@mastra/memory 1.26.1-alpha.0 → 1.26.1-alpha.3

package/dist/{src-Iw-V5CfD.js → src-MScpLVRh.js}

@@ -18521,33 +18521,62 @@ var TokenCounter = class TokenCounter {
 		if (isImageAttachment) await resolveImageDimensionsAsync(part);
 		return this.countAttachmentPartSync(part);
 	}
+	/**
+	* Count the name and the arguments of a tool call. Every state before the tool produces an
+	* output holds the same call signature in the context window, so all of those states share
+	* these cache kinds. `buildEstimateKey` hashes the text, so a shared kind stays correct and
+	* keeps the estimate warm while the invocation moves from one state to the next.
+	*/
+	countToolCallSignature(part, invocation) {
+		let tokens = 0;
+		let overheadDelta = 0;
+		if (invocation.toolName) tokens += this.readOrPersistPartEstimate(part, "tool-call-name", invocation.toolName);
+		if (invocation.args) if (typeof invocation.args === "string") tokens += this.readOrPersistPartEstimate(part, "tool-call-args", invocation.args);
+		else {
+			const argsJson = JSON.stringify(invocation.args);
+			tokens += this.readOrPersistPartEstimate(part, "tool-call-args-json", argsJson);
+			overheadDelta -= 12;
+		}
+		return {
+			tokens,
+			overheadDelta
+		};
+	}
 	countNonAttachmentPart(part) {
 		let overheadDelta = 0;
-		let toolResultDelta = 0;
+		let extraMessageDelta = 0;
 		if (part.type === "text") return {
 			tokens: this.readOrPersistPartEstimate(part, "text", part.text),
 			overheadDelta,
-			toolResultDelta
+			extraMessageDelta
 		};
 		if (part.type === "tool-invocation") {
 			const invocation = part.toolInvocation;
+			const state = invocation.state;
 			let tokens = 0;
-			if (invocation.state === "call" || invocation.state === "partial-call") {
-				if (invocation.toolName) tokens += this.readOrPersistPartEstimate(part, `tool-${invocation.state}-name`, invocation.toolName);
-				if (invocation.args) if (typeof invocation.args === "string") tokens += this.readOrPersistPartEstimate(part, `tool-${invocation.state}-args`, invocation.args);
-				else {
-					const argsJson = JSON.stringify(invocation.args);
-					tokens += this.readOrPersistPartEstimate(part, `tool-${invocation.state}-args-json`, argsJson);
-					overheadDelta -= 12;
-				}
+			if (state === "call" || state === "partial-call" || state === "approval-requested") {
+				const signature = this.countToolCallSignature(part, invocation);
+				return {
+					tokens: signature.tokens,
+					overheadDelta: overheadDelta + signature.overheadDelta,
+					extraMessageDelta
+				};
+			}
+			if (state === "approval-responded") {
+				extraMessageDelta++;
+				const signature = this.countToolCallSignature(part, invocation);
+				tokens += signature.tokens;
+				overheadDelta += signature.overheadDelta;
+				const reason = invocation.approval?.reason;
+				if (reason) tokens += this.readOrPersistPartEstimate(part, "tool-approval-reason", reason);
 				return {
 					tokens,
 					overheadDelta,
-					toolResultDelta
+					extraMessageDelta
 				};
 			}
-			if (invocation.state === "result") {
-				toolResultDelta++;
+			if (state === "result") {
+				extraMessageDelta++;
 				const { value: resultForCounting, usingStoredModelOutput } = this.resolveToolResultForTokenCounting(part, invocation.result);
 				if (resultForCounting !== void 0) {
 					const contentTokens = this.countMultimodalToolResultContent(part, resultForCounting);
@@ -18561,47 +18590,45 @@ var TokenCounter = class TokenCounter {
 				return {
 					tokens,
 					overheadDelta,
-					toolResultDelta
+					extraMessageDelta
 				};
 			}
-			if (invocation.state === "output-denied") {
-				toolResultDelta++;
+			if (state === "output-denied") {
+				extraMessageDelta++;
 				const reason = invocation.approval?.reason ?? "Tool call was not approved by the user";
 				tokens += this.readOrPersistPartEstimate(part, "tool-result-denied", reason);
 				return {
 					tokens,
 					overheadDelta,
-					toolResultDelta
+					extraMessageDelta
 				};
 			}
-			if (invocation.state === "output-error") {
