules: at most 5 words, at most 40 characters, no quotes, no markdown formatting.",
+    "Generate a very short title summarizing the TOPIC of this conversation. Rules: at most 5 words, at most 40 characters, no quotes, no markdown formatting. IMPORTANT: Summarize what the user is asking about — do NOT respond to the message, do NOT assess feasibility, and do NOT comment on your own capabilities.",
   ];
 
   if (context) {
@@ -369,7 +368,7 @@ function deriveFallbackTitle(context?: TitleContext): string | null {
 
 function buildRegenerationPrompt(recentMessages: MessageRow[]): string {
   const parts: string[] = [
-    "Generate a very short title for this conversation based on the recent messages below. Rules: at most 5 words, at most 40 characters, no quotes, no markdown formatting.",
+    "Generate a very short title summarizing the TOPIC of this conversation based on the recent messages below. Rules: at most 5 words, at most 40 characters, no quotes, no markdown formatting. IMPORTANT: Summarize what the user is asking about — do NOT respond to the messages, do NOT assess feasibility, and do NOT comment on your own capabilities.",
     "",
     "Recent messages:",
   ];

package/src/memory/db-init.ts

@@ -25,6 +25,7 @@ import {
   createCoreTables,
   createExternalConversationBindingsTables,
   createFollowupsTables,
+  createLifecycleEventsTable,
   createMediaAssetsTables,
   createMessagesFts,
   createNotificationTables,
@@ -41,6 +42,7 @@ import {
   migrateCallSessionMode,
   migrateCanonicalGuardianDeliveriesDestinationIndex,
   migrateCanonicalGuardianRequesterChatId,
+  migrateCapabilityCardColumns,
   migrateChannelInboundDeliveredSegments,
   migrateChannelInteractionColumns,
   migrateContactChannelsAccessFields,
@@ -49,14 +51,18 @@ import {
   migrateContactsNotesColumn,
   migrateContactsRolePrincipal,
   migrateConversationsThreadTypeIndex,
+  migrateCreateThreadStartersTable,
+  migrateCreateTraceEventsTable,
   migrateDropAccountsTable,
   migrateDropAssistantIdColumns,
+  migrateDropCapabilityCardState,
   migrateDropConflicts,
   migrateDropContactInteractionColumns,
   migrateDropEntityTables,
   migrateDropLegacyMemberGuardianTables,
   migrateDropLoopbackPortColumn,
   migrateDropMemorySegmentFts,
+  migrateDropOrphanedMediaTables,
   migrateDropRemindersTable,
   migrateDropUsageCompositeIndexes,
   migrateFkCascadeRebuilds,
@@ -67,6 +73,7 @@ import {
   migrateGuardianDeliveryConversationIndex,
   migrateGuardianPrincipalIdColumns,
   migrateGuardianPrincipalIdNotNull,
+  migrateGuardianTimestampsEpochMs,
   migrateGuardianVerificationPurpose,
   migrateGuardianVerificationSessions,
   migrateInviteCodeHashColumn,
@@ -79,7 +86,17 @@ import {
   migrateOAuthProvidersPingUrl,
   migrateReminderRoutingIntent,
   migrateRemindersToSchedules,
+  migrateRenameConversationTypeColumn,
+  migrateRenameCreatedBySessionIdColumns,
+  migrateRenameFollowupsThreadIdColumn,
+  migrateRenameGmailProviderKeyToGoogle,
   migrateRenameGuardianVerificationValues,
+  migrateRenameInboxThreadStateTable,
+  migrateRenameNotificationThreadColumns,
+  migrateRenameSequenceEnrollmentsThreadIdColumn,
+  migrateRenameSequenceStepsReplyKey,
+  migrateRenameSourceSessionIdColumn,
+  migrateRenameThreadStartersTable,
   migrateRenameVerificationSessionIdColumn,
   migrateRenameVerificationTable,
   migrateRenameVoiceToPhone,
@@ -219,7 +236,7 @@ export function initializeDb(): void {
   // 14c3. Guardian action supersession metadata (superseded_by_request_id, superseded_at) + session lookup index
   migrateGuardianActionSupersession(database);
 
