@mastra/memory 1.13.2-alpha.0 → 1.14.0-alpha.2

package/CHANGELOG.md

@@ -1,5 +1,78 @@
 # @mastra/memory
 
+## 1.14.0-alpha.2
+
+### Minor Changes
+
+- Added tracing support to Memory operations (recall, save, delete, update working memory). When an `observabilityContext` is provided, Memory creates `MEMORY_OPERATION` spans that capture operation type, message counts, embedding token usage, and vector result counts. Tracing is fully opt-in — existing usage without `observabilityContext` is unaffected. ([#14305](https://github.com/mastra-ai/mastra/pull/14305))
+
+  **Example usage:**
+
+  ```typescript
+  import { Memory } from '@mastra/memory';
+  import { InMemoryStore } from '@mastra/core/storage';
+
+  const memory = new Memory({ storage: new InMemoryStore() });
+
+  // Pass observabilityContext to create observable spans
+  await memory.recall({
+    threadId: 'thread-1',
+    observabilityContext: { tracingContext: { currentSpan: parentSpan } },
+  });
+
+  await memory.saveMessages({
+    messages: [userMessage, assistantMessage],
+    observabilityContext: { tracingContext: { currentSpan: parentSpan } },
+  });
+  ```
+
+- Added per-record config overrides for observation and reflection thresholds in Observational Memory. Each thread can now have its own `messageTokens` and `observationTokens` thresholds that override the instance-level defaults, without requiring a process restart or cache invalidation. If no per-record override is set, the instance-level config is used as before. ([#15102](https://github.com/mastra-ai/mastra/pull/15102))
+
+### Patch Changes
+
+- Fixed recall() to hide dynamic system reminder messages by default, with includeSystemReminders available when callers need raw reminder history. ([#15100](https://github.com/mastra-ai/mastra/pull/15100))
+
+- Updated dependencies [[`ac7baf6`](https://github.com/mastra-ai/mastra/commit/ac7baf66ef1db15e03975ef4ebb02724f015a391), [`0df8321`](https://github.com/mastra-ai/mastra/commit/0df832196eeb2450ab77ce887e8553abdd44c5a6), [`61109b3`](https://github.com/mastra-ai/mastra/commit/61109b34feb0e38d54bee4b8ca83eb7345b1d557), [`33f1ead`](https://github.com/mastra-ai/mastra/commit/33f1eadfa19c86953f593478e5fa371093b33779)]:
+  - @mastra/core@1.23.0-alpha.8
+
+## 1.14.0-alpha.1
+
+### Minor Changes
+
+- Added usage data to ObserveHooks callbacks and standalone reflect() return. ([#15047](https://github.com/mastra-ai/mastra/pull/15047))
+
+  **ObserveHooks:** `onObservationEnd` and `onReflectionEnd` now receive a result object containing token usage from the underlying LLM call. This enables reliable usage tracking across all observation and reflection paths (sync, async buffered, and resource-scoped).
+
+  **Standalone reflect():** `reflect()` now returns `{ reflected, record, usage? }` so callers can capture token usage without hooks.
+
+  **Examples**
+
+  ```ts
+  // Via hooks
+  await memory.observe({
+    threadId,
+    messages,
+    hooks: {
+      onObservationEnd: ({ usage }) => {
+        // usage: { inputTokens, outputTokens, totalTokens }
+      },
+      onReflectionEnd: ({ usage }) => {
+        // usage: { inputTokens, outputTokens, totalTokens }
+      },
+    },
+  });
+
+  // Via standalone reflect()
+  const { reflected, usage } = await memory.reflect(threadId, resourceId);
+  ```
+
+  Existing callbacks that accept no arguments continue to work without changes.
+
+### Patch Changes
+
+- Updated dependencies [[`fff91cf`](https://github.com/mastra-ai/mastra/commit/fff91cf914de0e731578aacebffdeebef82f0440)]:
+  - @mastra/core@1.23.0-alpha.4
+
 ## 1.13.2-alpha.0
 
 ### Patch Changes

package/dist/{chunk-C7PARRAD.js → chunk-GXDPND6K.js}

@@ -945,7 +945,7 @@ var ObservationStrategy = class _ObservationStrategy {
   static create;
   /**
    * Run the full observation lifecycle.
