@mastra/memory 1.9.0 → 1.9.1-alpha.1

package/CHANGELOG.md

@@ -1,5 +1,23 @@
 # @mastra/memory
 
+## 1.9.1-alpha.1
+
+### Patch Changes
+
+- Updated dependencies [[`423aa6f`](https://github.com/mastra-ai/mastra/commit/423aa6fd12406de6a1cc6b68e463d30af1d790fb), [`47358d9`](https://github.com/mastra-ai/mastra/commit/47358d960bb2b931321de7e798f341ab0df81f44), [`4bb5adc`](https://github.com/mastra-ai/mastra/commit/4bb5adc05c88e3a83fe1ea5ecb9eae6e17313124), [`4bb5adc`](https://github.com/mastra-ai/mastra/commit/4bb5adc05c88e3a83fe1ea5ecb9eae6e17313124)]:
+  - @mastra/core@1.16.0-alpha.3
+  - @mastra/schema-compat@1.2.7-alpha.1
+
+## 1.9.1-alpha.0
+
+### Patch Changes
+
+- Fixed observational memory reflection compression for `google/gemini-2.5-flash` by using stronger compression guidance and starting it at a higher compression level during reflection. `google/gemini-2.5-flash` is unusually good at generating long, faithful outputs. That made reflection retries more likely to preserve too much detail and miss the compression target, wasting tokens in the process. ([#14612](https://github.com/mastra-ai/mastra/pull/14612))
+
+- Updated dependencies [[`be37de4`](https://github.com/mastra-ai/mastra/commit/be37de4391bd1d5486ce38efacbf00ca51637262), [`f3ce603`](https://github.com/mastra-ai/mastra/commit/f3ce603fd76180f4a5be90b6dc786d389b6b3e98), [`2871451`](https://github.com/mastra-ai/mastra/commit/2871451703829aefa06c4a5d6eca7fd3731222ef), [`d3930ea`](https://github.com/mastra-ai/mastra/commit/d3930eac51c30b0ecf7eaa54bb9430758b399777)]:
+  - @mastra/core@1.16.0-alpha.2
+  - @mastra/schema-compat@1.2.7-alpha.0
+
 ## 1.9.0
 
 ### Minor Changes

package/dist/{chunk-LVV2RT42.cjs → chunk-CNOHXG5O.cjs}

@@ -1736,6 +1736,7 @@ User messages are extremely important. If the user asks a question or gives a ne
 
 ${instruction}` : ""}`;
 }
+var MAX_COMPRESSION_LEVEL = 4;
 var COMPRESSION_GUIDANCE = {
   0: "",
   1: `
@@ -1748,11 +1749,11 @@ Please re-process with slightly more compression:
 - Closer to the end, retain more fine details (recent context matters more)
 - Memory is getting long - use a more condensed style throughout
 - Combine related items more aggressively but do not lose important specific details of names, places, events, and people
+- Combine repeated similar tool calls (e.g. multiple file views, searches, or edits in the same area) into a single summary line describing what was explored/changed and the outcome
 - Preserve \u2705 completion markers \u2014 they are memory signals that tell the assistant what is already resolved and help prevent repeated work
 - Preserve the concrete resolved outcome captured by \u2705 markers so the assistant knows what exactly is done
-- For example if there is a long nested observation list about repeated tool calls, you can combine those into a single line and observe that the tool was called multiple times for x reason, and finally y outcome happened.
 
-Your current detail level was a 10/10, lets aim for a 8/10 detail level.
+Aim for a 8/10 detail level.
 `,
   2: `
 ## AGGRESSIVE COMPRESSION REQUIRED
@@ -1764,12 +1765,13 @@ Please re-process with much more aggressive compression:
 - Closer to the end, retain fine details (recent context matters more)
 - Memory is getting very long - use a significantly more condensed style throughout
 - Combine related items aggressively but do not lose important specific details of names, places, events, and people
+- Combine repeated similar tool calls (e.g. multiple file views, searches, or edits in the same area) into a single summary line describing what was explored/changed and the outcome
+- If the same file or module is mentioned across many observations, merge into one entry covering the full arc
 - Preserve \u2705 completion markers \u2014 they are memory signals that tell the assistant what is already resolved and help prevent repeated work
 - Preserve the concrete resolved outcome captured by \u2705 markers so the assistant knows what exactly is done
-- For example if there is a long nested observation list about repeated tool calls, you can combine those into a single line and observe that the tool was called multiple times for x reason, and finally y outcome happened.
 - Remove redundant information and merge overlapping observations
 
