@mastra/memory 0.0.0-studio-deploy-20260403185613 → 0.0.0-studio-deploy-20260403231316

package/CHANGELOG.md

@@ -1,11 +1,71 @@
 # @mastra/memory
 
-## 0.0.0-studio-deploy-20260403185613
+## 0.0.0-studio-deploy-20260403231316
 
 ### Patch Changes
 
-- Updated dependencies [[`c55b527`](https://github.com/mastra-ai/mastra/commit/c55b52758a31368b2077b0dbbc3badfe4063f560)]:
-  - @mastra/core@0.0.0-studio-deploy-20260403185613
+- Fixed observational memory buffering so sealed assistant chunks stay split instead of being merged back into one persisted message during long tool runs. ([#14995](https://github.com/mastra-ai/mastra/pull/14995))
+
+- Updated dependencies [[`c55b527`](https://github.com/mastra-ai/mastra/commit/c55b52758a31368b2077b0dbbc3badfe4063f560), [`7eb2596`](https://github.com/mastra-ai/mastra/commit/7eb25960d607e07468c9a10c5437abd2deaf1e9a)]:
+  - @mastra/core@0.0.0-studio-deploy-20260403231316
+
+## 1.13.1
+
+### Patch Changes
+
+- Fixed thread titles not persisting when generated during async buffered observation. Titles now update immediately when the observer produces them, rather than being lost until activation. ([#14992](https://github.com/mastra-ai/mastra/pull/14992))
+
+- Updated dependencies [[`cb15509`](https://github.com/mastra-ai/mastra/commit/cb15509b58f6a83e11b765c945082afc027db972), [`81e4259`](https://github.com/mastra-ai/mastra/commit/81e425939b4ceeb4f586e9b6d89c3b1c1f2d2fe7), [`951b8a1`](https://github.com/mastra-ai/mastra/commit/951b8a1b5ef7e1474c59dc4f2b9fc1a8b1e508b6), [`80c5668`](https://github.com/mastra-ai/mastra/commit/80c5668e365470d3a96d3e953868fd7a643ff67c), [`3d478c1`](https://github.com/mastra-ai/mastra/commit/3d478c1e13f17b80f330ac49d7aa42ef929b93ff), [`2b4ea10`](https://github.com/mastra-ai/mastra/commit/2b4ea10b053e4ea1ab232d536933a4a3c4cba999), [`a0544f0`](https://github.com/mastra-ai/mastra/commit/a0544f0a1e6bd52ac12676228967c1938e43648d), [`6039f17`](https://github.com/mastra-ai/mastra/commit/6039f176f9c457304825ff1df8c83b8e457376c0), [`06b928d`](https://github.com/mastra-ai/mastra/commit/06b928dfc2f5630d023467476cc5919dfa858d0a), [`6a8d984`](https://github.com/mastra-ai/mastra/commit/6a8d9841f2933456ee1598099f488d742b600054), [`c8c86aa`](https://github.com/mastra-ai/mastra/commit/c8c86aa1458017fbd1c0776fdc0c520d129df8a6)]:
+  - @mastra/core@1.22.0
+
+## 1.13.1-alpha.0
+
+### Patch Changes
+
+- Fixed thread titles not persisting when generated during async buffered observation. Titles now update immediately when the observer produces them, rather than being lost until activation. ([#14992](https://github.com/mastra-ai/mastra/pull/14992))
+
+- Updated dependencies [[`cb15509`](https://github.com/mastra-ai/mastra/commit/cb15509b58f6a83e11b765c945082afc027db972), [`80c5668`](https://github.com/mastra-ai/mastra/commit/80c5668e365470d3a96d3e953868fd7a643ff67c), [`3d478c1`](https://github.com/mastra-ai/mastra/commit/3d478c1e13f17b80f330ac49d7aa42ef929b93ff), [`6039f17`](https://github.com/mastra-ai/mastra/commit/6039f176f9c457304825ff1df8c83b8e457376c0), [`06b928d`](https://github.com/mastra-ai/mastra/commit/06b928dfc2f5630d023467476cc5919dfa858d0a), [`6a8d984`](https://github.com/mastra-ai/mastra/commit/6a8d9841f2933456ee1598099f488d742b600054)]:
+  - @mastra/core@1.22.0-alpha.2
+
+## 1.13.0
+
+### Minor Changes
+
+- Added `loadUnobservedMessages({ threadId, resourceId? })` as a public method on `ObservationalMemory`. ([#14921](https://github.com/mastra-ai/mastra/pull/14921))
+
+  This lets external consumers (e.g. the Mastra gateway proxy) load previously-stored messages that haven't been observed yet, without having to reimplement the internal storage query and part-level filtering logic. The method fetches the OM record, queries storage for messages after the `lastObservedAt` cursor, and applies part-level filtering so partially-observed messages only return their unobserved parts.
+
+  ```ts
+  const unobserved = await om.loadUnobservedMessages({
+    threadId: 'thread-123',
+    resourceId: 'user-456',
+  });
+  ```
+
+### Patch Changes
+
+- Updated dependencies [[`9a43b47`](https://github.com/mastra-ai/mastra/commit/9a43b476465e86c9aca381c2831066b5c33c999a), [`ec5c319`](https://github.com/mastra-ai/mastra/commit/ec5c3197a50d034cb8e9cc494eebfddc684b5d81), [`6517789`](https://github.com/mastra-ai/mastra/commit/65177895b74b5471fe2245c7292f0176d9b3385d), [`13f4327`](https://github.com/mastra-ai/mastra/commit/13f4327f052faebe199cefbe906d33bf90238767), [`9ad6aa6`](https://github.com/mastra-ai/mastra/commit/9ad6aa6dfe858afc6955d1df5f3f78c40bb96b9c), [`2862127`](https://github.com/mastra-ai/mastra/commit/2862127d0a7cbd28523120ad64fea067a95838e6), [`3d16814`](https://github.com/mastra-ai/mastra/commit/3d16814c395931373543728994ff45ac98093074), [`7f498d0`](https://github.com/mastra-ai/mastra/commit/7f498d099eacef64fd43ee412e3bd6f87965a8a6), [`8cf8a67`](https://github.com/mastra-ai/mastra/commit/8cf8a67b061b737cb06d501fb8c1967a98bbf3cb), [`d7827e3`](https://github.com/mastra-ai/mastra/commit/d7827e393937c6cb0c7a744dde4d31538cb542b7)]:
+  - @mastra/core@1.21.0
+
+## 1.13.0-alpha.0
+
+### Minor Changes
+
+- Added `loadUnobservedMessages({ threadId, resourceId? })` as a public method on `ObservationalMemory`. ([#14921](https://github.com/mastra-ai/mastra/pull/14921))
+
+  This lets external consumers (e.g. the Mastra gateway proxy) load previously-stored messages that haven't been observed yet, without having to reimplement the internal storage query and part-level filtering logic. The method fetches the OM record, queries storage for messages after the `lastObservedAt` cursor, and applies part-level filtering so partially-observed messages only return their unobserved parts.
+
+  ```ts
+  const unobserved = await om.loadUnobservedMessages({
+    threadId: 'thread-123',
+    resourceId: 'user-456',
+  });
+  ```
+
+### Patch Changes
+
+- Updated dependencies [[`13f4327`](https://github.com/mastra-ai/mastra/commit/13f4327f052faebe199cefbe906d33bf90238767)]:
+  - @mastra/core@1.21.0-alpha.1
 
