@vellumai/assistant 0.5.5 → 0.5.7

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

@@ -133,6 +133,8 @@ export async function generateAndPersistConversationTitle(
   const result = await runBtwSidechain({
     content: prompt,
     provider,
+    systemPrompt: buildTitleSystemPrompt(),
+    tools: [],
     maxTokens: config.daemon.titleGenerationMaxTokens,
     modelIntent: "latency-optimized",
     signal,
@@ -236,6 +238,8 @@ export async function regenerateConversationTitle(
   const result = await runBtwSidechain({
     content: prompt,
     provider,
+    systemPrompt: buildTitleSystemPrompt(),
+    tools: [],
     maxTokens: config.daemon.titleGenerationMaxTokens,
     modelIntent: "latency-optimized",
     signal,
@@ -277,14 +281,30 @@ export function queueRegenerateConversationTitle(
 
 // ── Internal helpers ─────────────────────────────────────────────────
 
+/**
+ * Dedicated system prompt for title generation. Replaces the default
+ * assistant system prompt that btw-sidechain would otherwise inject,
+ * which caused the model to respond to the conversation content instead
+ * of titling it.
+ */
+function buildTitleSystemPrompt(): string {
+  return [
+    "You generate short conversation titles. Output ONLY the title text — no explanation, no quotes, no markdown, no preamble.",
+    "",
+    "Rules:",
+    "- Maximum 5 words and 40 characters",
+    "- Summarize the TOPIC the user is asking about",
+    "- Do NOT respond to the conversation content",
+    "- Do NOT assess feasibility or comment on capabilities",
+  ].join("\n");
+}
+
 function buildTitlePrompt(
   context?: TitleContext,
   userMessage?: string,
   assistantResponse?: string,
 ): string {
-  const parts: string[] = [
-    "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.",
-  ];
+  const parts: string[] = [];
 
   if (context) {
     const hints: string[] = [];
@@ -295,12 +315,12 @@ function buildTitlePrompt(
     if (context.metadataHints?.length)
       hints.push(`Hints: ${context.metadataHints.join(", ")}`);
     if (hints.length > 0) {
-      parts.push("", "Metadata:", ...hints);
+      parts.push("Metadata:", ...hints, "");
     }
   }
 
   if (userMessage) {
-    parts.push("", `User: ${truncate(userMessage, 200, "")}`);
+    parts.push(`User: ${truncate(userMessage, 200, "")}`);
   }
   if (assistantResponse) {
     parts.push(`Assistant: ${truncate(assistantResponse, 200, "")}`);
@@ -338,16 +358,61 @@ function deriveFallbackTitle(context?: TitleContext): string | null {
   return null;
 }
 
+/**
+ * Extract only human-authored text from stored message content for title
+ * generation. Unlike extractTextFromStoredMessageContent (which includes
+ * tool metadata like "Tool use (...): {...}"), this only extracts:
+ * - `text` blocks (the actual conversation content)
+ * - `tool_result` string content (topical signal from tool responses)
+ *   — web_search_tool_result is skipped (structured search data, not topical)
+ *
+ * Returns empty string for content-block arrays with no extractable text,
+ * preventing raw JSON from polluting the title prompt.
+ */
+function extractTextForTitle(raw: string): string {
+  try {
+    const parsed = JSON.parse(raw);
+    if (typeof parsed === "string") return parsed;
+    if (!Array.isArray(parsed)) return raw;
+    const texts: string[] = [];
+    for (const block of parsed) {
+      if (!block || typeof block !== "object") continue;
+      if (block.type === "text" && typeof block.text === "string") {
+        texts.push(block.text);
+        // guard:allow-tool-result-only — web_search_tool_result has structured
+        // search result arrays, not useful for title generation; only plain
+        // tool_result string content carries topical signal.
+      } else if (block.type === "tool_result") {
+        if (typeof block.content === "string") {
+          texts.push(block.content);
+        } else if (Array.isArray(block.content)) {
+          for (const nested of block.content) {
+            if (
+              nested &&
+              typeof nested === "object" &&
+              nested.type === "text" &&
+              typeof nested.text === "string"
+            ) {
+              texts.push(nested.text);
+            }
+          }
+        }
+      }
+    }
+    return texts.join("\n");
+  } catch {
+    return raw;
+  }
+}
+
 function buildRegenerationPrompt(recentMessages: MessageRow[]): string {
-  const parts: string[] = [
-    "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:",
-  ];
+  const parts: string[] = ["Recent messages:"];
 
