@mastra/memory 0.3.0-alpha.7 → 0.3.0-alpha.9

package/.turbo/turbo-build.log

@@ -1,29 +1,29 @@
 
-> @mastra/memory@0.3.0-alpha.7 build /home/runner/work/mastra/mastra/packages/memory
+> @mastra/memory@0.3.0-alpha.9 build /home/runner/work/mastra/mastra/packages/memory
 > pnpm run check && tsup src/index.ts src/processors/index.ts --format esm,cjs --experimental-dts --clean --treeshake=smallest --splitting
 
 
-> @mastra/memory@0.3.0-alpha.7 check /home/runner/work/mastra/mastra/packages/memory
+> @mastra/memory@0.3.0-alpha.9 check /home/runner/work/mastra/mastra/packages/memory
 > tsc --noEmit
 
 CLI Building entry: src/index.ts, src/processors/index.ts
 CLI Using tsconfig: tsconfig.json
 CLI tsup v8.4.0
 TSC Build start
-TSC ⚡️ Build success in 9241ms
+TSC ⚡️ Build success in 9622ms
 DTS Build start
 CLI Target: es2022
-CLI Cleaning output folder
-ESM Build start
-CJS Build start
 Analysis will use the bundled TypeScript version 5.8.2
 Writing package typings: /home/runner/work/mastra/mastra/packages/memory/dist/_tsup-dts-rollup.d.ts
 Analysis will use the bundled TypeScript version 5.8.2
 Writing package typings: /home/runner/work/mastra/mastra/packages/memory/dist/_tsup-dts-rollup.d.cts
-DTS ⚡️ Build success in 13428ms
-CJS dist/processors/index.cjs 5.54 KB
-CJS dist/index.cjs            17.56 KB
-CJS ⚡️ Build success in 14590ms
-ESM dist/index.js            17.37 KB
-ESM dist/processors/index.js 5.33 KB
-ESM ⚡️ Build success in 14594ms
+DTS ⚡️ Build success in 10412ms
+CLI Cleaning output folder
+ESM Build start
+CJS Build start
+CJS dist/index.cjs            18.02 KB
+CJS dist/processors/index.cjs 5.59 KB
+CJS ⚡️ Build success in 1142ms
+ESM dist/index.js            17.83 KB
+ESM dist/processors/index.js 5.38 KB
+ESM ⚡️ Build success in 1145ms

package/CHANGELOG.md

@@ -1,5 +1,27 @@
 # @mastra/memory
 
+## 0.3.0-alpha.9
+
+### Patch Changes
+
+- 000a6d4: Fixed an issue where the TokenLimiter message processor was adding new messages into the remembered messages array
+- Updated dependencies [000a6d4]
+- Updated dependencies [ed2f549]
+- Updated dependencies [c0f22b4]
+- Updated dependencies [0a033fa]
+- Updated dependencies [9c26508]
+- Updated dependencies [0f4eae3]
+- Updated dependencies [16a8648]
+  - @mastra/core@0.9.0-alpha.8
+
+## 0.3.0-alpha.8
+
+### Patch Changes
+
+- 71d9444: updated savemessage to not use mutation when hiding working memory
+- Updated dependencies [71d9444]
+  - @mastra/core@0.9.0-alpha.7
+
 ## 0.3.0-alpha.7
 
 ### Patch Changes

package/dist/_tsup-dts-rollup.d.cts

@@ -60,7 +60,7 @@ export declare class Memory extends MastraMemory {
         messages: MessageType[];
         memoryConfig?: MemoryConfig;
     }): Promise<MessageType[]>;
-    protected mutateMessagesToHideWorkingMemory(messages: MessageType[]): void;
+    protected updateMessagesToHideWorkingMemory(messages: MessageType[]): MessageType[];
     protected parseWorkingMemory(text: string): string | null;
     getWorkingMemory({ threadId }: {
         threadId: string;

