@mastra/memory 0.2.6-alpha.2 → 0.2.6-alpha.4

package/.turbo/turbo-build.log

@@ -1,27 +1,27 @@
 
-> @mastra/memory@0.2.6-alpha.2 build /home/runner/work/mastra/mastra/packages/memory
+> @mastra/memory@0.2.6-alpha.4 build /home/runner/work/mastra/mastra/packages/memory
 > pnpm run check && tsup src/index.ts --format esm,cjs --experimental-dts --clean --treeshake=smallest --splitting
 
 
-> @mastra/memory@0.2.6-alpha.2 check /home/runner/work/mastra/mastra/packages/memory
+> @mastra/memory@0.2.6-alpha.4 check /home/runner/work/mastra/mastra/packages/memory
 > tsc --noEmit
 
 CLI Building entry: src/index.ts
 CLI Using tsconfig: tsconfig.json
 CLI tsup v8.4.0
 TSC Build start
-TSC ⚡️ Build success in 10016ms
+TSC ⚡️ Build success in 13659ms
 DTS Build start
 CLI Target: es2022
 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 11685ms
+DTS ⚡️ Build success in 5635ms
 CLI Cleaning output folder
 ESM Build start
 CJS Build start
-ESM dist/index.js 12.85 KB
-ESM ⚡️ Build success in 605ms
-CJS dist/index.cjs 12.88 KB
-CJS ⚡️ Build success in 605ms
+ESM dist/index.js 13.53 KB
+ESM ⚡️ Build success in 315ms
+CJS dist/index.cjs 13.56 KB
+CJS ⚡️ Build success in 316ms

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # @mastra/memory
 
+## 0.2.6-alpha.4
+
+### Patch Changes
+
+- 394dfad: Removed working memory tool calls from thread history after the working memory has been updated. This is to prevent updates from polluting the context history and confusing agents. They should only see the most recent copy of working memory.
+  Also made memory.getWorkingMemory() public since it's useful for testing, debugging, and building UIs.
+
+## 0.2.6-alpha.3
+
+### Patch Changes
+
+- 05095e9: Fixed an issue where very long messages would cause Memory semantic recall to throw errors
+- Updated dependencies [b3b34f5]
+- Updated dependencies [a4686e8]
+  - @mastra/core@0.7.0-alpha.3
+  - @mastra/rag@0.1.14-alpha.3
+
 ## 0.2.6-alpha.2
 
 ### Patch Changes

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

@@ -45,13 +45,14 @@ export declare class Memory extends MastraMemory {
         metadata: Record<string, unknown>;
     }): Promise<StorageThreadType>;
     deleteThread(threadId: string): Promise<void>;
+    private embedMessageContent;
     saveMessages({ messages, memoryConfig, }: {
         messages: MessageType[];
         memoryConfig?: MemoryConfig;
     }): Promise<MessageType[]>;
     protected mutateMessagesToHideWorkingMemory(messages: MessageType[]): void;
     protected parseWorkingMemory(text: string): string | null;
-    protected getWorkingMemory({ threadId }: {
+    getWorkingMemory({ threadId }: {
         threadId: string;
     }): Promise<string | null>;
     private saveWorkingMemory;

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

@@ -45,13 +45,14 @@ export declare class Memory extends MastraMemory {
         metadata: Record<string, unknown>;
     }): Promise<StorageThreadType>;
     deleteThread(threadId: string): Promise<void>;
+    private embedMessageContent;
     saveMessages({ messages, memoryConfig, }: {
         messages: MessageType[];
         memoryConfig?: MemoryConfig;
     }): Promise<MessageType[]>;
     protected mutateMessagesToHideWorkingMemory(messages: MessageType[]): void;
     protected parseWorkingMemory(text: string): string | null;
-    protected getWorkingMemory({ threadId }: {
+    getWorkingMemory({ threadId }: {
         threadId: string;
     }): Promise<string | null>;
     private saveWorkingMemory;

package/dist/index.cjs

@@ -2,6 +2,7 @@
 
 var core = require('@mastra/core');
 var memory = require('@mastra/core/memory');
+var rag = require('@mastra/rag');
 var ai = require('ai');
 var zod = require('zod');
 