   for (const msg of recentMessages) {
+    const text = extractTextForTitle(msg.content);
+    if (!text) continue;
     const role = msg.role === "user" ? "User" : "Assistant";
-    parts.push(`${role}: ${truncate(msg.content, 200, "")}`);
+    parts.push(`${role}: ${truncate(text, 200, "")}`);
   }
 
   return parts.join("\n");

package/src/memory/db-init.ts

@@ -35,12 +35,14 @@ import {
   createTasksAndWorkItemsTables,
   createWatchersAndLogsTables,
   migrateAssistantContactMetadata,
+  migrateBackfillAudioAttachmentMimeTypes,
   migrateBackfillContactInteractionStats,
   migrateBackfillGuardianPrincipalId,
   migrateBackfillInlineAttachmentsToDisk,
   migrateBackfillUsageCacheAccounting,
   migrateCallSessionInviteMetadata,
   migrateCallSessionMode,
+  migrateCallSessionSkipDisclosure,
   migrateCanonicalGuardianDeliveriesDestinationIndex,
   migrateCanonicalGuardianRequesterChatId,
   migrateCapabilityCardColumns,
@@ -51,6 +53,7 @@ import {
   migrateContactsAssistantId,
   migrateContactsNotesColumn,
   migrateContactsRolePrincipal,
+  migrateContactsUserFileColumn,
   migrateConversationForkLineage,
   migrateConversationsThreadTypeIndex,
   migrateCreateThreadStartersTable,
@@ -66,6 +69,7 @@ import {
   migrateDropMemorySegmentFts,
   migrateDropOrphanedMediaTables,
   migrateDropRemindersTable,
+  migrateDropSimplifiedMemory,
   migrateDropUsageCompositeIndexes,
   migrateFkCascadeRebuilds,
   migrateGuardianActionFollowup,
@@ -82,10 +86,7 @@ import {
   migrateInviteContactId,
   migrateLlmRequestLogMessageId,
   migrateLlmRequestLogProvider,
-  migrateMemoryArchiveTables,
-  migrateMemoryBriefState,
   migrateMemoryItemSupersession,
-  migrateMemoryReducerCheckpoints,
   migrateMessagesFtsBackfill,
   migrateNormalizePhoneIdentities,
   migrateNotificationDeliveryThreadDecision,
@@ -488,17 +489,20 @@ export function initializeDb(): void {
   // 84. Add nullable conversation fork lineage columns and parent lookup index
   migrateConversationForkLineage(database);
 
-  // 85. Memory brief state tables (time_contexts, open_loops) for simplified memory system
-  migrateMemoryBriefState(database);
+  // 85. Add quiet flag to schedule jobs
+  migrateScheduleQuietFlag(database);
 
-  // 86. Memory archive tables (observations, chunks, episodes) for simplified memory v1
-  migrateMemoryArchiveTables(database);
+  // 86. Drop simplified-memory tables and reducer checkpoint columns
+  migrateDropSimplifiedMemory(database);
 
-  // 87. Add memory reducer checkpoint columns to conversations
-  migrateMemoryReducerCheckpoints(database);
+  // 87. Add skip_disclosure column to call_sessions for per-call disclosure control
+  migrateCallSessionSkipDisclosure(database);
 
-  // 88. Add quiet flag to schedule jobs
-  migrateScheduleQuietFlag(database);
+  // 88. Backfill correct MIME types for audio attachments stored as application/octet-stream
+  migrateBackfillAudioAttachmentMimeTypes(database);
+
+  // 89. Add user_file column to contacts for per-user persona file mapping
+  migrateContactsUserFileColumn(database);
 
   validateMigrationState(database);
 

package/src/memory/indexer.ts

@@ -5,7 +5,6 @@ import { getConfig } from "../config/loader.js";
 import type { MemoryConfig } from "../config/types.js";
 import type { TrustClass } from "../runtime/actor-trust-resolver.js";
 import { getLogger } from "../util/logger.js";
-import { computeChunkContentHash } from "./archive-store.js";
 import { getDb } from "./db.js";
 import { selectedBackendSupportsMultimodal } from "./embedding-backend.js";
 import { enqueueMemoryJob } from "./jobs-store.js";
@@ -13,7 +12,7 @@ import {
   extractMediaBlockMeta,
   extractTextFromStoredMessageContent,
 } from "./message-content.js";
-import { memoryChunks, memoryObservations, memorySegments } from "./schema.js";
+import { memorySegments } from "./schema.js";
 import { segmentText } from "./segmenter.js";
 
 const log = getLogger("memory-indexer");
@@ -54,12 +53,7 @@ export async function indexMessageNow(
     input.provenanceTrustClass === undefined;
 