-Your current detail level was a 10/10, lets aim for a 6/10 detail level.
+Aim for a 6/10 detail level.
 `,
   3: `
 ## CRITICAL COMPRESSION REQUIRED
@@ -1780,13 +1782,32 @@ Please re-process with maximum compression:
 - Summarize the oldest observations (first 50-70%) into brief high-level paragraphs \u2014 only key facts, decisions, and outcomes
 - For the most recent observations (last 30-50%), retain important details but still use a condensed style
 - Ruthlessly merge related observations \u2014 if 10 observations are about the same topic, combine into 1-2 lines
+- Combine all tool call sequences (file views, searches, edits, builds) into outcome-only summaries \u2014 drop individual steps entirely
 - Drop procedural details (tool calls, retries, intermediate steps) \u2014 keep only final outcomes
 - Drop observations that are no longer relevant or have been superseded by newer information
 - Preserve \u2705 completion markers \u2014 they are memory signals that tell the assistant what is already resolved and help prevent repeated work
 - Preserve the concrete resolved outcome captured by \u2705 markers so the assistant knows what exactly is done
 - Preserve: names, dates, decisions, errors, user preferences, and architectural choices
 
-Your current detail level was a 10/10, lets aim for a 4/10 detail level.
+Aim for a 4/10 detail level.
+`,
+  4: `
+## EXTREME COMPRESSION REQUIRED
+
+Multiple compression attempts have failed. The content may already be dense from a prior reflection.
+
+You MUST dramatically reduce the number of observations while keeping the standard observation format (date groups with bullet points and priority emojis):
+- Tool call observations are the biggest source of bloat. Collapse ALL tool call sequences into outcome-only observations \u2014 e.g. 10 observations about viewing/searching/editing files become 1 observation about what was actually learned or achieved (e.g. "Investigated auth module and found token validation was skipping expiry check")
+- Never preserve individual tool calls (viewed file X, searched for Y, ran build) \u2014 only preserve what was discovered or accomplished
+- Consolidate many related observations into single, more generic observations
+- Merge all same-day date groups into at most 2-3 date groups per day
+- For older content, each topic or task should be at most 1-2 observations capturing the key outcome
+- For recent content, retain more detail but still merge related items aggressively
+- If multiple observations describe incremental progress on the same task, keep only the final state
+- Preserve \u2705 completion markers and their outcomes but merge related completions into fewer lines
+- Preserve: user preferences, key decisions, architectural choices, and unresolved issues
+
+Aim for a 2/10 detail level. Fewer, more generic observations are better than many specific ones that exceed the budget.
 `
 };
 function buildReflectorPrompt(observations, manualPrompt, compressionLevel, skipContinuationHints) {
@@ -3912,6 +3933,22 @@ Async buffering is enabled by default \u2014 this opt-out is only needed when us
       modelId: resolved.modelId
     };
   }
+  /**
+   * Get the default compression start level based on model behavior.
+   * gemini-2.5-flash is a faithful transcriber that needs explicit pressure to compress effectively.
+   */
+  async getCompressionStartLevel(requestContext) {
+    try {
+      const resolved = await this.resolveModelContext(this.reflectionConfig.model, requestContext);
+      const modelId = resolved?.modelId ?? "";
+      if (modelId.includes("gemini-2.5-flash")) {
+        return 2;
+      }
+      return 1;
+    } catch {
+      return 1;
+    }
+  }
   getRuntimeModelContext(model) {
     if (!model?.modelId) {
       return void 0;
@@ -4709,8 +4746,9 @@ ${unreflectedContent}` : bufferedReflection;
     const originalTokens = this.tokenCounter.countObservations(observations);
     const targetThreshold = observationTokensThreshold ?? getMaxThreshold(this.reflectionConfig.observationTokens);