-   * @returns `true` if a full observation cycle completed; `false` if skipped (stale lock) or async-buffer failure was swallowed.
+   * @returns Result with `observed` flag and optional `usage` from the observer LLM call.
    * @throws On sync/resource-scoped observer failure after failed markers (same as pre–Option-A contract).
    */
   async run() {
@@ -955,7 +955,7 @@ var ObservationStrategy = class _ObservationStrategy {
       if (this.needsLock) {
         const fresh = await this.storage.getObservationalMemory(record.threadId, record.resourceId);
         if (fresh?.lastObservedAt && record.lastObservedAt && fresh.lastObservedAt > record.lastObservedAt) {
-          return false;
+          return { observed: false };
         }
       }
       const { messages, existingObservations } = await this.prepare();
@@ -976,7 +976,7 @@ var ObservationStrategy = class _ObservationStrategy {
           observabilityContext: this.opts.observabilityContext
         });
       }
-      return true;
+      return { observed: true, usage: output.usage };
     } catch (error) {
       await this.emitFailedMarkers(cycleId, error);
       if (!this.rethrowOnFailure) {
@@ -995,7 +995,7 @@ var ObservationStrategy = class _ObservationStrategy {
         });
         if (abortSignal?.aborted) throw error;
         omError("[OM] Observation failed", error);
-        return false;
+        return { observed: false };
       }
       omError("[OM] Observation failed", error);
       throw error;
@@ -4193,7 +4193,7 @@ var ReflectorRunner = class {
   /**
    * Start an async buffered reflection in the background.
    */
-  startAsyncBufferedReflection(record, observationTokens, lockKey, writer, requestContext, observabilityContext) {
+  startAsyncBufferedReflection(record, observationTokens, lockKey, writer, requestContext, observabilityContext, reflectionHooks) {
     const bufferKey = this.buffering.getReflectionBufferKey(lockKey);
     if (this.buffering.isAsyncBufferingInProgress(bufferKey)) {
       return;
@@ -4203,7 +4203,10 @@ var ReflectorRunner = class {
     this.storage.setBufferingReflectionFlag(record.id, true).catch((err) => {
       omError("[OM] Failed to set buffering reflection flag", err);
     });
-    const asyncOp = this.doAsyncBufferedReflection(record, bufferKey, writer, requestContext, observabilityContext).catch(async (error) => {
+    reflectionHooks?.onReflectionStart?.();
+    const asyncOp = this.doAsyncBufferedReflection(record, bufferKey, writer, requestContext, observabilityContext).then((usage) => {
+      reflectionHooks?.onReflectionEnd?.({ usage });
+    }).catch(async (error) => {
       if (writer) {
         const failedMarker = createBufferingFailedMarker({
           cycleId: `reflect-buf-${Date.now()}-${Math.random().toString(36).slice(2, 11)}`,
@@ -4219,6 +4222,10 @@ var ReflectorRunner = class {
         await this.persistMarkerToStorage(failedMarker, record.threadId ?? "", record.resourceId ?? void 0);
       }
       omError("[OM] Async buffered reflection failed", error);
+      reflectionHooks?.onReflectionEnd?.({
+        usage: void 0,
+        error: error instanceof Error ? error : new Error(String(error))
+      });
       BufferingCoordinator.lastBufferedBoundary.delete(bufferKey);
     }).finally(() => {
       BufferingCoordinator.asyncBufferingOps.delete(bufferKey);
@@ -4312,6 +4319,7 @@ var ReflectorRunner = class {
       });
       await this.persistMarkerToStorage(endMarker, currentRecord.threadId ?? "", currentRecord.resourceId ?? void 0);
     }
+    return reflectResult.usage;
   }
   /**
    * Try to activate buffered reflection when threshold is reached.