-  // 14d. Index on conversations.thread_type for frequent WHERE filters
+  // 14d. Index on conversations.conversation_type for frequent WHERE filters
   migrateConversationsThreadTypeIndex(database);
 
   // 14e. Index on guardian_action_deliveries.destination_conversation_id for conversation-based lookups
@@ -250,7 +267,7 @@ export function initializeDb(): void {
   // 22. Scoped approval grants (channel-agnostic one-time-use grants)
   createScopedApprovalGrantsTable(database);
 
-  // 23. Thread decision audit columns on notification_deliveries
+  // 23. Conversation decision audit columns on notification_deliveries
   migrateNotificationDeliveryThreadDecision(database);
 
   // 24. Canonical guardian requests and deliveries (unified cross-source guardian domain)
@@ -388,6 +405,57 @@ export function initializeDb(): void {
   // 63. Drop loopback_port column from oauth_providers (moved to code-side behavior registry)
   migrateDropLoopbackPortColumn(database);
 
+  // 64. Drop orphaned media tables (CREATE TABLE removed in #16739, clean up existing databases)
+  migrateDropOrphanedMediaTables(database);
+
+  // 65. Convert guardian timestamps from ISO 8601 text to epoch ms integers
+  migrateGuardianTimestampsEpochMs(database);
+
+  // 66. Rename assistant_inbox_thread_state → assistant_inbox_conversation_state
+  migrateRenameInboxThreadStateTable(database);
+
+  // 67. Rename thread_type → conversation_type in conversations table
+  migrateRenameConversationTypeColumn(database);
+
+  // 68. Rename notification_deliveries thread columns → conversation columns
+  migrateRenameNotificationThreadColumns(database);
+
+  // 69. Rename followups.thread_id → conversation_id
+  migrateRenameFollowupsThreadIdColumn(database);
+
+  // 70. Rename sequence_enrollments.thread_id → conversation_id
+  migrateRenameSequenceEnrollmentsThreadIdColumn(database);
+
+  // 71. Rename replyToThread → replyInSameConversation in sequence steps JSON blobs
+  migrateRenameSequenceStepsReplyKey(database);
+
+  // 72. Rename integration:gmail → integration:google across OAuth tables
+  migrateRenameGmailProviderKeyToGoogle(database);
+
+  // 73. Create thread_starters table for personalized empty-thread suggestions (renamed in migration 77)
+  migrateCreateThreadStartersTable(database);
+
+  // 74. Add capability card columns to thread_starters + category relevance table
+  migrateCapabilityCardColumns(database);
+
+  // 75. Rename created_by_session_id → source_conversation_id in verification sessions and invites
+  migrateRenameCreatedBySessionIdColumns(database);
+
+  // 76. Rename source_session_id → source_context_id in notification_events
+  migrateRenameSourceSessionIdColumn(database);
+
+  // 77. Rename thread_starters → conversation_starters table and indexes
+  migrateRenameThreadStartersTable(database);
+
+  // 78. Lifecycle events table for app_open / hatch telemetry
+  createLifecycleEventsTable(database);
+
+  // 79. Remove deleted capability-card state while keeping conversation starter chips
+  migrateDropCapabilityCardState(database);
+
+  // 80. Trace events table for persistent trace/activity storage across sessions
+  migrateCreateTraceEventsTable(database);
+
   validateMigrationState(database);
 
   if (process.env.BUN_TEST === "1") {

package/src/memory/embedding-backend.ts

@@ -2,7 +2,7 @@ import { createHash } from "node:crypto";
 
 import { getOllamaBaseUrlEnv } from "../config/env.js";
 import type { AssistantConfig } from "../config/types.js";
-import { getSecureKeyAsync } from "../security/secure-keys.js";
+import { getProviderKeyAsync } from "../security/secure-keys.js";
 import { getLogger } from "../util/logger.js";
 import { GeminiEmbeddingBackend } from "./embedding-gemini.js";
 import { OllamaEmbeddingBackend } from "./embedding-ollama.js";
@@ -203,6 +203,15 @@ function getCachedOrCreate<T extends EmbeddingBackend>(
   return instance;
 }
 