     let totalUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
-    let currentLevel = compressionStartLevel ?? 0;
-    const maxLevel = 3;
+    const startLevel = compressionStartLevel ?? 0;
+    let currentLevel = startLevel;
+    const maxLevel = Math.min(MAX_COMPRESSION_LEVEL, startLevel + 3);
     let parsed = { observations: "", suggestedContinuation: void 0 };
     let reflectedTokens = 0;
     let attemptNumber = 0;
@@ -4781,13 +4819,17 @@ ${unreflectedContent}` : bufferedReflection;
         omDebug(`[OM:callReflector] degenerate output persists at maxLevel=${maxLevel}, breaking`);
         break;
       }
+      if (currentLevel >= maxLevel) {
+        break;
+      }
+      const nextLevel = currentLevel + 1;
       if (streamContext?.writer) {
         const failedMarker = createObservationFailedMarker({
           cycleId: streamContext.cycleId,
           operationType: "reflection",
           startedAt: streamContext.startedAt,
           tokensAttempted: originalTokens,
-          error: `Did not compress below threshold (${originalTokens} \u2192 ${reflectedTokens}, target: ${targetThreshold}), retrying at level ${currentLevel + 1}`,
+          error: `Did not compress below threshold (${originalTokens} \u2192 ${reflectedTokens}, target: ${targetThreshold}), retrying at level ${nextLevel}`,
           recordId: streamContext.recordId,
           threadId: streamContext.threadId
         });
@@ -4808,7 +4850,7 @@ ${unreflectedContent}` : bufferedReflection;
         await streamContext.writer.custom(startMarker).catch(() => {
         });
       }
-      currentLevel = Math.min(currentLevel + 1, maxLevel);
+      currentLevel = nextLevel;
     }
     return {
       observations: parsed.observations,
@@ -6709,9 +6751,6 @@ ${bufferedObservations}`;
     omDebug(
       `[OM:reflect] doAsyncBufferedReflection: slicing observations for reflection \u2014 totalLines=${totalLines}, avgTokPerLine=${avgTokensPerLine.toFixed(1)}, activationPointTokens=${activationPointTokens}, linesToReflect=${linesToReflect}/${totalLines}, sliceTokenEstimate=${sliceTokenEstimate}, compressionTarget=${compressionTarget}`
     );
-    omDebug(
-      `[OM:reflect] doAsyncBufferedReflection: starting reflector call, recordId=${currentRecord.id}, observationTokens=${sliceTokenEstimate}, compressionTarget=${compressionTarget} (inputTokens), activeObsLength=${activeObservations.length}, reflectedLineCount=${reflectedObservationLineCount}`
-    );
     if (writer) {
       const startMarker = createBufferingStartMarker({
         cycleId,
@@ -6725,6 +6764,9 @@ ${bufferedObservations}`;
       void writer.custom(startMarker).catch(() => {
       });
     }
+    omDebug(
+      `[OM:reflect] doAsyncBufferedReflection: starting reflector call, recordId=${currentRecord.id}, observationTokens=${sliceTokenEstimate}, compressionTarget=${compressionTarget} (inputTokens), activeObsLength=${activeObservations.length}, reflectedLineCount=${reflectedObservationLineCount}`
+    );
     const reflectResult = await this.callReflector(
       activeObservations,
       void 0,
@@ -6736,8 +6778,7 @@ ${bufferedObservations}`;
       // No abort signal for background ops
       true,
       // Skip continuation hints for async buffering
-      1,
-      // Start at compression level 1 for buffered reflection
+      await this.getCompressionStartLevel(requestContext),
       requestContext
     );
     const reflectionTokenCount = this.tokenCounter.countObservations(reflectResult.observations);
@@ -7598,5 +7639,5 @@ exports.stripEphemeralAnchorIds = stripEphemeralAnchorIds;
 exports.stripObservationGroups = stripObservationGroups;
 exports.truncateStringByTokens = truncateStringByTokens;
 exports.wrapInObservationGroup = wrapInObservationGroup;
-//# sourceMappingURL=chunk-LVV2RT42.cjs.map
-//# sourceMappingURL=chunk-LVV2RT42.cjs.map
+//# sourceMappingURL=chunk-CNOHXG5O.cjs.map
+//# sourceMappingURL=chunk-CNOHXG5O.cjs.map