-				toolResultDelta++;
-				const errorText = invocation.errorText;
-				const errorMessage = typeof errorText === "string" ? errorText : "Tool execution failed";
+			if (state === "output-error") {
+				extraMessageDelta++;
+				const errorMessage = typeof invocation.errorText === "string" ? invocation.errorText : "Tool execution failed";
 				tokens += this.readOrPersistPartEstimate(part, "tool-result-error", errorMessage);
 				return {
 					tokens,
 					overheadDelta,
-					toolResultDelta
+					extraMessageDelta
 				};
 			}
-			throw new Error(`Unhandled tool-invocation state '${part.toolInvocation?.state}' in token counting for part type '${part.type}'`);
 		}
 		if (typeof part.type === "string" && part.type.startsWith("data-")) return {
 			tokens: 0,
 			overheadDelta,
-			toolResultDelta
+			extraMessageDelta
 		};
 		if (part.type === "reasoning") return {
 			tokens: 0,
 			overheadDelta,
-			toolResultDelta
+			extraMessageDelta
 		};
 		const serialized = serializePartForTokenCounting(part);
 		return {
 			tokens: this.readOrPersistPartEstimate(part, `part-${part.type}`, serialized),
 			overheadDelta,
-			toolResultDelta
+			extraMessageDelta
 		};
 	}
 	/**
@@ -18610,7 +18637,7 @@ var TokenCounter = class TokenCounter {
 	countMessage(message) {
 		let payloadTokens = this.countString(message.role);
 		let overhead = TokenCounter.TOKENS_PER_MESSAGE;
-		let toolResultCount = 0;
+		let extraMessageCount = 0;
 		if (typeof message.content === "string") payloadTokens += this.readOrPersistMessageEstimate(message, "message-content", message.content);
 		else if (message.content && typeof message.content === "object") {
 			if (message.content.content && !Array.isArray(message.content.parts)) payloadTokens += this.readOrPersistMessageEstimate(message, "content-content", message.content.content);
@@ -18623,16 +18650,16 @@ var TokenCounter = class TokenCounter {
 				const result = this.countNonAttachmentPart(part);
 				payloadTokens += result.tokens;
 				overhead += result.overheadDelta;
-				toolResultCount += result.toolResultDelta;
+				extraMessageCount += result.extraMessageDelta;
 			}
 		}
-		if (toolResultCount > 0) overhead += toolResultCount * TokenCounter.TOKENS_PER_MESSAGE;
+		if (extraMessageCount > 0) overhead += extraMessageCount * TokenCounter.TOKENS_PER_MESSAGE;
 		return Math.round(payloadTokens + overhead);
 	}
 	async countMessageAsync(message) {
 		let payloadTokens = this.countString(message.role);
 		let overhead = TokenCounter.TOKENS_PER_MESSAGE;
-		let toolResultCount = 0;
+		let extraMessageCount = 0;
 		if (typeof message.content === "string") payloadTokens += this.readOrPersistMessageEstimate(message, "message-content", message.content);
 		else if (message.content && typeof message.content === "object") {
 			if (message.content.content && !Array.isArray(message.content.parts)) payloadTokens += this.readOrPersistMessageEstimate(message, "content-content", message.content.content);
@@ -18645,10 +18672,10 @@ var TokenCounter = class TokenCounter {
 				const result = this.countNonAttachmentPart(part);
 				payloadTokens += result.tokens;
 				overhead += result.overheadDelta;
-				toolResultCount += result.toolResultDelta;
+				extraMessageCount += result.extraMessageDelta;
 			}
 		}
-		if (toolResultCount > 0) overhead += toolResultCount * TokenCounter.TOKENS_PER_MESSAGE;
+		if (extraMessageCount > 0) overhead += extraMessageCount * TokenCounter.TOKENS_PER_MESSAGE;
 		return Math.round(payloadTokens + overhead);
 	}
 	/**
@@ -18959,9 +18986,10 @@ async function listThreadsForResource({ memory, resourceId, currentThreadId, pag
 		hasMore
 	};
 }
+const SEARCH_NOT_CONFIGURED_MESSAGE = "Search is not configured. Enable it with `retrieval: { vector: true }` and configure a vector store and embedder on your Memory instance.";
 async function searchMessagesForResource({ memory, resourceId, currentThreadId, query, topK = 10, maxTokens = DEFAULT_MAX_RESULT_TOKENS, before, after, threadScope }) {