 ## 1.12.1
 

package/dist/{chunk-HLGFIN4J.cjs → chunk-4FMHSWZD.cjs}

@@ -1406,8 +1406,6 @@ var SyncObservationStrategy = class extends ObservationStrategy {
     }
   }
 };
-
-// src/processors/observational-memory/observation-strategies/async-buffer.ts
 var AsyncBufferObservationStrategy = class extends ObservationStrategy {
   startedAt;
   cycleId;
@@ -1498,6 +1496,30 @@ var AsyncBufferObservationStrategy = class extends ObservationStrategy {
       lastBufferedAtTime: processed.lastObservedAt
     });
     await this.indexObservationGroups(processed.observations, threadId, resourceId, processed.lastObservedAt);
+    const newTitle = processed.threadTitle?.trim();
+    if (newTitle && newTitle.length >= 3) {
+      const thread = await this.storage.getThreadById({ threadId });
+      if (thread) {
+        const oldTitle = thread.title?.trim();
+        if (newTitle !== oldTitle) {
+          const newMetadata = memory.setThreadOMMetadata(thread.metadata, {
+            threadTitle: processed.threadTitle
+          });
+          await this.storage.updateThread({
+            id: threadId,
+            title: newTitle,
+            metadata: newMetadata
+          });
+          const marker = createThreadUpdateMarker({
+            cycleId: this.cycleId,
+            threadId,
+            oldTitle,
+            newTitle
+          });
+          await this.streamMarker(marker);
+        }
+      }
+    }
   }
   async emitEndMarkers(_cycleId, processed) {
     if (!processed.observations || !this.opts.writer) return;
@@ -1952,6 +1974,7 @@ var ObservationStep = class {
     let buffered = false;
     let reflected = false;
     let didThresholdCleanup = false;
+    let observerExchange;
     if (this.stepNumber === 0) {
       const step0Messages = messageList.get.all.db();
       const activation = await om.activate({
@@ -2007,6 +2030,26 @@ var ObservationStep = class {
     if (statusSnapshot.shouldBuffer && !hasIncompleteToolCalls) {
       const allMessages = messageList.get.all.db();
       const unobservedMessages = om.getUnobservedMessages(allMessages, statusSnapshot.record);
+      const candidates = om.getUnobservedMessages(unobservedMessages, statusSnapshot.record, {
+        excludeBuffered: true
+      });
+      if (candidates.length > 0) {
+        om.sealMessagesForBuffering(candidates);
+        try {
+          await this.turn.hooks?.onBufferChunkSealed?.();
+        } catch (error) {
+          omDebug(
+            `[OM:buffer] onBufferChunkSealed hook failed: ${error instanceof Error ? error.message : String(error)}`
+          );
+        }
+        if (this.turn.memory) {
+          await this.turn.memory.persistMessages(candidates);
+        }
+        messageList.removeByIds(candidates.map((msg) => msg.id));
+        for (const msg of candidates) {
+          messageList.add(msg, "memory");
+        }
+      }
       void om.buffer({
         threadId,
         resourceId,
@@ -2015,23 +2058,7 @@ var ObservationStep = class {
         record: statusSnapshot.record,
         writer: this.turn.writer,
         requestContext: this.turn.requestContext,
-        observabilityContext: this.turn.observabilityContext,
-        beforeBuffer: async (candidates) => {
-          if (candidates.length === 0) {
-            return;
-          }
-          om.sealMessagesForBuffering(candidates);
-          try {
-            await this.turn.hooks?.onBufferChunkSealed?.();
-          } catch (error) {
-            omDebug(
-              `[OM:buffer] onBufferChunkSealed hook failed: ${error instanceof Error ? error.message : String(error)}`
-            );
-          }
-          if (this.turn.memory) {
-            await this.turn.memory.persistMessages(candidates);
-          }
-        }
+        observabilityContext: this.turn.observabilityContext
       }).catch((err) => {
         omDebug(`[OM:buffer] fire-and-forget buffer failed: ${err?.message}`);
       });
@@ -2050,6 +2077,7 @@ var ObservationStep = class {
       if (statusSnapshot.shouldObserve && !hasIncompleteToolCalls) {
         const preObsGeneration = this.turn.record.generationCount;
         const obsResult = await this.runThresholdObservation();
+        observerExchange = obsResult.observerExchange;
         if (obsResult.succeeded) {
           observed = true;
           didThresholdCleanup = true;
@@ -2103,6 +2131,7 @@ var ObservationStep = class {
     }
     this._context = {
       systemMessage,
+      observerExchange,
       activated,
       observed,
       buffered,
@@ -2170,7 +2199,11 @@ var ObservationStep = class {
       writer: this.turn.writer,
       observabilityContext: this.turn.observabilityContext
     });
-    return { succeeded: obsResult.observed, record: obsResult.record };
+    return {
+      succeeded: obsResult.observed,
+      record: obsResult.record,
+      observerExchange: om.observer.lastExchange
+    };
   }
 };
 