+/** Check if a backend instance already exists in the cache. */
+function getCached(
+  provider: string,
+  model: string,
+  extras?: string[],
+): EmbeddingBackend | undefined {
+  return backendCache.get(cacheKey(provider, model, extras));
+}
+
 function geminiCacheExtras(config: AssistantConfig): string[] {
   const extras: string[] = [];
   if (config.memory.embeddings.geminiTaskType) {
@@ -250,7 +259,10 @@ export async function selectEmbeddingBackend(
     };
   }
   if (requested === "ollama") {
-    const ollamaKey = (await getSecureKeyAsync("ollama")) ?? undefined;
+    // Check cache first to avoid unnecessary async key fetch on cache hits
+    const cached = getCached("ollama", config.memory.embeddings.ollamaModel);
+    if (cached) return { backend: cached, reason: null };
+    const ollamaKey = (await getProviderKeyAsync("ollama")) ?? undefined;
     return {
       backend: getCachedOrCreate(
         "ollama",
@@ -286,7 +298,13 @@ export async function selectEmbeddingBackend(
           reason: null,
         };
       case "openai": {
-        const openaiKey = await getSecureKeyAsync("openai");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOpenai = getCached(
+          "openai",
+          config.memory.embeddings.openaiModel,
+        );
+        if (cachedOpenai) return { backend: cachedOpenai, reason: null };
+        const openaiKey = await getProviderKeyAsync("openai");
         if (!openaiKey) continue;
         return {
           backend: getCachedOrCreate(
@@ -302,7 +320,14 @@ export async function selectEmbeddingBackend(
         };
       }
       case "gemini": {
-        const geminiKey = await getSecureKeyAsync("gemini");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedGemini = getCached(
+          "gemini",
+          config.memory.embeddings.geminiModel,
+          geminiCacheExtras(config),
+        );
+        if (cachedGemini) return { backend: cachedGemini, reason: null };
+        const geminiKey = await getProviderKeyAsync("gemini");
         if (!geminiKey) continue;
         return {
           backend: getCachedOrCreate(
@@ -323,8 +348,14 @@ export async function selectEmbeddingBackend(
         };
       }
       case "ollama": {
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOllama = getCached(
+          "ollama",
+          config.memory.embeddings.ollamaModel,
+        );
+        if (cachedOllama) return { backend: cachedOllama, reason: null };
         if (!(await isOllamaConfigured(config))) continue;
-        const ollamaKey = (await getSecureKeyAsync("ollama")) ?? undefined;
+        const ollamaKey = (await getProviderKeyAsync("ollama")) ?? undefined;
         return {
           backend: getCachedOrCreate(
             "ollama",
@@ -539,7 +570,16 @@ async function selectFallbackBackends(
     if (provider === exclude) continue;
     switch (provider) {
       case "openai": {
-        const openaiKey = await getSecureKeyAsync("openai");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOpenai = getCached(
+          "openai",
+          config.memory.embeddings.openaiModel,
+        );
+        if (cachedOpenai) {
+          backends.push(cachedOpenai);
+          break;
+        }
+        const openaiKey = await getProviderKeyAsync("openai");
         if (openaiKey) {
           backends.push(
             getCachedOrCreate(
@@ -556,7 +596,17 @@ async function selectFallbackBackends(
         break;
       }
       case "gemini": {
-        const geminiKey = await getSecureKeyAsync("gemini");
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedGemini = getCached(
+          "gemini",
+          config.memory.embeddings.geminiModel,
+          geminiCacheExtras(config),
+        );
+        if (cachedGemini) {
+          backends.push(cachedGemini);
+          break;
+        }
+        const geminiKey = await getProviderKeyAsync("gemini");
         if (geminiKey) {
           backends.push(
             getCachedOrCreate(
@@ -577,9 +627,18 @@ async function selectFallbackBackends(
         }
         break;
       }
-      case "ollama":
+      case "ollama": {
+        // Check cache first to avoid unnecessary async key fetch on cache hits
+        const cachedOllama = getCached(
+          "ollama",
+          config.memory.embeddings.ollamaModel,
+        );
+        if (cachedOllama) {
+          backends.push(cachedOllama);
+          break;
+        }
         if (await isOllamaConfigured(config)) {
-          const ollamaKey = (await getSecureKeyAsync("ollama")) ?? undefined;
+          const ollamaKey = (await getProviderKeyAsync("ollama")) ?? undefined;
           backends.push(
             getCachedOrCreate(
               "ollama",
@@ -595,6 +654,7 @@ async function selectFallbackBackends(
           );
         }
         break;
+      }
     }
   }
   return backends;
@@ -626,8 +686,8 @@ export async function selectedBackendSupportsMultimodal(
 