 	if (!memory.searchMessages) return {
-		results: "Search is not configured. Enable it with `retrieval: { vector: true }` and configure a vector store and embedder on your Memory instance.",
+		results: SEARCH_NOT_CONFIGURED_MESSAGE,
 		count: 0
 	};
 	const MAX_TOPK = 20;
@@ -19507,9 +19535,16 @@ async function recallThreadFromStart({ memory, threadId, resourceId, page = 1, l
 }
 const recallTool = (_memoryConfig, options) => {
 	const isResourceScope = (options?.retrievalScope ?? "thread") === "resource";
+	const searchEnabled = options?.searchEnabled ?? true;
+	const description = isResourceScope ? `Browse conversation history. Use mode="threads" to list all threads for the current user. Use mode="messages" (default) to browse messages in the current thread or pass threadId to browse another thread in the active resource. When mode="messages" has no cursor or threadId, it defaults to the current thread and says so at the top of the result. If you pass only a cursor, it must belong to the current thread.${searchEnabled ? " Use mode=\"search\" to find messages by content across all threads." : ""}` : `Browse conversation history in the current thread. Use mode="messages" (default) to page through messages near a cursor.${searchEnabled ? " Use mode=\"search\" to find messages by content in this thread." : ""} Use mode="threads" to get the current thread's ID and title.`;
+	const modeEnum = searchEnabled ? [
+		"messages",
+		"threads",
+		"search"
+	] : ["messages", "threads"];
 	return createTool({
 		id: "recall",
-		description: isResourceScope ? "Browse conversation history. Use mode=\"threads\" to list all threads for the current user. Use mode=\"messages\" (default) to browse messages in the current thread or pass threadId to browse another thread in the active resource. When mode=\"messages\" has no cursor or threadId, it defaults to the current thread and says so at the top of the result. If you pass only a cursor, it must belong to the current thread. Use mode=\"search\" to find messages by content across all threads." : "Browse conversation history in the current thread. Use mode=\"messages\" (default) to page through messages near a cursor. Use mode=\"search\" to find messages by content in this thread. Use mode=\"threads\" to get the current thread's ID and title.",
+		description,
 		inputSchema: {
 			$schema: "http://json-schema.org/draft-07/schema#",
 			type: "object",
@@ -19517,12 +19552,8 @@ const recallTool = (_memoryConfig, options) => {
 				...isResourceScope ? {
 					mode: {
 						type: "string",
-						enum: [
-							"messages",
-							"threads",
-							"search"
-						],
-						description: "What to retrieve. \"messages\" (default) pages through message history. \"threads\" lists all threads for the current user. \"search\" finds messages by semantic similarity across all threads."
+						enum: modeEnum,
+						description: `What to retrieve. "messages" (default) pages through message history. "threads" lists all threads for the current user.${searchEnabled ? " \"search\" finds messages by semantic similarity across all threads." : ""}`
 					},
 					threadId: {
 						type: "string",
@@ -19539,18 +19570,14 @@ const recallTool = (_memoryConfig, options) => {
 					}
 				} : { mode: {
 					type: "string",
-					enum: [
-						"messages",
-						"threads",
-						"search"
-					],
-					description: "What to retrieve. \"messages\" (default) pages through message history. \"threads\" returns info about the current thread. \"search\" finds messages by semantic similarity in this thread."
+					enum: modeEnum,
+					description: `What to retrieve. "messages" (default) pages through message history. "threads" returns info about the current thread.${searchEnabled ? " \"search\" finds messages by semantic similarity in this thread." : ""}`
 				} },
-				query: {
+				...searchEnabled ? { query: {
 					type: "string",
 					minLength: 1,
 					description: "Search query for mode=\"search\". Finds messages semantically similar to this text."