   const text = extractTextFromStoredMessageContent(input.content);
-  const hasText = text.length > 0;
-  const candidateMediaMeta = extractMediaBlockMeta(input.content).filter(
-    (b) => b.type === "image",
-  );
-  const hasMedia = candidateMediaMeta.length > 0;
-  if (!hasText && !hasMedia) {
+  if (text.length === 0) {
     enqueueMemoryJob("build_conversation_summary", {
       conversationId: input.conversationId,
     });
@@ -68,13 +62,11 @@ export async function indexMessageNow(
 
   const db = getDb();
   const now = Date.now();
-  const segments = hasText
-    ? segmentText(
-        text,
-        config.segmentation.targetTokens,
-        config.segmentation.overlapTokens,
-      )
-    : [];
+  const segments = segmentText(
+    text,
+    config.segmentation.targetTokens,
+    config.segmentation.overlapTokens,
+  );
   const shouldExtract =
     input.role === "user" ||
     (input.role === "assistant" && config.extraction.extractFromAssistant);
@@ -84,6 +76,9 @@ export async function indexMessageNow(
   // overhead for messages on non-multimodal backends.
   // selectedBackendSupportsMultimodal requires async key resolution, so we
   // skip it entirely for text-only messages.
+  const candidateMediaMeta = extractMediaBlockMeta(input.content).filter(
+    (b) => b.type === "image",
+  );
   const mediaBlocks =
     candidateMediaMeta.length > 0 &&
     (await selectedBackendSupportsMultimodal(getConfig()))
@@ -93,10 +88,7 @@ export async function indexMessageNow(
   // Wrap all segment inserts and job enqueues in a single transaction so they
   // either all succeed or all roll back, preventing partial/orphaned state.
   let skippedEmbedJobs = 0;
-  let skippedChunkEmbedJobs = 0;
-  const scopeId = input.scopeId ?? "default";
   db.transaction((tx) => {
-    // ── Legacy segment path (kept intact for parallel validation) ───
     for (const segment of segments) {
       const segmentId = buildSegmentId(input.messageId, segment.segmentIndex);
       const hash = createHash("sha256").update(segment.text).digest("hex");
@@ -117,7 +109,7 @@ export async function indexMessageNow(
           segmentIndex: segment.segmentIndex,
           text: segment.text,
           tokenEstimate: segment.tokenEstimate,
-          scopeId,
+          scopeId: input.scopeId ?? "default",
           contentHash: hash,
           createdAt: input.createdAt,
           updatedAt: now,
@@ -127,7 +119,7 @@ export async function indexMessageNow(
           set: {
             text: segment.text,
             tokenEstimate: segment.tokenEstimate,
-            scopeId,
+            scopeId: input.scopeId ?? "default",
             contentHash: hash,
             updatedAt: now,
           },
@@ -141,65 +133,6 @@ export async function indexMessageNow(
       }
     }
 
-    // ── Archive chunk dual-write (mirrors segment boundaries) ──────
-    // Create a single observation per message, then create one chunk per
-    // segment using the same segmentation boundaries. Chunks are
-    // deduplicated by (scopeId, contentHash) via onConflictDoNothing so
-    // unchanged content does not enqueue duplicate embed_chunk jobs.
-    const observationId = buildObservationId(input.messageId);
-    tx.insert(memoryObservations)
-      .values({
-        id: observationId,
-        scopeId,
-        conversationId: input.conversationId,
-        messageId: input.messageId,
-        role: input.role,
-        content: hasText ? text : input.content,
-        modality: hasMedia ? "multimodal" : "text",
-        source: null,
-        createdAt: input.createdAt,
-      })
-      .onConflictDoNothing({ target: memoryObservations.id })
-      .run();
-
-    for (const segment of segments) {
-      const chunkId = buildChunkId(input.messageId, segment.segmentIndex);
-      const chunkHash = computeChunkContentHash(scopeId, segment.text);
-
-      // Check if this chunk already exists with the same content hash
-      const existingChunk = tx
-        .select({ contentHash: memoryChunks.contentHash })
-        .from(memoryChunks)
-        .where(eq(memoryChunks.id, chunkId))
-        .get();
-
-      tx.insert(memoryChunks)
-        .values({
-          id: chunkId,
-          scopeId,
-          observationId,
-          content: segment.text,
-          tokenEstimate: segment.tokenEstimate,
-          contentHash: chunkHash,
-          createdAt: input.createdAt,
-        })
-        .onConflictDoUpdate({
-          target: memoryChunks.id,
-          set: {
-            content: segment.text,
-            tokenEstimate: segment.tokenEstimate,
-            contentHash: chunkHash,
-          },
-        })
-        .run();
-
-      if (existingChunk?.contentHash === chunkHash) {
-        skippedChunkEmbedJobs++;
-      } else {
-        enqueueMemoryJob("embed_chunk", { chunkId, scopeId }, Date.now(), tx);
-      }
-    }
-
     // Enqueue embed_attachment jobs for image content blocks when the
     // embedding provider supports multimodal (Gemini only).
     for (const block of mediaBlocks) {
@@ -214,7 +147,7 @@ export async function indexMessageNow(
     if (shouldExtract && isTrustedActor && !input.automated) {
       enqueueMemoryJob(
         "extract_items",
-        { messageId: input.messageId, scopeId },
+        { messageId: input.messageId, scopeId: input.scopeId ?? "default" },
         Date.now(),
         tx,
       );
@@ -233,12 +166,6 @@ export async function indexMessageNow(
     );
   }
 