@@ -4432,7 +4440,8 @@ ${unreflectedContent}` : freshRecord.bufferedReflection;
           lockKey,
           writer,
           requestContext,
-          observabilityContext
+          observabilityContext,
+          reflectionHooks
         );
       }
     }
@@ -4466,7 +4475,8 @@ ${unreflectedContent}` : freshRecord.bufferedReflection;
           lockKey,
           writer,
           requestContext,
-          observabilityContext
+          observabilityContext,
+          reflectionHooks
         );
         return;
       }
@@ -4505,6 +4515,8 @@ ${unreflectedContent}` : freshRecord.bufferedReflection;
       recordId: record.id,
       threadId
     } : void 0;
+    let reflectionUsage;
+    let reflectionError;
     try {
       const compressionStartLevel = await this.getCompressionStartLevel(requestContext);
       const reflectResult = await this.call(
@@ -4518,6 +4530,7 @@ ${unreflectedContent}` : freshRecord.bufferedReflection;
         requestContext,
         observabilityContext
       );
+      reflectionUsage = reflectResult.usage;
       const reflectionTokenCount = this.tokenCounter.countObservations(reflectResult.observations);
       await this.storage.createReflectionGeneration({
         currentRecord: record,
@@ -4562,13 +4575,14 @@ ${unreflectedContent}` : freshRecord.bufferedReflection;
         await writer.custom(failedMarker).catch(() => {
         });
       }
+      reflectionError = error instanceof Error ? error : new Error(String(error));
       if (abortSignal?.aborted) {
         throw error;
       }
       omError("[OM] Reflection failed", error);
     } finally {
       await this.storage.setReflectingFlag(record.id, false);
-      reflectionHooks?.onReflectionEnd?.();
+      reflectionHooks?.onReflectionEnd?.({ usage: reflectionUsage, error: reflectionError });
       unregisterOp(record.id, "reflecting");
     }
   }
@@ -6244,6 +6258,47 @@ Async buffering is enabled by default \u2014 this opt-out is only needed when us
       }
     }
   }
+  /**
+   * Resolve the effective messageTokens for a record.
+   * Only explicit per-record overrides (stored under `_overrides`) win;
+   * the initial config snapshot written by getOrCreateRecord() is ignored
+   * so that later instance-level changes still take effect.
+   *
+   * Overrides that fall below the instance-level buffering floor
+   * (bufferTokens / absolute bufferActivation) are clamped to the
+   * instance threshold to preserve buffering invariants.
+   */
+  getEffectiveMessageTokens(record) {
+    const overrides = record.config?._overrides;
+    const recordTokens = overrides?.observation?.messageTokens;
+    if (recordTokens) {
+      const maxOverride = getMaxThreshold(recordTokens);
+      const bufferTokens = this.observationConfig.bufferTokens;
+      if (bufferTokens && maxOverride <= bufferTokens) {
+        return this.observationConfig.messageTokens;
+      }
+      const bufferActivation = this.observationConfig.bufferActivation;
+      if (bufferActivation && bufferActivation >= 1e3 && maxOverride <= bufferActivation) {
+        return this.observationConfig.messageTokens;
+      }
+      return recordTokens;
+    }
+    return this.observationConfig.messageTokens;
+  }
+  /**
+   * Resolve the effective reflection observationTokens for a record.
+   * Only explicit per-record overrides (stored under `_overrides`) win;
+   * the initial config snapshot is ignored so instance-level changes
+   * still take effect for existing records.
+   */
+  getEffectiveReflectionTokens(record) {
+    const overrides = record.config?._overrides;
+    const recordTokens = overrides?.reflection?.observationTokens;
+    if (recordTokens) {
+      return recordTokens;
+    }
+    return this.reflectionConfig.observationTokens;
+  }
   /**
    * Check whether the unobserved message tokens meet the observation threshold.