@@ -63,7 +64,7 @@ var Memory = class extends memory.MastraMemory {
     threadConfig
   }) {
     if (resourceId) await this.validateThreadIsOwnedByResource(threadId, resourceId);
-    let vectorResults = null;
+    const vectorResults = [];
     this.logger.debug(`Memory query() with:`, {
       threadId,
       selectBy,
@@ -78,19 +79,22 @@ var Memory = class extends memory.MastraMemory {
       messageRange: config?.semanticRecall?.messageRange ?? { before: 2, after: 2 }
     };
     if (config?.semanticRecall && selectBy?.vectorSearchString && this.vector && !!selectBy.vectorSearchString) {
-      const { embedding } = await ai.embed({
-        value: selectBy.vectorSearchString,
-        model: this.embedder
-      });
       const { indexName } = await this.createEmbeddingIndex();
-      vectorResults = await this.vector.query({
-        indexName,
-        queryVector: embedding,
-        topK: vectorConfig.topK,
-        filter: {
-          thread_id: threadId
-        }
-      });
+      const { embeddings } = await this.embedMessageContent(selectBy.vectorSearchString);
+      await Promise.all(
+        embeddings.map(async (embedding) => {
+          vectorResults.push(
+            ...await this.vector.query({
+              indexName,
+              queryVector: embedding,
+              topK: vectorConfig.topK,
+              filter: {
+                thread_id: threadId
+              }
+            })
+          );
+        })
+      );
     }
     const rawMessages = await this.storage.__getMessages({
       threadId,
@@ -176,6 +180,23 @@ var Memory = class extends memory.MastraMemory {
   async deleteThread(threadId) {
     await this.storage.__deleteThread({ threadId });
   }
+  async embedMessageContent(content) {
+    const doc = rag.MDocument.fromText(content);
+    const chunks = await doc.chunk({
+      strategy: "token",
+      size: 4096,
+      overlap: 20
+    });
+    const { embeddings } = await ai.embedMany({
+      values: chunks.map((chunk) => chunk.text),
+      model: this.embedder,
+      maxRetries: 3
+    });
+    return {
+      embeddings,
+      chunks
+    };
+  }
   async saveMessages({
     messages,
     memoryConfig
@@ -187,17 +208,14 @@ var Memory = class extends memory.MastraMemory {
       const { indexName } = await this.createEmbeddingIndex();
       for (const message of messages) {
         if (typeof message.content !== `string` || message.content === "") continue;
-        const { embedding } = await ai.embed({ value: message.content, model: this.embedder, maxRetries: 3 });
+        const { embeddings, chunks } = await this.embedMessageContent(message.content);
         await this.vector.upsert({
           indexName,
-          vectors: [embedding],
-          metadata: [
-            {
-              text: message.content,
-              message_id: message.id,
-              thread_id: message.threadId
-            }
-          ]
+          vectors: embeddings,
+          metadata: chunks.map(() => ({
+            message_id: message.id,
+            thread_id: message.threadId
+          }))
         });
       }
     }
@@ -205,7 +223,7 @@ var Memory = class extends memory.MastraMemory {
   }
   mutateMessagesToHideWorkingMemory(messages) {
     const workingMemoryRegex = /<working_memory>([^]*?)<\/working_memory>/g;
-    for (const message of messages) {
+    for (const [index, message] of messages.entries()) {
       if (typeof message?.content === `string`) {
         message.content = message.content.replace(workingMemoryRegex, ``).trim();
       } else if (Array.isArray(message?.content)) {
@@ -213,6 +231,9 @@ var Memory = class extends memory.MastraMemory {
           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];
+          }
         }
       }
     }

package/dist/index.js

@@ -1,6 +1,7 @@
 import { deepMerge } from '@mastra/core';
 import { MastraMemory } from '@mastra/core/memory';
-import { embed } from 'ai';
+import { MDocument } from '@mastra/rag';
+import { embedMany } from 'ai';
 import { z } from 'zod';
 