@@ -3496,6 +3529,8 @@ var ObserverRunner = class {
   observedMessageIds;
   resolveModel;
   tokenCounter;
+  /** Captured prompt/response from the last observer call (for repro capture). */
+  lastExchange;
   constructor(opts) {
     this.observationConfig = opts.observationConfig;
     this.observedMessageIds = opts.observedMessageIds;
@@ -3571,15 +3606,38 @@ var ObserverRunner = class {
     };
     let result = await doGenerate();
     let parsed = parseObserverOutput(result.text);
+    let retriedDueToDegenerate = false;
     if (parsed.degenerate) {
       omDebug(`[OM:callObserver] degenerate repetition detected, retrying once`);
       result = await doGenerate();
       parsed = parseObserverOutput(result.text);
+      retriedDueToDegenerate = true;
       if (parsed.degenerate) {
         omDebug(`[OM:callObserver] degenerate repetition on retry, failing`);
         throw new Error("Observer produced degenerate output after retry");
       }
     }
+    const systemPrompt = buildObserverSystemPrompt(
+      false,
+      this.observationConfig.instruction,
+      this.observationConfig.threadTitle
+    );
+    this.lastExchange = {
+      systemPrompt,
+      observerMessages,
+      rawOutput: result.text,
+      parsedResult: {
+        observations: parsed.observations,
+        currentTask: parsed.currentTask,
+        suggestedContinuation: parsed.suggestedContinuation,
+        threadTitle: parsed.threadTitle,
+        degenerate: parsed.degenerate
+      },
+      model: String(resolvedModel.model),
+      inputTokens,
+      isMultiThread: false,
+      retriedDueToDegenerate
+    };
     const usage = result.totalUsage ?? result.usage;
     return {
       observations: parsed.observations,
@@ -3646,15 +3704,36 @@ var ObserverRunner = class {
     };
     let result = await doGenerate();
     let parsed = parseMultiThreadObserverOutput(result.text);
+    let retriedDueToDegenerate = false;
     if (parsed.degenerate) {
       omDebug(`[OM:callMultiThreadObserver] degenerate repetition detected, retrying once`);
       result = await doGenerate();
       parsed = parseMultiThreadObserverOutput(result.text);
+      retriedDueToDegenerate = true;
       if (parsed.degenerate) {
         omDebug(`[OM:callMultiThreadObserver] degenerate repetition on retry, failing`);
         throw new Error("Multi-thread observer produced degenerate output after retry");
       }
     }
+    const systemPrompt = buildObserverSystemPrompt(
+      true,
+      this.observationConfig.instruction,
+      this.observationConfig.threadTitle
+    );
+    this.lastExchange = {
+      systemPrompt,
+      observerMessages,
+      rawOutput: result.text,
+      parsedResult: {
+        observations: Array.from(parsed.threads.values()).map((t) => t.observations).join("\n"),
+        threadTitle: Array.from(parsed.threads.values()).map((t) => t.threadTitle).filter(Boolean).join(", "),
+        degenerate: parsed.degenerate
+      },
+      model: String(resolvedModel.model),
+      inputTokens,
+      isMultiThread: true,
+      retriedDueToDegenerate
+    };
     const results = /* @__PURE__ */ new Map();
     for (const [threadId, threadResult] of parsed.threads) {
       results.set(threadId, {
@@ -6715,7 +6794,7 @@ ${suggestedResponse}
    * In resource scope mode, loads messages for the entire resource (all threads).
    * In thread scope mode, loads messages for just the current thread.
    */
-  async loadUnobservedMessages(threadId, resourceId, lastObservedAt) {
+  async loadMessagesFromStorage(threadId, resourceId, lastObservedAt) {
     const startDate = lastObservedAt ? new Date(lastObservedAt.getTime() + 1) : void 0;
     let result;
     if (this.scope === "resource" && resourceId) {
@@ -7403,7 +7482,7 @@ ${grouped}` : grouped;
     if (opts.messages) {
       unobservedMessages = this.getUnobservedMessages(opts.messages, record);
     } else {
-      const rawMessages = await this.loadUnobservedMessages(
+      const rawMessages = await this.loadMessagesFromStorage(
         threadId,
         resourceId,
         record.lastObservedAt ? new Date(record.lastObservedAt) : void 0
@@ -7516,6 +7595,27 @@ ${grouped}` : grouped;
     const record = await this.getOrCreateRecord(threadId, resourceId);
     return this.getUnobservedMessages(messages, record);
   }
+  /**
+   * Load unobserved messages from storage for a thread/resource.
+   *
+   * Fetches the OM record, queries storage for messages after the