    */
@@ -6251,7 +6306,7 @@ Async buffering is enabled by default \u2014 this opt-out is only needed when us
     const { record, unobservedTokens, extraTokens = 0 } = opts;
     const pendingTokens = (record.pendingMessageTokens ?? 0) + unobservedTokens + extraTokens;
     const currentObservationTokens = record.observationTokenCount ?? 0;
-    const threshold = calculateDynamicThreshold(this.observationConfig.messageTokens, currentObservationTokens);
+    const threshold = calculateDynamicThreshold(this.getEffectiveMessageTokens(record), currentObservationTokens);
     return pendingTokens >= threshold;
   }
   /**
@@ -7400,7 +7455,7 @@ ${grouped}` : grouped;
       const projectedMessageRemoval = calculateProjectedMessageRemoval(
         bufferedChunks,
         this.observationConfig.bufferActivation ?? 1,
-        getMaxThreshold(this.observationConfig.messageTokens),
+        getMaxThreshold(this.getEffectiveMessageTokens(record)),
         pendingTokens
       );
       let obsBufferStatus = "idle";
@@ -7489,7 +7544,7 @@ ${grouped}` : grouped;
       otherThreadTokens = otherContext ? this.tokenCounter.countString(otherContext) : 0;
     }
     const pendingTokens = Math.max(0, contextWindowTokens + otherThreadTokens);
-    const threshold = calculateDynamicThreshold(this.observationConfig.messageTokens, currentObservationTokens);
+    const threshold = calculateDynamicThreshold(this.getEffectiveMessageTokens(record), currentObservationTokens);
     const bufferedChunks = getBufferedChunks(record);
     const bufferedChunkCount = bufferedChunks.length;
     const bufferedChunkTokens = bufferedChunks.reduce((sum, chunk) => sum + (chunk.messageTokens ?? 0), 0);
@@ -7506,11 +7561,12 @@ ${grouped}` : grouped;
       );
     }
     const shouldObserve = pendingTokens >= threshold;
-    const reflectThreshold = getMaxThreshold(this.reflectionConfig.observationTokens);
+    const reflectThreshold = getMaxThreshold(this.getEffectiveReflectionTokens(record));
     const shouldReflect = currentObservationTokens >= reflectThreshold;
     const canActivate = bufferedChunkCount > 0;
-    const isSharedBudget = typeof this.observationConfig.messageTokens !== "number";
-    const totalBudget = isSharedBudget ? this.observationConfig.messageTokens.max : 0;
+    const effectiveMessageTokens = this.getEffectiveMessageTokens(record);
+    const isSharedBudget = typeof effectiveMessageTokens !== "number";
+    const totalBudget = isSharedBudget ? effectiveMessageTokens.max : 0;
     const effectiveObservationTokensThreshold = isSharedBudget ? Math.max(totalBudget - threshold, 1e3) : reflectThreshold;
     const unbufferedPendingTokens = Math.max(0, pendingTokens - bufferedChunkTokens);
     return {
@@ -7825,7 +7881,7 @@ ${grouped}` : grouped;
     if (!freshChunks.length) {
       return { activated: false, record };
     }
-    const messageTokensThreshold = getMaxThreshold(this.observationConfig.messageTokens);
+    const messageTokensThreshold = getMaxThreshold(this.getEffectiveMessageTokens(freshRecord));
     const bufferActivation = this.observationConfig.bufferActivation ?? 0.7;
     const activationRatio = resolveActivationRatio(bufferActivation, messageTokensThreshold);
     const totalChunkMessageTokens = freshChunks.reduce((sum, c) => sum + (c.messageTokens ?? 0), 0);
@@ -7913,6 +7969,7 @@ ${grouped}` : grouped;
     const lockKey = this.buffering.getLockKey(threadId, resourceId);
     const reflectionHooks = hooks ? { onReflectionStart: hooks.onReflectionStart, onReflectionEnd: hooks.onReflectionEnd } : void 0;
     let observed = false;
+    let observationUsage;
     let generationBefore = -1;
     await this.withLock(lockKey, async () => {
       const freshRecord = await this.getOrCreateRecord(threadId, resourceId);
@@ -7929,8 +7986,9 @@ ${grouped}` : grouped;
         return;
       }
       hooks?.onObservationStart?.();
+      let observationError;
       try {
-        observed = await ObservationStrategy.create(this, {
+        const result = await ObservationStrategy.create(this, {
           record: freshRecord,
           threadId,
           resourceId,
@@ -7940,8 +7998,13 @@ ${grouped}` : grouped;
           writer: opts.writer,
           observabilityContext: opts.observabilityContext
         }).run();
+        observed = result.observed;
+        observationUsage = result.usage;
+      } catch (error) {
+        observationError = error instanceof Error ? error : new Error(String(error));