-				},
+				} } : {},
 				cursor: {
 					type: "string",
 					minLength: 1,
@@ -19611,6 +19638,10 @@ const recallTool = (_memoryConfig, options) => {
 			if (!memory) throw new Error("Memory instance is required for recall");
 			if (explicitThreadId === "current" && !currentThreadId) throw new Error("Could not resolve current thread.");
 			if (mode === "search") {
+				if (!searchEnabled) return {
+					results: SEARCH_NOT_CONFIGURED_MESSAGE,
+					count: 0
+				};
 				if (!query) throw new Error("query is required for mode=\"search\"");
 				if (!resourceId) throw new Error("Resource ID is required for recall");
 				return searchMessagesForResource({
@@ -19747,6 +19778,7 @@ function deepMergeWorkingMemory(existing, update) {
 	for (const key of Object.keys(update)) {
 		const updateValue = update[key];
 		const existingValue = result[key];
+		if (updateValue === void 0) continue;
 		if (updateValue === null) delete result[key];
 		else if (Array.isArray(updateValue)) result[key] = updateValue;
 		else if (typeof updateValue === "object" && updateValue !== null && typeof existingValue === "object" && existingValue !== null && !Array.isArray(existingValue)) result[key] = deepMergeWorkingMemory(existingValue, updateValue);
@@ -19799,7 +19831,7 @@ const updateWorkingMemoryTool = (memoryConfig) => {
 			version: 1,
 			vendor: "mastra",
 			validate: (value) => {
-				const memoryValue = !!value && typeof value === "object" && !Array.isArray(value) && "memory" in value ? value.memory : stripNullsFromOptional(value, jsonSchema);
+				const memoryValue = stripNullsFromOptional(!!value && typeof value === "object" && !Array.isArray(value) && "memory" in value ? value.memory : value, jsonSchema);
 				const result = validateMemory(memoryValue);
 				return result instanceof Promise ? result.then(toWrappedResult) : toWrappedResult(result);
 			},
@@ -19816,6 +19848,7 @@ const updateWorkingMemoryTool = (memoryConfig) => {
 		id: "update-working-memory",
 		description: schema ? useStateSignals ? `${stateSignalsPreamble} Data is merged with existing memory — only include fields you want to add or update.` : `Update the working memory with new information. Data is merged with existing memory - only include fields you want to add or update. To preserve existing data, omit the field entirely. Arrays are replaced entirely when provided, so pass the complete array or omit it to keep the existing values.` : useStateSignals ? `${stateSignalsPreamble} Pass the full updated Markdown blob as a string in the memory field.` : `Update the working memory with new information. Any data not included will be overwritten. Always pass data as string to the memory field. Never pass an object.`,
 		inputSchema,
+		...usesMergeSemantics ? { strict: false } : {},
 		execute: async (inputData, context) => {
 			const workingMemoryInput = inputData;
 			const threadId = context?.agent?.threadId;
@@ -20572,6 +20605,24 @@ function getUnobservedPartsPreservingToolCallPairs(message) {
 	return preservedCalls.length > 0 ? [...preservedCalls, ...unobservedParts] : unobservedParts;
 }
 /**
+* Get the messages Observational Memory is allowed to work with.
+*
+* Messages supplied through the `context` option are per-run ephemeral input. Core's
+* persistence contract already treats them as never-persist: `MessageStateManager` routes
+* them into `userContextMessages`, and `drainUnsavedMessages` only drains input/response.
+*
+* OM builds its windows from `get.all.db()`, which includes context messages, and then
+* seals and persists candidates directly — turning ephemeral context into durable user
+* messages. Excluding them here keeps OM's window, sealing, persistence and token
+* accounting consistent with that contract.
+*/
+function getObservableMessages(messageList) {
+	const allMessages = messageList.get.all.db();
+	const contextMessageIds = messageList.makeMessageSourceChecker().context;
+	if (contextMessageIds.size === 0) return allMessages;
+	return allMessages.filter((message) => !contextMessageIds.has(message.id));
+}
+/**
 * Safely extract buffered observation chunks from a record.
 * Handles both array and JSON-string formats, returning empty array if malformed.