+   * lastObservedAt cursor, then applies part-level filtering so
+   * partially-observed messages only include their unobserved parts.
+   *
+   * Use this when you need to load stored conversation history that
+   * hasn't been observed yet (e.g. in a stateless gateway proxy that
+   * only receives the latest message from the HTTP request).
+   */
+  async loadUnobservedMessages(opts) {
+    const { threadId, resourceId } = opts;
+    const record = await this.getOrCreateRecord(threadId, resourceId);
+    const rawMessages = await this.loadMessagesFromStorage(
+      threadId,
+      resourceId,
+      record.lastObservedAt ? new Date(record.lastObservedAt) : void 0
+    );
+    return this.getUnobservedMessages(rawMessages, record);
+  }
   /**
    * Create a buffered observation chunk without merging into active observations.
    *
@@ -7577,7 +7677,7 @@ ${grouped}` : grouped;
       if (opts.messages) {
         candidateMessages = this.getUnobservedMessages(opts.messages, record, { excludeBuffered: true });
       } else {
-        const rawMessages = await this.loadUnobservedMessages(
+        const rawMessages = await this.loadMessagesFromStorage(
           threadId,
           resourceId,
           record.lastObservedAt ? new Date(record.lastObservedAt) : void 0
@@ -7782,13 +7882,18 @@ ${grouped}` : grouped;
       const activatedChunks = freshChunks.filter((c) => activationResult.activatedCycleIds.includes(c.cycleId));
       const lastActivated = activatedChunks[activatedChunks.length - 1];
       if (lastActivated) {
+        const chunkThreadTitle = lastActivated.threadTitle;
         const newMetadata = memory.setThreadOMMetadata(thread.metadata, {
           suggestedResponse: lastActivated.suggestedContinuation,
-          currentTask: lastActivated.currentTask
+          currentTask: lastActivated.currentTask,
+          threadTitle: chunkThreadTitle
         });
+        const oldTitle = thread.title?.trim();
+        const newTitle = chunkThreadTitle?.trim();
+        const shouldUpdateThreadTitle = !!newTitle && newTitle.length >= 3 && newTitle !== oldTitle;
         await this.storage.updateThread({
           id: threadId,
-          title: thread.title ?? "",
+          title: shouldUpdateThreadTitle ? newTitle : thread.title ?? "",
           metadata: newMetadata
         });
       }
@@ -7819,7 +7924,7 @@ ${grouped}` : grouped;
     await this.withLock(lockKey, async () => {
       const freshRecord = await this.getOrCreateRecord(threadId, resourceId);
       generationBefore = freshRecord.generationCount;
-      const unobservedMessages = messages ? this.getUnobservedMessages(messages, freshRecord) : await this.loadUnobservedMessages(
+      const unobservedMessages = messages ? this.getUnobservedMessages(messages, freshRecord) : await this.loadMessagesFromStorage(
         threadId,
         resourceId,
         freshRecord.lastObservedAt ? new Date(freshRecord.lastObservedAt) : void 0
@@ -8223,6 +8328,20 @@ function writeProcessInputStepReproCapture(params) {
 `
       );
     }
+    if (params.observerExchange) {
+      const serialized = safeCaptureJsonOrError(params.observerExchange);
+      if (serialized.ok) {
+        fs.writeFileSync(path.join(captureDir, "observer-exchange.json"), `${JSON.stringify(serialized.value, null, 2)}
+`);
+      } else {
+        captureErrors.push({ fileName: "observer-exchange.json", error: serialized.error });
+        fs.writeFileSync(
+          path.join(captureDir, "observer-exchange.json"),
+          `${JSON.stringify({ __captureError: serialized.error }, null, 2)}
+`
+        );
+      }
+    }
     if (captureErrors.length > 0) {
       fs.writeFileSync(path.join(captureDir, "capture-error.json"), `${JSON.stringify(captureErrors, null, 2)}
 `);
@@ -8387,7 +8506,8 @@ var ObservationalMemoryProcessor = class {
             postBufferedChunks: [],
             postContextTokenCount: finalTotalPending,
             messageList,
-            details: {}
+            details: {},
+            observerExchange: ctx.observerExchange
           });
         }
       }
@@ -8475,5 +8595,5 @@ exports.stripEphemeralAnchorIds = stripEphemeralAnchorIds;
 exports.stripObservationGroups = stripObservationGroups;
 exports.truncateStringByTokens = truncateStringByTokens;
 exports.wrapInObservationGroup = wrapInObservationGroup;
-//# sourceMappingURL=chunk-HLGFIN4J.cjs.map
-//# sourceMappingURL=chunk-HLGFIN4J.cjs.map
+//# sourceMappingURL=chunk-4FMHSWZD.cjs.map
+//# sourceMappingURL=chunk-4FMHSWZD.cjs.map