 async function isOllamaConfigured(config: AssistantConfig): Promise<boolean> {
   return (
-    config.provider === "ollama" ||
-    Boolean(await getSecureKeyAsync("ollama")) ||
+    config.services.inference.provider === "ollama" ||
+    Boolean(await getProviderKeyAsync("ollama")) ||
     Boolean(getOllamaBaseUrlEnv())
   );
 }

package/src/memory/external-conversation-store.ts

@@ -178,7 +178,7 @@ export function getBindingByChannelChat(
 
 /**
  * Remove an external binding by channel + external chat ID.
- * Used when disconnecting a synced thread by its channel identifiers.
+ * Used when disconnecting a synced conversation by its channel identifiers.
  */
 export function deleteBindingByChannelChat(
   sourceChannel: string,

package/src/memory/guardian-action-store.ts

@@ -724,7 +724,7 @@ export function getPendingDeliveriesByDestination(
 
 /**
  * Look up all pending deliveries by destination conversation ID.
- * Used for disambiguation when a reused vellum thread has multiple active
+ * Used for disambiguation when a reused vellum conversation has multiple active
  * guardian requests.
  */
 export function getPendingDeliveriesByConversation(
@@ -803,7 +803,7 @@ export function getExpiredDeliveriesByDestination(
 
 /**
  * Look up all expired deliveries by destination conversation ID.
- * Used for disambiguation when a reused vellum thread has multiple expired
+ * Used for disambiguation when a reused vellum conversation has multiple expired
  * guardian requests eligible for follow-up.
  */
 export function getExpiredDeliveriesByConversation(
@@ -885,7 +885,7 @@ export function getFollowupDeliveriesByDestination(
 /**
  * Look up all deliveries for requests in `awaiting_guardian_choice` follow-up
  * state by destination conversation ID. Used for disambiguation when a reused
- * vellum thread has multiple follow-up guardian requests.
+ * vellum conversation has multiple follow-up guardian requests.
  */
 export function getFollowupDeliveriesByConversation(
   conversationId: string,

package/src/memory/indexer.ts

@@ -31,6 +31,8 @@ export interface IndexMessageInput {
    * profile mutations are triggered.
    */
   provenanceTrustClass?: TrustClass;
+  /** When true, the message was auto-sent by the client (e.g. wake-up greeting) and should not trigger memory extraction. */
+  automated?: boolean;
 }
 
 export interface IndexMessageResult {
@@ -68,13 +70,18 @@ export async function indexMessageNow(
   const shouldExtract =
     input.role === "user" ||
     (input.role === "assistant" && config.extraction.extractFromAssistant);
-  // Check if the resolved embedding backend supports multimodal input.
-  // Only enqueue embed_attachment jobs when it does (currently Gemini only).
-  const supportsMultimodal =
-    await selectedBackendSupportsMultimodal(getConfig());
-  const mediaBlocks = supportsMultimodal
-    ? extractMediaBlocks(input.content).filter((b) => b.type === "image")
-    : [];
+  // Check if the message has any image blocks before probing the backend.
+  // extractMediaBlocks is synchronous and cheap, while
+  // selectedBackendSupportsMultimodal requires async key resolution that
+  // would add unnecessary latency for text-only messages.
+  const candidateMediaBlocks = extractMediaBlocks(input.content).filter(
+    (b) => b.type === "image",
+  );
+  const mediaBlocks =
+    candidateMediaBlocks.length > 0 &&
+    (await selectedBackendSupportsMultimodal(getConfig()))
+      ? candidateMediaBlocks
+      : [];
 
   // Wrap all segment inserts and job enqueues in a single transaction so they
   // either all succeed or all roll back, preventing partial/orphaned state.
@@ -135,7 +142,7 @@ export async function indexMessageNow(
       );
     }
 