 // src/index.ts
@@ -61,7 +62,7 @@ var Memory = class extends MastraMemory {
     threadConfig
   }) {
     if (resourceId) await this.validateThreadIsOwnedByResource(threadId, resourceId);
-    let vectorResults = null;
+    const vectorResults = [];
     this.logger.debug(`Memory query() with:`, {
       threadId,
       selectBy,
@@ -76,19 +77,22 @@ var Memory = class extends MastraMemory {
       messageRange: config?.semanticRecall?.messageRange ?? { before: 2, after: 2 }
     };
     if (config?.semanticRecall && selectBy?.vectorSearchString && this.vector && !!selectBy.vectorSearchString) {
-      const { embedding } = await embed({
-        value: selectBy.vectorSearchString,
-        model: this.embedder
-      });
       const { indexName } = await this.createEmbeddingIndex();
-      vectorResults = await this.vector.query({
-        indexName,
-        queryVector: embedding,
-        topK: vectorConfig.topK,
-        filter: {
-          thread_id: threadId
-        }
-      });
+      const { embeddings } = await this.embedMessageContent(selectBy.vectorSearchString);
+      await Promise.all(
+        embeddings.map(async (embedding) => {
+          vectorResults.push(
+            ...await this.vector.query({
+              indexName,
+              queryVector: embedding,
+              topK: vectorConfig.topK,
+              filter: {
+                thread_id: threadId
+              }
+            })
+          );
+        })
+      );
     }
     const rawMessages = await this.storage.__getMessages({
       threadId,
@@ -174,6 +178,23 @@ var Memory = class extends MastraMemory {
   async deleteThread(threadId) {
     await this.storage.__deleteThread({ threadId });
   }
+  async embedMessageContent(content) {
+    const doc = MDocument.fromText(content);
+    const chunks = await doc.chunk({
+      strategy: "token",
+      size: 4096,
+      overlap: 20
+    });
+    const { embeddings } = await embedMany({
+      values: chunks.map((chunk) => chunk.text),
+      model: this.embedder,
+      maxRetries: 3
+    });
+    return {
+      embeddings,
+      chunks
+    };
+  }
   async saveMessages({
     messages,
     memoryConfig
@@ -185,17 +206,14 @@ var Memory = class extends MastraMemory {
       const { indexName } = await this.createEmbeddingIndex();
       for (const message of messages) {
         if (typeof message.content !== `string` || message.content === "") continue;
-        const { embedding } = await embed({ value: message.content, model: this.embedder, maxRetries: 3 });
+        const { embeddings, chunks } = await this.embedMessageContent(message.content);
         await this.vector.upsert({
           indexName,
-          vectors: [embedding],
-          metadata: [
-            {
-              text: message.content,
-              message_id: message.id,
-              thread_id: message.threadId
-            }
-          ]
+          vectors: embeddings,
+          metadata: chunks.map(() => ({
+            message_id: message.id,
+            thread_id: message.threadId
+          }))
         });
       }
     }
@@ -203,7 +221,7 @@ var Memory = class extends MastraMemory {
   }
   mutateMessagesToHideWorkingMemory(messages) {
     const workingMemoryRegex = /<working_memory>([^]*?)<\/working_memory>/g;
-    for (const message of messages) {
+    for (const [index, message] of messages.entries()) {
       if (typeof message?.content === `string`) {
         message.content = message.content.replace(workingMemoryRegex, ``).trim();
       } else if (Array.isArray(message?.content)) {
@@ -211,6 +229,9 @@ var Memory = class extends MastraMemory {
           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];
+          }
         }
       }
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mastra/memory",
-  "version": "0.2.6-alpha.2",
+  "version": "0.2.6-alpha.4",
   "description": "",
   "type": "module",
   "main": "./dist/index.js",
@@ -49,7 +49,8 @@
     "postgres": "^3.4.5",
     "redis": "^4.7.0",
     "zod": "^3.24.2",
-    "@mastra/core": "^0.7.0-alpha.2"
+    "@mastra/core": "^0.7.0-alpha.3",
+    "@mastra/rag": "^0.1.14-alpha.3"
   },
   "devDependencies": {
     "@microsoft/api-extractor": "^7.52.1",

package/src/index.ts

@@ -3,7 +3,8 @@ import type { AiMessageType, CoreMessage, CoreTool } from '@mastra/core';
 import { MastraMemory } from '@mastra/core/memory';
 import type { MessageType, MemoryConfig, SharedMemoryConfig, StorageThreadType } from '@mastra/core/memory';
 import type { StorageGetMessagesArg } from '@mastra/core/storage';
-import { embed } from 'ai';
+import { MDocument } from '@mastra/rag';
+import { embedMany } from 'ai';
 import { updateWorkingMemoryTool } from './tools/working-memory';
 