-  if (skippedChunkEmbedJobs > 0) {
-    log.debug(
-      `Skipped ${skippedChunkEmbedJobs}/${segments.length} embed_chunk jobs (content unchanged)`,
-    );
-  }
-
   if (!isTrustedActor && shouldExtract) {
     log.info(
       `Skipping extraction jobs for untrusted actor (trustClass=${input.provenanceTrustClass})`,
@@ -250,11 +177,9 @@ export async function indexMessageNow(
   }
 
   const extractionGated = !isTrustedActor || !!input.automated;
-  const segmentEmbedJobs = segments.length - skippedEmbedJobs;
-  const chunkEmbedJobs = segments.length - skippedChunkEmbedJobs;
   const enqueuedJobs =
-    segmentEmbedJobs +
-    chunkEmbedJobs +
+    segments.length -
+    skippedEmbedJobs +
     mediaBlocks.length +
     (shouldExtract && !extractionGated ? 2 : 1);
   return {
@@ -288,19 +213,3 @@ export function getRecentSegmentsForConversation(
 function buildSegmentId(messageId: string, segmentIndex: number): string {
   return `${messageId}:${segmentIndex}`;
 }
-
-/**
- * Deterministic observation ID derived from the messageId so repeated
- * indexer runs for the same message converge on the same observation row.
- */
-function buildObservationId(messageId: string): string {
-  return `obs:${messageId}`;
-}
-
-/**
- * Deterministic chunk ID derived from the messageId and segment index so
- * the dual-write path mirrors the legacy segment identity scheme exactly.
- */
-function buildChunkId(messageId: string, segmentIndex: number): string {
-  return `chunk:${messageId}:${segmentIndex}`;
-}

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 { resolveGuardianPersona } from "../prompts/persona-resolver.js";
 import { buildCoreIdentityContext } from "../prompts/system-prompt.js";
 import {
   createTimeout,
@@ -150,6 +151,7 @@ function buildExtractionSystemPrompt(
     statement: string;
   }>,
   messageRole: string,
+  userPersona?: string | null,
 ): string {
   // Build the fixed instruction body first so we can measure it and allocate
   // the remaining budget to identity context.
@@ -206,7 +208,9 @@ IMPORTANT: The message below is from the ASSISTANT, not the user. Do NOT attribu
   // ceiling, preventing oversized prompts from exceeding the provider input
   // window (which would cause sendMessage to error and fall back to
   // lower-quality pattern-based extraction).
-  const rawIdentityContext = buildCoreIdentityContext();
+  const rawIdentityContext = buildCoreIdentityContext(
+    userPersona ? { userPersona } : undefined,
+  );
 
   let prompt = "";
   if (rawIdentityContext) {
@@ -316,6 +320,7 @@ async function extractItemsWithLLM(
   extractionConfig: MemoryExtractionConfig,
   scopeId: string,
   messageRole: string,
+  userPersona?: string | null,
 ): Promise<ExtractedItem[]> {
   const provider = await getConfiguredProvider();
   if (!provider) {
@@ -334,6 +339,7 @@ async function extractItemsWithLLM(
       const systemPrompt = buildExtractionSystemPrompt(
         existingItems,
         messageRole,
+        userPersona,
       );
 
       const messagePrefix =
@@ -532,12 +538,20 @@ export async function extractAndUpsertMemoryItemsForMessage(
   const config = getConfig();
   const extractionConfig = config.memory.extraction;
   const effectiveScopeId = scopeId ?? "default";
+
+  // Resolve the guardian's persona to provide personality-aware extraction
+  // context. Currently uses the guardian persona for all conversations —
+  // non-guardian conversations are rare and the guardian's persona provides
+  // better extraction context than none.
+  const userPersona = resolveGuardianPersona();
+
   const extracted = extractionConfig.useLLM
     ? await extractItemsWithLLM(
         text,
         extractionConfig,
         effectiveScopeId,
         message.role,
+        userPersona,
       )
     : extractItemsPatternBased(text, effectiveScopeId);
 