-    if (shouldExtract && isTrustedActor) {
+    if (shouldExtract && isTrustedActor && !input.automated) {
       enqueueMemoryJob(
         "extract_items",
         { messageId: input.messageId, scopeId: input.scopeId ?? "default" },
@@ -163,7 +170,11 @@ export async function indexMessageNow(
     );
   }
 
-  const extractionGated = !isTrustedActor;
+  if (input.automated && shouldExtract) {
+    log.info("Skipping extraction jobs for automated message");
+  }
+
+  const extractionGated = !isTrustedActor || !!input.automated;
   const enqueuedJobs =
     segments.length -
     skippedEmbedJobs +

package/src/memory/invite-store.ts

@@ -23,7 +23,7 @@ export interface IngressInvite {
   id: string;
   sourceChannel: string;
   tokenHash: string;
-  createdBySessionId: string | null;
+  sourceConversationId: string | null;
   note: string | null;
   maxUses: number;
   useCount: number;
@@ -73,7 +73,7 @@ function rowToInvite(
     id: row.id,
     sourceChannel: row.sourceChannel,
     tokenHash: row.tokenHash,
-    createdBySessionId: row.createdBySessionId,
+    sourceConversationId: row.sourceConversationId,
     note: row.note,
     maxUses: row.maxUses,
     useCount: row.useCount,
@@ -101,7 +101,7 @@ function rowToInvite(
 export function createInvite(params: {
   sourceChannel: string;
   contactId: string;
-  createdBySessionId?: string;
+  sourceConversationId?: string;
   note?: string;
   maxUses?: number;
   expiresInMs?: number;
@@ -124,7 +124,7 @@ export function createInvite(params: {
     id,
     sourceChannel: params.sourceChannel,
     tokenHash: tokenH,
-    createdBySessionId: params.createdBySessionId ?? null,
+    sourceConversationId: params.sourceConversationId ?? null,
     note: params.note ?? null,
     maxUses: params.maxUses ?? 1,
     useCount: 0,

package/src/memory/items-extractor.ts

@@ -3,6 +3,7 @@ import { v4 as uuid } from "uuid";
 
 import { getConfig } from "../config/loader.js";
 import type { MemoryExtractionConfig } from "../config/types.js";
+import { buildCoreIdentityContext } from "../prompts/system-prompt.js";
 import {
   createTimeout,
   extractToolUse,
@@ -11,6 +12,7 @@ import {
 } from "../providers/provider-send-message.js";
 import { getLogger } from "../util/logger.js";
 import { truncate } from "../util/truncate.js";
+import { maybeEnqueueConversationStartersJob } from "./conversation-starters-cadence.js";
 import { getDb } from "./db.js";
 import { computeMemoryFingerprint } from "./fingerprint.js";
 import { enqueueMemoryJob } from "./jobs-store.js";
@@ -140,8 +142,18 @@ function buildExtractionSystemPrompt(
     subject: string;
     statement: string;
   }>,
+  messageRole: string,
 ): string {
-  let prompt = `You are a memory extraction system. Given a message from a conversation, extract structured memory items that would be valuable to remember for future interactions.
+  // Inject identity context so extracted memories use real names instead of
+  // generic "User ..." labels.
+  const identityContext = buildCoreIdentityContext();
+
+  let prompt = "";
+  if (identityContext) {
+    prompt += `# Identity Context\n\n${identityContext}\n\n---\n\n`;
+  }
+
+  prompt += `You are a memory extraction system. Given a message from a conversation, extract structured memory items that would be valuable to remember for future interactions.
 
 Extract items in these categories:
 - identity: Personal info (name, role, location, timezone, background), notable facts, relationships between people/teams/systems
@@ -174,6 +186,12 @@ Rules:
 - Prefer fewer high-quality items over many low-quality ones.
 - If the message contains no memorable information, return an empty array.`;
 