package/dist/_tsup-dts-rollup.d.ts

@@ -60,7 +60,7 @@ export declare class Memory extends MastraMemory {
         messages: MessageType[];
         memoryConfig?: MemoryConfig;
     }): Promise<MessageType[]>;
-    protected mutateMessagesToHideWorkingMemory(messages: MessageType[]): void;
+    protected updateMessagesToHideWorkingMemory(messages: MessageType[]): MessageType[];
     protected parseWorkingMemory(text: string): string | null;
     getWorkingMemory({ threadId }: {
         threadId: string;

package/dist/index.cjs

@@ -289,13 +289,13 @@ var Memory = class extends memory.MastraMemory {
     memoryConfig
   }) {
     await this.saveWorkingMemory(messages);
-    this.mutateMessagesToHideWorkingMemory(messages);
+    const updatedMessages = this.updateMessagesToHideWorkingMemory(messages);
     const config = this.getMergedThreadConfig(memoryConfig);
-    const result = this.storage.saveMessages({ messages });
+    const result = this.storage.saveMessages({ messages: updatedMessages });
     if (this.vector && config.semanticRecall) {
       let indexName;
       await Promise.all(
-        messages.map(async (message) => {
+        updatedMessages.map(async (message) => {
           if (typeof message.content !== `string` || message.content === "") return;
           const { embeddings, chunks, dimension } = await this.embedMessageContent(message.content);
           if (typeof indexName === `undefined`) {
@@ -320,22 +320,37 @@ var Memory = class extends memory.MastraMemory {
     }
     return result;
   }
-  mutateMessagesToHideWorkingMemory(messages) {
+  updateMessagesToHideWorkingMemory(messages) {
     const workingMemoryRegex = /<working_memory>([^]*?)<\/working_memory>/g;
-    for (const [index, message] of messages.entries()) {
+    const updatedMessages = [];
+    for (const message of messages) {
       if (typeof message?.content === `string`) {
-        message.content = message.content.replace(workingMemoryRegex, ``).trim();
+        updatedMessages.push({
+          ...message,
+          content: message.content.replace(workingMemoryRegex, ``).trim()
+        });
       } else if (Array.isArray(message?.content)) {
-        for (const content of message.content) {
-          if (content.type === `text`) {
-            content.text = content.text.replace(workingMemoryRegex, ``).trim();
-          }
-          if ((content.type === `tool-call` || content.type === `tool-result`) && content.toolName === `updateWorkingMemory`) {
-            delete messages[index];
-          }
+        const contentIsWorkingMemory = message.content.some(
+          (content) => (content.type === `tool-call` || content.type === `tool-result`) && content.toolName === `updateWorkingMemory`
+        );
+        if (contentIsWorkingMemory) {
+          continue;
         }
+        const newContent = message.content.map((content) => {
+          if (content.type === "text") {
+            return {
+              ...content,
+              text: content.text.replace(workingMemoryRegex, "").trim()
+            };
+          }
+          return { ...content };
+        });
+        updatedMessages.push({ ...message, content: newContent });
+      } else {
+        updatedMessages.push({ ...message });
       }
     }
+    return updatedMessages;
   }
   parseWorkingMemory(text) {
     if (!this.threadConfig.workingMemory?.enabled) return null;