 */
@@ -20584,7 +20635,7 @@ function getUnobservedPartsPreservingToolCallPairs(message) {
 */
 function filterObservedMessages(opts) {
 	const { messageList, record } = opts;
-	const allMessages = messageList.get.all.db();
+	const allMessages = getObservableMessages(messageList);
 	const useMarkerBoundaryPruning = opts.useMarkerBoundaryPruning ?? true;
 	const preserveMessageIds = opts.preserveMessageIds ?? /* @__PURE__ */ new Set();
 	const observedIds = new Set(Array.isArray(record?.observedMessageIds) ? record.observedMessageIds : []);
@@ -21137,7 +21188,7 @@ var ObservationStrategy = class ObservationStrategy {
 	*/
 	async persistMarkerToMessage(marker, messageList, threadId, resourceId) {
 		if (!messageList) return false;
-		const allMsgs = messageList.get.all.db();
+		const allMsgs = getObservableMessages(messageList);
 		for (let i = allMsgs.length - 1; i >= 0; i--) {
 			const msg = allMsgs[i];
 			if (msg?.role === "assistant" && msg.content?.parts && Array.isArray(msg.content.parts)) {
@@ -22016,7 +22067,7 @@ var ObservationStep = class {
 		let didThresholdCleanup = false;
 		let observerExchange;
 		if (this.stepNumber === 0) {
-			const step0Messages = messageList.get.all.db();
+			const step0Messages = getObservableMessages(messageList);
 			const activation = await om.activate({
 				threadId,
 				resourceId,
@@ -22048,7 +22099,7 @@ var ObservationStep = class {
 				currentModel: this.turn.actorModelContext,
 				requestContext: this.turn.requestContext,
 				observabilityContext: this.turn.observabilityContext,
-				lastActivityAt: getLastActivityFromMessages(messageList.get.all.db()),
+				lastActivityAt: getLastActivityFromMessages(getObservableMessages(messageList)),
 				reflectionHooks: om.composeHooks(void 0, {
 					threadId,
 					resourceId,
@@ -22058,7 +22109,7 @@ var ObservationStep = class {
 			await this.turn.refreshRecord();
 			if (this.turn.record.generationCount > preReflectGeneration) reflected = true;
 		}
-		const allMsgsForToolCheck = messageList.get.all.db();
+		const allMsgsForToolCheck = getObservableMessages(messageList);
 		const lastMessage = allMsgsForToolCheck[allMsgsForToolCheck.length - 1];
 		const pendingStepMessages = [...messageList.get.input.db(), ...messageList.get.response.db()];
 		const latestStepParts = [...getLatestStepParts(lastMessage?.content?.parts ?? []), ...pendingStepMessages.flatMap((msg) => getLatestStepParts(msg.content?.parts ?? []))];
@@ -22067,10 +22118,10 @@ var ObservationStep = class {
 		let statusSnapshot = await om.getStatus({
 			threadId,
 			resourceId,
-			messages: messageList.get.all.db()
+			messages: getObservableMessages(messageList)
 		});
 		if (statusSnapshot.shouldBuffer && !hasIncompleteToolCalls) {
-			const allMessages = messageList.get.all.db();
+			const allMessages = getObservableMessages(messageList);
 			const unobservedMessages = om.getUnobservedMessages(allMessages, statusSnapshot.record);
 			const candidates = om.getUnobservedMessages(unobservedMessages, statusSnapshot.record, { excludeBuffered: true });
 			if (candidates.length > 0) {
@@ -22162,7 +22213,7 @@ var ObservationStep = class {
 			statusSnapshot = await om.getStatus({
 				threadId,
 				resourceId,
-				messages: messageList.get.all.db()
+				messages: getObservableMessages(messageList)
 			});
 		}
 		const otherThreadsContext = await this.turn.refreshOtherThreadsContext();
@@ -22211,7 +22262,7 @@ var ObservationStep = class {
 		const { threadId, resourceId, messageList } = this.turn;
 		const om = this.turn.om;
 		await om.waitForBuffering(threadId, resourceId);
-		const observableMessages = this.seededResponseMessage ? messageList.get.all.db().filter((msg) => msg.id !== this.turn.responseMessageId) : messageList.get.all.db();
+		const observableMessages = this.seededResponseMessage ? getObservableMessages(messageList).filter((msg) => msg.id !== this.turn.responseMessageId) : getObservableMessages(messageList);
 		const freshStatus = await om.getStatus({
 			threadId,
 			resourceId,
@@ -22241,7 +22292,7 @@ var ObservationStep = class {
 					currentModel: this.turn.actorModelContext,
 					requestContext: this.turn.requestContext,
 					observabilityContext: this.turn.observabilityContext,