+        throw error;
       } finally {
-        hooks?.onObservationEnd?.();
+        hooks?.onObservationEnd?.({ usage: observationUsage, error: observationError });
       }
     });
     const record = await this.getOrCreateRecord(threadId, resourceId);
@@ -7962,12 +8025,12 @@ ${grouped}` : grouped;
   async reflect(threadId, resourceId, prompt, requestContext, observabilityContext) {
     const record = await this.getOrCreateRecord(threadId, resourceId);
     if (!record.activeObservations) {
-      return { reflected: false, record };
+      return { reflected: false, record, usage: void 0 };
     }
     await this.storage.setReflectingFlag(record.id, true);
     registerOp(record.id, "reflecting");
     try {
-      const reflectThreshold = getMaxThreshold(this.reflectionConfig.observationTokens);
+      const reflectThreshold = getMaxThreshold(this.getEffectiveReflectionTokens(record));
       const reflectResult = await this.reflector.call(
         record.activeObservations,
         prompt,
@@ -7987,11 +8050,11 @@ ${grouped}` : grouped;
         tokenCount: reflectionTokenCount
       });
       const updatedRecord = await this.getOrCreateRecord(threadId, resourceId);
-      return { reflected: true, record: updatedRecord };
+      return { reflected: true, record: updatedRecord, usage: reflectResult.usage };
     } catch (error) {
       omError("[OM] reflect() failed", error);
       const latestRecord = await this.getOrCreateRecord(threadId, resourceId);
-      return { reflected: false, record: latestRecord };
+      return { reflected: false, record: latestRecord, usage: void 0 };
     } finally {
       await this.storage.setReflectingFlag(record.id, false);
       unregisterOp(record.id, "reflecting");
@@ -8012,12 +8075,40 @@ ${grouped}` : grouped;
     const ids = this.getStorageIds(threadId, resourceId);
     return this.storage.getObservationalMemory(ids.threadId, ids.resourceId);
   }
+  /**
+   * Update per-record config overrides for observation and/or reflection thresholds.
+   * The provided config is deep-merged into the record's `_overrides` key,
+   * so you only need to specify the fields you want to change.
+   *
+   * Overrides that violate buffering invariants (e.g. messageTokens below
+   * bufferTokens) are silently ignored at read time — the helpers fall back
+   * to the instance-level config.
+   *
+   * @example
+   * ```ts
+   * await om.updateRecordConfig('thread-1', undefined, {
+   *   observation: { messageTokens: 2000 },
+   *   reflection: { observationTokens: 8000 },
+   * });
+   * ```
+   */
+  async updateRecordConfig(threadId, resourceId, config) {
+    const ids = this.getStorageIds(threadId, resourceId);
+    const record = await this.storage.getObservationalMemory(ids.threadId, ids.resourceId);
+    if (!record) {
+      throw new Error(`No observational memory record found for thread ${ids.threadId}`);
+    }
+    await this.storage.updateObservationalMemoryConfig({
+      id: record.id,
+      config: { _overrides: config }
+    });
+  }
   /**
    * Get observation history (previous generations)
    */
-  async getHistory(threadId, resourceId, limit) {
+  async getHistory(threadId, resourceId, limit, options) {
     const ids = this.getStorageIds(threadId, resourceId);
-    return this.storage.getObservationalMemoryHistory(ids.threadId, ids.resourceId, limit);
+    return this.storage.getObservationalMemoryHistory(ids.threadId, ids.resourceId, limit, options);
   }
   /**
    * Clear all memory for a specific thread/resource
@@ -8563,5 +8654,5 @@ function getObservationsAsOf(activeObservations, asOf) {
 }
 
 export { ModelByInputTokens, OBSERVER_SYSTEM_PROMPT, ObservationalMemory, ObservationalMemoryProcessor, TokenCounter, buildObserverPrompt, buildObserverSystemPrompt, combineObservationGroupRanges, deriveObservationGroupProvenance, extractCurrentTask, formatMessagesForObserver, formatToolResultForObserver, getObservationsAsOf, hasCurrentTaskSection, injectAnchorIds, optimizeObservationsForContext, parseAnchorId, parseObservationGroups, parseObserverOutput, reconcileObservationGroupsFromReflection, renderObservationGroupsForReflection, resolveToolResultValue, stripEphemeralAnchorIds, stripObservationGroups, truncateStringByTokens, wrapInObservationGroup };
-//# sourceMappingURL=chunk-C7PARRAD.js.map
-//# sourceMappingURL=chunk-C7PARRAD.js.map
+//# sourceMappingURL=chunk-GXDPND6K.js.map
+//# sourceMappingURL=chunk-GXDPND6K.js.map