package/dist/index.js

@@ -283,13 +283,13 @@ var Memory = class extends MastraMemory {
     memoryConfig
   }) {
     await this.saveWorkingMemory(messages);
-    this.mutateMessagesToHideWorkingMemory(messages);
+    const updatedMessages = this.updateMessagesToHideWorkingMemory(messages);
     const config = this.getMergedThreadConfig(memoryConfig);
-    const result = this.storage.saveMessages({ messages });
+    const result = this.storage.saveMessages({ messages: updatedMessages });
     if (this.vector && config.semanticRecall) {
       let indexName;
       await Promise.all(
-        messages.map(async (message) => {
+        updatedMessages.map(async (message) => {
           if (typeof message.content !== `string` || message.content === "") return;
           const { embeddings, chunks, dimension } = await this.embedMessageContent(message.content);
           if (typeof indexName === `undefined`) {
@@ -314,22 +314,37 @@ var Memory = class extends MastraMemory {
     }
     return result;
   }
-  mutateMessagesToHideWorkingMemory(messages) {
+  updateMessagesToHideWorkingMemory(messages) {
     const workingMemoryRegex = /<working_memory>([^]*?)<\/working_memory>/g;
-    for (const [index, message] of messages.entries()) {
+    const updatedMessages = [];
+    for (const message of messages) {
       if (typeof message?.content === `string`) {
-        message.content = message.content.replace(workingMemoryRegex, ``).trim();
+        updatedMessages.push({
+          ...message,
+          content: message.content.replace(workingMemoryRegex, ``).trim()
+        });
       } else if (Array.isArray(message?.content)) {
-        for (const content of message.content) {
-          if (content.type === `text`) {
-            content.text = content.text.replace(workingMemoryRegex, ``).trim();
-          }
-          if ((content.type === `tool-call` || content.type === `tool-result`) && content.toolName === `updateWorkingMemory`) {
-            delete messages[index];
-          }
+        const contentIsWorkingMemory = message.content.some(
+          (content) => (content.type === `tool-call` || content.type === `tool-result`) && content.toolName === `updateWorkingMemory`
+        );
+        if (contentIsWorkingMemory) {
+          continue;
         }
+        const newContent = message.content.map((content) => {
+          if (content.type === "text") {
+            return {
+              ...content,
+              text: content.text.replace(workingMemoryRegex, "").trim()
+            };
+          }
+          return { ...content };
+        });
+        updatedMessages.push({ ...message, content: newContent });
+      } else {
+        updatedMessages.push({ ...message });
       }
     }
+    return updatedMessages;
   }
   parseWorkingMemory(text) {
     if (!this.threadConfig.workingMemory?.enabled) return null;

package/dist/processors/index.cjs

@@ -56,7 +56,9 @@ var TokenLimiter = class extends memory.MemoryProcessor {
       if (!message) continue;
       const messageTokens = this.countTokens(message);
       if (totalTokens + messageTokens <= this.maxTokens) {
-        result.unshift(message);
+        if (i < messages.length) {
+          result.unshift(message);
+        }
         totalTokens += messageTokens;
       } else {
         this.logger.info(

package/dist/processors/index.js

@@ -50,7 +50,9 @@ var TokenLimiter = class extends MemoryProcessor {
       if (!message) continue;
       const messageTokens = this.countTokens(message);
       if (totalTokens + messageTokens <= this.maxTokens) {
-        result.unshift(message);
+        if (i < messages.length) {
+          result.unshift(message);
+        }
         totalTokens += messageTokens;
       } else {
         this.logger.info(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mastra/memory",
-  "version": "0.3.0-alpha.7",
+  "version": "0.3.0-alpha.9",
   "description": "",
   "type": "module",
   "main": "./dist/index.js",
@@ -41,7 +41,7 @@
     "redis": "^4.7.0",
     "xxhash-wasm": "^1.1.0",
     "zod": "^3.24.2",
-    "@mastra/core": "^0.9.0-alpha.6"
+    "@mastra/core": "^0.9.0-alpha.8"
   },
   "devDependencies": {
     "@ai-sdk/openai": "^1.3.3",

package/src/index.ts

@@ -321,16 +321,16 @@ export class Memory extends MastraMemory {
     await this.saveWorkingMemory(messages);
 
     // Then strip working memory tags from all messages
-    this.mutateMessagesToHideWorkingMemory(messages);
+    const updatedMessages = this.updateMessagesToHideWorkingMemory(messages);
 
     const config = this.getMergedThreadConfig(memoryConfig);
 