-					lastActivityAt: getLastActivityFromMessages(messageList.get.all.db()),
+					lastActivityAt: getLastActivityFromMessages(getObservableMessages(messageList)),
 					reflectionHooks: om.composeHooks(void 0, {
 						threadId,
 						resourceId,
@@ -22267,7 +22318,7 @@ var ObservationStep = class {
 		});
 		if (obsResult.observed) {
 			const observedMessageIds = new Set(obsResult.record.observedMessageIds ?? []);
-			const liveMessages = messageList.get.all.db();
+			const liveMessages = getObservableMessages(messageList);
 			let latestObservedIndex = -1;
 			for (let i = liveMessages.length - 1; i >= 0; i--) {
 				const message = liveMessages[i];
@@ -22442,7 +22493,7 @@ var ObservationTurn = class {
 		const asyncObservationEnabled = this.om.buffering.isAsyncObservationEnabled();
 		const bufferOnIdle = this.om.getObservationConfig().bufferOnIdle;
 		if (asyncObservationEnabled && bufferOnIdle) {
-			const allMessages = this.messageList.get.all.db();
+			const allMessages = getObservableMessages(this.messageList);
 			const record = this._record;
 			const unobservedMessages = this.om.getUnobservedMessages(allMessages, record);
 			if (unobservedMessages.length > 0) this.om.buffer({
@@ -22812,7 +22863,7 @@ function getCurrentModel$1(model) {
 	return formatModelContext$1(model?.provider, model?.modelId);
 }
 function getLastModelFromMessageList(messageList) {
-	const messages = messageList?.get.all.db();
+	const messages = messageList ? getObservableMessages(messageList) : void 0;
 	if (!messages) return void 0;
 	for (let i = messages.length - 1; i >= 0; i--) {
 		const message = messages[i];
@@ -24396,7 +24447,7 @@ var ObservationalMemory = class ObservationalMemory {
 	*/
 	async persistMarkerToMessage(marker, messageList, threadId, resourceId) {
 		if (!messageList) return;
-		const allMsgs = messageList.get.all.db();
+		const allMsgs = getObservableMessages(messageList);
 		for (let i = allMsgs.length - 1; i >= 0; i--) {
 			const msg = allMsgs[i];
 			if (msg?.role === "assistant" && msg.content?.parts && Array.isArray(msg.content.parts)) {
@@ -25074,7 +25125,7 @@ ${formattedMessages}
 	async cleanupMessages(opts) {
 		const { threadId, resourceId, observedMessageIds, retentionFloor, preserveMessageIds } = opts;
 		const messageList = this.isMessageList(opts.messages) ? opts.messages : void 0;
-		const allMsgs = messageList ? messageList.get.all.db() : opts.messages;
+		const allMsgs = messageList ? getObservableMessages(messageList) : opts.messages;
 		let markerIdx = -1;
 		let markerMsg = null;
 		for (let i = allMsgs.length - 1; i >= 0; i--) {
@@ -26190,7 +26241,7 @@ function isTemporalGapMarkerForMessage(message, targetMessageId) {
 async function insertTemporalGapMarkers({ messageList, sendSignal }) {
 	const latestInputMessage = messageList.get.input.db().filter((message) => Boolean(message)).at(-1);
 	if (!latestInputMessage || isTemporalGapMarker(latestInputMessage)) return;
-	const allMessages = messageList.get.all.db().filter((message) => Boolean(message));
+	const allMessages = getObservableMessages(messageList).filter((message) => Boolean(message));
 	const latestInputIndex = allMessages.findIndex((message) => message.id === latestInputMessage.id);
 	if (latestInputIndex <= 0) return;
 	if (allMessages.some((message) => isTemporalGapMarkerForMessage(message, latestInputMessage.id))) return;
@@ -26414,7 +26465,7 @@ var ObservationalMemoryProcessor = class {
 					threadId,
 					resourceId
 				});
-				const allDbMsgs = messageList.get.all.db();
+				const allDbMsgs = getObservableMessages(messageList);
 				const tokenCounter = this.engine.getTokenCounter();
 				const contextTokens = await tokenCounter.countMessagesAsync(allDbMsgs);
 				const otherThreadsContext = this.turn.context.otherThreadsContext;
@@ -26784,6 +26835,18 @@ var Memory = class extends MastraMemory {
 	_omEngine;
 	_omEngineInstance;
 	_mastraInstance;
+	/**
+	* Every vector cleanup that deleteThread or deleteMessages started in the background.
+	* Callers do not wait for the cleanup, so this handle is the only join point.
+	*/
+	pendingVectorCleanup = Promise.resolve();
+	/**
+	* Adds a background vector cleanup to the join handle.
+	* The handle keeps the earlier cleanups, so it settles only after all of them end.