+  if (messageRole === "assistant") {
+    prompt += `
+
+IMPORTANT: The message below is from the ASSISTANT, not the user. Do NOT attribute the assistant's own statements, feelings, self-descriptions, or introspection to the user. Only extract facts about the user, the world, or the project that the assistant is referencing or relaying — NOT the assistant's own identity, uncertainty, or behavior. If the assistant is simply talking about itself (e.g., introducing itself, expressing uncertainty about its own purpose), extract nothing.`;
+  }
+
   if (existingItems.length > 0) {
     prompt += `\n\nExisting memory items (use these to identify supersession targets — set \`supersedes\` to the item ID if the new information replaces one of these):\n`;
     for (const item of existingItems) {
@@ -268,6 +286,7 @@ async function extractItemsWithLLM(
   text: string,
   extractionConfig: MemoryExtractionConfig,
   scopeId: string,
+  messageRole: string,
 ): Promise<ExtractedItem[]> {
   const provider = await getConfiguredProvider();
   if (!provider) {
@@ -283,10 +302,17 @@ async function extractItemsWithLLM(
     try {
       // Query existing items to give the LLM supersession context
       const existingItems = queryExistingItemsForContext(scopeId, text);
-      const systemPrompt = buildExtractionSystemPrompt(existingItems);
+      const systemPrompt = buildExtractionSystemPrompt(
+        existingItems,
+        messageRole,
+      );
 
+      const messagePrefix =
+        messageRole === "assistant"
+          ? "[This message is from the assistant]\n\n"
+          : "";
       const response = await provider.sendMessage(
-        [userMessage(text)],
+        [userMessage(`${messagePrefix}${text}`)],
         [
           {
             name: "store_memory_items",
@@ -478,7 +504,12 @@ export async function extractAndUpsertMemoryItemsForMessage(
   const extractionConfig = config.memory.extraction;
   const effectiveScopeId = scopeId ?? "default";
   const extracted = extractionConfig.useLLM
-    ? await extractItemsWithLLM(text, extractionConfig, effectiveScopeId)
+    ? await extractItemsWithLLM(
+        text,
+        extractionConfig,
+        effectiveScopeId,
+        message.role,
+      )
     : extractItemsPatternBased(text, effectiveScopeId);
 
   if (extracted.length === 0) return 0;
@@ -683,6 +714,19 @@ export async function extractAndUpsertMemoryItemsForMessage(
     { messageId, extracted: extracted.length, upserted },
     "Extracted memory items from message",
   );
+
+  // Trigger conversation starters generation when new items are upserted
+  if (upserted > 0) {
+    try {
+      maybeEnqueueConversationStartersJob(effectiveScopeId);
+    } catch (err) {
+      log.warn(
+        { err: err instanceof Error ? err.message : String(err) },
+        "Failed to check conversation starters cadence",
+      );
+    }
+  }
+
   return upserted;
 }
 

package/src/memory/job-handlers/backfill.ts

@@ -19,16 +19,22 @@ import { messages } from "../schema.js";
 const BACKFILL_CHECKPOINT_KEY = "memory:backfill:last_created_at";
 const BACKFILL_CHECKPOINT_ID_KEY = "memory:backfill:last_message_id";
 
-function parseProvenanceTrustClass(
-  rawMetadata: string | null,
-): TrustClass | undefined {
-  if (!rawMetadata) return undefined;
+function parseMessageMetadata(rawMetadata: string | null): {
+  provenanceTrustClass: TrustClass | undefined;
+  automated: boolean | undefined;
+} {
+  if (!rawMetadata)
+    return { provenanceTrustClass: undefined, automated: undefined };
   try {
     const parsed = messageMetadataSchema.safeParse(JSON.parse(rawMetadata));
-    if (!parsed.success) return undefined;
-    return parsed.data.provenanceTrustClass;
+    if (!parsed.success)
+      return { provenanceTrustClass: undefined, automated: undefined };
+    return {
+      provenanceTrustClass: parsed.data.provenanceTrustClass,
+      automated: parsed.data.automated,
+    };
   } catch {
-    return undefined;
+    return { provenanceTrustClass: undefined, automated: undefined };
   }
 }
 
@@ -73,7 +79,7 @@ export async function backfillJob(
         scopeId = getConversationMemoryScopeId(message.conversationId);
         scopeCache.set(message.conversationId, scopeId);
       }
-      const provenanceTrustClass = parseProvenanceTrustClass(
+      const { provenanceTrustClass, automated } = parseMessageMetadata(
         message.metadata ?? null,
       );
       await indexMessageNow(
@@ -85,6 +91,7 @@ export async function backfillJob(
           createdAt: message.createdAt,
           scopeId,
           provenanceTrustClass,
+          automated,
         },
         config.memory,
       );