package/src/memory/job-handlers/conversation-starters.ts

@@ -9,6 +9,7 @@ import { and, desc, eq } from "drizzle-orm";
 import { v4 as uuid } from "uuid";
 
 import { loadSkillCatalog } from "../../config/skills.js";
+import { resolveGuardianPersona } from "../../prompts/persona-resolver.js";
 import { buildCoreIdentityContext } from "../../prompts/system-prompt.js";
 import {
   createTimeout,
@@ -171,7 +172,9 @@ async function generateStarters(scopeId: string): Promise<GeneratedStarter[]> {
 
   // Truncate identity context to prevent oversized prompts when SOUL.md /
   // IDENTITY.md / USER.md are large.
-  const rawIdentityContext = buildCoreIdentityContext();
+  const rawIdentityContext = buildCoreIdentityContext({
+    userPersona: resolveGuardianPersona(),
+  });
   const identityContext = rawIdentityContext
     ? truncate(rawIdentityContext, 2000, "\n…[truncated]")
     : null;

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

@@ -11,10 +11,7 @@ import type { MemoryJob } from "../jobs-store.js";
 import { extractMediaBlocks } from "../message-content.js";
 import {
   mediaAssets,
-  memoryChunks,
-  memoryEpisodes,
   memoryItems,
-  memoryObservations,
   memorySegments,
   memorySummaries,
   messages,
@@ -93,26 +90,6 @@ export async function embedSummaryJob(
   );
 }
 
-export async function embedChunkJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
-  const chunkId = asString(job.payload.chunkId);
-  if (!chunkId) return;
-  const db = getDb();
-  const chunk = db
-    .select()
-    .from(memoryChunks)
-    .where(eq(memoryChunks.id, chunkId))
-    .get();
-  if (!chunk) return;
-  await embedAndUpsert(config, "chunk", chunk.id, chunk.content, {
-    observation_id: chunk.observationId,
-    created_at: chunk.createdAt,
-    memory_scope_id: chunk.scopeId,
-  });
-}
-
 export async function embedMediaJob(
   job: MemoryJob,
   config: AssistantConfig,
@@ -146,40 +123,6 @@ export async function embedMediaJob(
   });
 }
 
-export async function embedObservationJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
-  const observationId = asString(job.payload.observationId);
-  const chunkId = asString(job.payload.chunkId);
-  if (!observationId || !chunkId) return;
-
-  const db = getDb();
-  const observation = db
-    .select()
-    .from(memoryObservations)
-    .where(eq(memoryObservations.id, observationId))
-    .get();
-  if (!observation) return;
-
-  const chunk = db
-    .select()
-    .from(memoryChunks)
-    .where(eq(memoryChunks.id, chunkId))
-    .get();
-  if (!chunk) return;
-
-  await embedAndUpsert(config, "observation", chunk.id, chunk.content, {
-    observation_id: observationId,
-    conversation_id: observation.conversationId,
-    role: observation.role,
-    modality: observation.modality,
-    source: observation.source,
-    created_at: observation.createdAt,
-    memory_scope_id: observation.scopeId,
-  });
-}
-
 export async function embedAttachmentJob(
   job: MemoryJob,
   config: AssistantConfig,
@@ -216,25 +159,3 @@ export async function embedAttachmentJob(
     memory_scope_id: memoryScopeId,
   });
 }
-
-export async function embedEpisodeJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
-  const episodeId = asString(job.payload.episodeId);
-  if (!episodeId) return;
-  const db = getDb();
-  const episode = db
-    .select()
-    .from(memoryEpisodes)
-    .where(eq(memoryEpisodes.id, episodeId))
-    .get();
-  if (!episode) return;
-  const text = `[episode] ${episode.title}: ${episode.summary}`;
-  await embedAndUpsert(config, "episode", episode.id, text, {
-    conversation_id: episode.conversationId,
-    created_at: episode.startAt,
-    last_seen_at: episode.endAt,
-    memory_scope_id: episode.scopeId,
-  });
-}

package/src/memory/job-utils.ts

@@ -142,7 +142,7 @@ export function truncate(text: string, max: number): string {
 
 export async function embedAndUpsert(
   config: AssistantConfig,
-  targetType: "segment" | "item" | "summary" | "observation" | "chunk" | "episode" | "media",
+  targetType: "segment" | "item" | "summary" | "media",
   targetId: string,
   input: EmbeddingInput,
   extraPayload?: Record<string, unknown>,