+	*/
+	trackVectorCleanup(cleanup) {
+		this.pendingVectorCleanup = Promise.allSettled([this.pendingVectorCleanup, cleanup]).then(() => void 0);
+	}
 	/** The shared ObservationalMemory engine. Lazily created on first access. */
 	get omEngine() {
 		if (!this._omEngine) this._omEngine = this._initOMEngine().then((engine) => {
@@ -27048,26 +27111,40 @@ var Memory = class extends MastraMemory {
 		const thread = await memoryStore.getThreadById({ threadId });
 		await memoryStore.deleteThread({ threadId });
 		if (thread?.resourceId && memoryStore.supportsObservationalMemory) await memoryStore.clearObservationalMemory(threadId, thread.resourceId);
-		if (this.vector) this.deleteThreadVectors(threadId);
+		if (this.vector) this.trackVectorCleanup(this.deleteThreadVectors(threadId));
+	}
+	/**
+	* Prefix shared by every message index. The index for the default embedding
+	* dimension is named with the bare prefix; other dimensions add a suffix.
+	*/
+	get messageIndexPrefix() {
+		return this.getEmbeddingIndexName();
 	}
 	/**
-	* Lists all vector indexes that match the memory messages prefix.
-	* Handles separator differences across vector store backends (e.g. '_' vs '-').
+	* Prefix shared by every observation index. Each observation index adds a dimension suffix.
 	*/
-	async getMemoryVectorIndexes() {
+	get observationIndexPrefix() {
+		return `memory${this.vector?.indexSeparator ?? "_"}observations`;
+	}
+	/**
+	* Lists the vector indexes whose name starts with one of the given prefixes.
+	* Index names can carry a dimension suffix, so discovery matches on the prefix.
+	*/
+	async getMemoryVectorIndexes(prefixes) {
 		if (!this.vector) return [];
-		const prefix = `memory${this.vector.indexSeparator ?? "_"}messages`;
-		return (await this.vector.listIndexes()).filter((name) => name.startsWith(prefix));
+		return (await this.vector.listIndexes()).filter((name) => prefixes.some((prefix) => name.startsWith(prefix)));
 	}
 	/**
 	* Deletes all vector embeddings associated with a thread.
 	* This is called internally by deleteThread to clean up orphaned vectors.
+	* Both message and observation vectors are removed, so no text of the deleted
+	* thread stays reachable through resource-scoped retrieval.
 	*
 	* @param threadId - The ID of the thread whose vectors should be deleted
 	*/
 	async deleteThreadVectors(threadId) {
 		try {
-			const memoryIndexes = await this.getMemoryVectorIndexes();
+			const memoryIndexes = await this.getMemoryVectorIndexes([this.messageIndexPrefix, this.observationIndexPrefix]);
 			await Promise.all(memoryIndexes.map(async (indexName) => {
 				try {
 					await this.vector.deleteVectors({
@@ -27075,14 +27152,14 @@ var Memory = class extends MastraMemory {
 						filter: { thread_id: threadId }
 					});
 				} catch {
-					this.logger.debug("Failed to delete vectors for thread, skipping", {
+					this.logger.warn("Failed to delete vectors of the deleted thread from index", {
 						threadId,
 						indexName
 					});
 				}
 			}));
 		} catch {
-			this.logger.debug("Failed to clean up vectors for thread", { threadId });
+			this.logger.warn("Failed to clean up vectors of the deleted thread", { threadId });
 		}
 	}
 	async updateWorkingMemory({ threadId, resourceId, workingMemory, memoryConfig, observabilityContext }) {
@@ -27755,7 +27832,7 @@ Notes:
 	getObservationEmbeddingIndexName(dimensions) {
 		const usedDimensions = dimensions ?? 384;
 		const separator = this.vector?.indexSeparator ?? "_";
-		return `memory${separator}observations${separator}${usedDimensions}`;
+		return `${this.observationIndexPrefix}${separator}${usedDimensions}`;
 	}
 	async createObservationEmbeddingIndex(dimensions) {
 		const usedDimensions = dimensions ?? 384;
@@ -28001,7 +28078,10 @@ Notes:
 			tools[name] = tool;