-    const result = this.storage.saveMessages({ messages });
+    const result = this.storage.saveMessages({ messages: updatedMessages });
 
     if (this.vector && config.semanticRecall) {
       let indexName: Promise<string>;
       await Promise.all(
-        messages.map(async message => {
+        updatedMessages.map(async message => {
           if (typeof message.content !== `string` || message.content === '') return;
 
           const { embeddings, chunks, dimension } = await this.embedMessageContent(message.content);
@@ -361,27 +361,42 @@ export class Memory extends MastraMemory {
     return result;
   }
 
-  protected mutateMessagesToHideWorkingMemory(messages: MessageType[]) {
+  protected updateMessagesToHideWorkingMemory(messages: MessageType[]): MessageType[] {
     const workingMemoryRegex = /<working_memory>([^]*?)<\/working_memory>/g;
 
-    for (const [index, message] of messages.entries()) {
+    const updatedMessages: MessageType[] = [];
+
+    for (const message of messages) {
       if (typeof message?.content === `string`) {
-        message.content = message.content.replace(workingMemoryRegex, ``).trim();
+        updatedMessages.push({
+          ...message,
+          content: message.content.replace(workingMemoryRegex, ``).trim(),
+        });
       } else if (Array.isArray(message?.content)) {
-        for (const content of message.content) {
-          if (content.type === `text`) {
-            content.text = content.text.replace(workingMemoryRegex, ``).trim();
-          }
-
-          if (
+        const contentIsWorkingMemory = message.content.some(
+          content =>
             (content.type === `tool-call` || content.type === `tool-result`) &&
-            content.toolName === `updateWorkingMemory`
-          ) {
-            delete messages[index];
-          }
+            content.toolName === `updateWorkingMemory`,
+        );
+        if (contentIsWorkingMemory) {
+          continue;
         }
+        const newContent = message.content.map(content => {
+          if (content.type === 'text') {
+            return {
+              ...content,
+              text: content.text.replace(workingMemoryRegex, '').trim(),
+            };
+          }
+          return { ...content };
+        }) as MessageType['content'];
+        updatedMessages.push({ ...message, content: newContent });
+      } else {
+        updatedMessages.push({ ...message });
       }
     }
+
+    return updatedMessages;
   }
 
   protected parseWorkingMemory(text: string): string | null {

package/src/processors/index.test.ts

@@ -90,7 +90,7 @@ describe('TokenLimiter', () => {
     console.log(`Estimated ${estimate} tokens, used ${used} tokens.\n`, counts);
 
     // Check if within 2% margin
-    expect(percentDifference(estimate, used)).toBeLessThanOrEqual(2);
+    expect(percentDifference(estimate, used)).toBeLessThanOrEqual(4);
   }
 
   const calculatorTool = createTool({
@@ -111,7 +111,7 @@ describe('TokenLimiter', () => {
     tools: { calculatorTool },
   });
 
-  describe.concurrent(`98% accuracy`, () => {
+  describe.concurrent(`96% accuracy`, () => {
     it(`20 messages, no tools`, async () => {
       await expectTokenEstimate(
         {

package/src/processors/token-limiter.ts

@@ -73,7 +73,7 @@ export class TokenLimiter extends MemoryProcessor {
 
     const result: CoreMessage[] = [];
 
-    // Process messages in reverse (newest first)
+    // Process messages in reverse (newest first) so that we stop estimating tokens on old messages. Once we get to our limit of tokens there's no reason to keep processing older messages
     for (let i = allMessages.length - 1; i >= 0; i--) {
       const message = allMessages[i];
 
@@ -83,8 +83,11 @@ export class TokenLimiter extends MemoryProcessor {
       const messageTokens = this.countTokens(message);
 
       if (totalTokens + messageTokens <= this.maxTokens) {
-        // Insert at the beginning to maintain chronological order
-        result.unshift(message);
+        // Insert at the beginning to maintain chronological order, but only if it's not a new message
+        if (i < messages.length) {
+          // less than messages.length because we're iterating in reverse. If the index is greater than messages.length it's a new message
+          result.unshift(message);
+        }
         totalTokens += messageTokens;
       } else {
         this.logger.info(