 /**
@@ -43,14 +44,12 @@ export class Memory extends MastraMemory {
   }: StorageGetMessagesArg): Promise<{ messages: CoreMessage[]; uiMessages: AiMessageType[] }> {
     if (resourceId) await this.validateThreadIsOwnedByResource(threadId, resourceId);
 
-    let vectorResults:
-      | null
-      | {
-          id: string;
-          score: number;
-          metadata?: Record<string, any>;
-          vector?: number[];
-        }[] = null;
+    const vectorResults: {
+      id: string;
+      score: number;
+      metadata?: Record<string, any>;
+      vector?: number[];
+    }[] = [];
 
     this.logger.debug(`Memory query() with:`, {
       threadId,
@@ -72,21 +71,23 @@ export class Memory extends MastraMemory {
           };
 
     if (config?.semanticRecall && selectBy?.vectorSearchString && this.vector && !!selectBy.vectorSearchString) {
-      const { embedding } = await embed({
-        value: selectBy.vectorSearchString,
-        model: this.embedder,
-      });
-
       const { indexName } = await this.createEmbeddingIndex();
-
-      vectorResults = await this.vector.query({
-        indexName,
-        queryVector: embedding,
-        topK: vectorConfig.topK,
-        filter: {
-          thread_id: threadId,
-        },
-      });
+      const { embeddings } = await this.embedMessageContent(selectBy.vectorSearchString);
+
+      await Promise.all(
+        embeddings.map(async embedding => {
+          vectorResults.push(
+            ...(await this.vector.query({
+              indexName,
+              queryVector: embedding,
+              topK: vectorConfig.topK,
+              filter: {
+                thread_id: threadId,
+              },
+            })),
+          );
+        }),
+      );
     }
 
     // Get raw messages from storage
@@ -220,6 +221,27 @@ export class Memory extends MastraMemory {
     // }
   }
 
+  private async embedMessageContent(content: string) {
+    const doc = MDocument.fromText(content);
+
+    const chunks = await doc.chunk({
+      strategy: 'token',
+      size: 4096,
+      overlap: 20,
+    });
+
+    const { embeddings } = await embedMany({
+      values: chunks.map(chunk => chunk.text),
+      model: this.embedder,
+      maxRetries: 3,
+    });
+
+    return {
+      embeddings,
+      chunks,
+    };
+  }
+
   async saveMessages({
     messages,
     memoryConfig,
@@ -240,17 +262,16 @@ export class Memory extends MastraMemory {
 
       for (const message of messages) {
         if (typeof message.content !== `string` || message.content === '') continue;
-        const { embedding } = await embed({ value: message.content, model: this.embedder, maxRetries: 3 });
+
+        const { embeddings, chunks } = await this.embedMessageContent(message.content);
+
         await this.vector.upsert({
           indexName,
-          vectors: [embedding],
-          metadata: [
-            {
-              text: message.content,
-              message_id: message.id,
-              thread_id: message.threadId,
-            },
-          ],
+          vectors: embeddings,
+          metadata: chunks.map(() => ({
+            message_id: message.id,
+            thread_id: message.threadId,
+          })),
         });
       }
     }
@@ -260,7 +281,8 @@ export class Memory extends MastraMemory {
 
   protected mutateMessagesToHideWorkingMemory(messages: MessageType[]) {
     const workingMemoryRegex = /<working_memory>([^]*?)<\/working_memory>/g;
-    for (const message of messages) {
+
+    for (const [index, message] of messages.entries()) {
       if (typeof message?.content === `string`) {
         message.content = message.content.replace(workingMemoryRegex, ``).trim();
       } else if (Array.isArray(message?.content)) {
@@ -268,6 +290,13 @@ export class Memory extends MastraMemory {
           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];
+          }
         }
       }
     }
@@ -287,7 +316,7 @@ export class Memory extends MastraMemory {
     return null;
   }
 
-  protected async getWorkingMemory({ threadId }: { threadId: string }): Promise<string | null> {
+  public async getWorkingMemory({ threadId }: { threadId: string }): Promise<string | null> {
     if (!this.threadConfig.workingMemory?.enabled) return null;
 
     // Get thread from storage