 		}
 		const omConfig = normalizeObservationalMemoryConfig(mergedConfig.observationalMemory);
-		if (omConfig?.retrieval) tools.recall = recallTool(mergedConfig, { retrievalScope: typeof omConfig.retrieval === "object" ? omConfig.retrieval.scope ?? "resource" : "resource" });
+		if (omConfig?.retrieval) tools.recall = recallTool(mergedConfig, {
+			retrievalScope: typeof omConfig.retrieval === "object" ? omConfig.retrieval.scope ?? "resource" : "resource",
+			searchEnabled: this.hasRetrievalSearch(omConfig.retrieval)
+		});
 		return tools;
 	}
 	/**
@@ -28057,7 +28137,7 @@ Notes:
 				}));
 				const messageIdsNeedingDeletion = /* @__PURE__ */ new Set([...messageIdsWithClearedContent, ...messageIdsWithNewEmbeddings]);
 				if (messageIdsNeedingDeletion.size > 0) try {
-					const memoryIndexes = await this.getMemoryVectorIndexes();
+					const memoryIndexes = await this.getMemoryVectorIndexes([this.messageIndexPrefix]);
 					const idsToDelete = [...messageIdsNeedingDeletion];
 					await Promise.all(memoryIndexes.map(async (indexName) => {
 						for (let i = 0; i < idsToDelete.length; i += VECTOR_DELETE_BATCH_SIZE) {
@@ -28116,7 +28196,7 @@ Notes:
 		const span = this.createMemorySpan("delete", observabilityContext, void 0, { messageCount: messageIds.length });
 		try {
 			await (await this.getMemoryStore()).deleteMessages(messageIds);
-			if (this.vector) this.deleteMessageVectors(messageIds);
+			if (this.vector) this.trackVectorCleanup(this.deleteMessageVectors(messageIds));
 			span?.end({
 				output: { success: true },
 				attributes: { messageCount: messageIds.length }
@@ -28132,12 +28212,14 @@ Notes:
 	/**
 	* Deletes vector embeddings for specific messages.
 	* This is called internally by deleteMessages to clean up orphaned vectors.
+	* Only the message indexes are touched, because observation vectors can hold
+	* text of other messages of the thread.
 	*
 	* @param messageIds - The IDs of the messages whose vectors should be deleted
 	*/
 	async deleteMessageVectors(messageIds) {
 		try {
-			const memoryIndexes = await this.getMemoryVectorIndexes();
+			const memoryIndexes = await this.getMemoryVectorIndexes([this.messageIndexPrefix]);
 			await Promise.all(memoryIndexes.map(async (indexName) => {
 				for (let i = 0; i < messageIds.length; i += VECTOR_DELETE_BATCH_SIZE) {
 					const batch = messageIds.slice(i, i + VECTOR_DELETE_BATCH_SIZE);
@@ -28531,4 +28613,4 @@ Notes:
 //#endregion
 export { extractCurrentTask as A, OBSERVATION_CONTEXT_INSTRUCTIONS as B, WorkingMemoryExtractor as C, OBSERVER_SYSTEM_PROMPT as D, TokenCounter as E, injectAnchorIds as F, OBSERVATION_CONTINUATION_HINT as H, parseAnchorId as I, stripEphemeralAnchorIds as L, hasCurrentTaskSection as M, optimizeObservationsForContext as N, buildObserverPrompt as O, parseObserverOutput as P, Extractor as R, deepMergeWorkingMemory as S, summarizeConversation as T, OBSERVATION_CONTEXT_PROMPT as V, reconcileObservationGroupsFromReflection as _, extractWorkingMemoryContent as a, wrapInObservationGroup as b, WORKING_MEMORY_STATE_ID as c, getObservationsAsOf as d, ObservationalMemoryProcessor as f, parseObservationGroups as g, deriveObservationGroupProvenance as h, WorkingMemory as i, formatMessagesForObserver as j, buildObserverSystemPrompt as k, WORKING_MEMORY_STATE_PROCESSOR_ID as l, combineObservationGroupRanges as m, MessageHistory$1 as n, extractWorkingMemoryTags as o, ObservationalMemory as p, SemanticRecall as r, removeWorkingMemoryTags as s, Memory as t, WorkingMemoryStateProcessor as u, renderObservationGroupsForReflection as v, SUMMARIZE_THREAD_DEFAULTS as w, ModelByInputTokens as x, stripObservationGroups as y, OBSERVATIONAL_MEMORY_DEFAULTS as z };
 
-//# sourceMappingURL=src-Iw-V5CfD.js.map
+//# sourceMappingURL=src-MScpLVRh.js.map