@mastra/memory 0.3.2-alpha.5 → 0.3.2-alpha.6

package/.turbo/turbo-build.log

@@ -1,29 +1,29 @@
 
-> @mastra/memory@0.3.2-alpha.5 build /home/runner/work/mastra/mastra/packages/memory
+> @mastra/memory@0.3.2-alpha.6 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@0.3.2-alpha.5 check /home/runner/work/mastra/mastra/packages/memory
+> @mastra/memory@0.3.2-alpha.6 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 8312ms
+TSC ⚡️ Build success in 9693ms
 DTS Build start
 CLI Target: es2022
 Analysis will use the bundled TypeScript version 5.8.3
 Writing package typings: /home/runner/work/mastra/mastra/packages/memory/dist/_tsup-dts-rollup.d.ts
 Analysis will use the bundled TypeScript version 5.8.3
 Writing package typings: /home/runner/work/mastra/mastra/packages/memory/dist/_tsup-dts-rollup.d.cts
-DTS ⚡️ Build success in 6581ms
+DTS ⚡️ Build success in 11004ms
 CLI Cleaning output folder
 ESM Build start
 CJS Build start
 ESM dist/index.js            18.27 KB
 ESM dist/processors/index.js 5.38 KB
-ESM ⚡️ Build success in 393ms
-CJS dist/index.cjs            18.45 KB
+ESM ⚡️ Build success in 718ms
 CJS dist/processors/index.cjs 5.59 KB
-CJS ⚡️ Build success in 394ms
+CJS dist/index.cjs            18.45 KB
+CJS ⚡️ Build success in 719ms

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # @mastra/memory
 
+## 0.3.2-alpha.6
+
+### Patch Changes
+
+- 544767d: Improved token estimation in TokenLimiter from 96% accuracy back to 99%
+
 ## 0.3.2-alpha.5
 
 ### Patch Changes

package/dist/processors/index.cjs

@@ -16,9 +16,9 @@ var TokenLimiter = class extends memory.MemoryProcessor {
   // Token overheads per OpenAI's documentation
   // See: https://cookbook.openai.com/examples/how_to_count_tokens_with_tiktoken#6-counting-tokens-for-chat-completions-api-calls
   // Every message follows <|start|>{role/name}\n{content}<|end|>
-  TOKENS_PER_MESSAGE = 3;
+  TOKENS_PER_MESSAGE = 3.8;
   // tokens added for each message (start & end tokens)
-  TOKENS_PER_TOOL = 2;
+  TOKENS_PER_TOOL = 2.2;
   // empirical adjustment for tool calls
   TOKENS_PER_CONVERSATION = 25;
   // fixed overhead for the conversation

package/dist/processors/index.js

@@ -10,9 +10,9 @@ var TokenLimiter = class extends MemoryProcessor {
   // Token overheads per OpenAI's documentation
   // See: https://cookbook.openai.com/examples/how_to_count_tokens_with_tiktoken#6-counting-tokens-for-chat-completions-api-calls
   // Every message follows <|start|>{role/name}\n{content}<|end|>
-  TOKENS_PER_MESSAGE = 3;
+  TOKENS_PER_MESSAGE = 3.8;
   // tokens added for each message (start & end tokens)
-  TOKENS_PER_TOOL = 2;
+  TOKENS_PER_TOOL = 2.2;
   // empirical adjustment for tool calls
   TOKENS_PER_CONVERSATION = 25;
   // fixed overhead for the conversation

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mastra/memory",
-  "version": "0.3.2-alpha.5",
+  "version": "0.3.2-alpha.6",
   "description": "",
   "type": "module",
   "main": "./dist/index.js",

package/src/processors/index.test.ts

@@ -72,7 +72,7 @@ describe('TokenLimiter', () => {
       estimatedTokens += testLimiter.countTokens(message as CoreMessage);
     }
 
-    return estimatedTokens;
+    return Number(estimatedTokens.toFixed(2));
   }
 
   function percentDifference(a: number, b: number) {
@@ -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(4);
+    expect(percentDifference(estimate, used)).toBeLessThanOrEqual(2);
   }
 
   const calculatorTool = createTool({
@@ -111,7 +111,7 @@ describe('TokenLimiter', () => {
     tools: { calculatorTool },
   });
 
-  describe.concurrent(`96% accuracy`, () => {
+  describe.concurrent(`98% accuracy`, () => {
     it(`20 messages, no tools`, async () => {
       await expectTokenEstimate(
         {
@@ -134,10 +134,10 @@ describe('TokenLimiter', () => {
       );
     });
 
-    it(`4 messages, 0 tools`, async () => {
+    it(`20 messages, 0 tools`, async () => {
       await expectTokenEstimate(
         {
-          messageCount: 2,
+          messageCount: 10,
           toolFrequency: 0,
           threadId: '3',
         },

package/src/processors/token-limiter.ts

@@ -25,8 +25,8 @@ export class TokenLimiter extends MemoryProcessor {
   // Token overheads per OpenAI's documentation
   // See: https://cookbook.openai.com/examples/how_to_count_tokens_with_tiktoken#6-counting-tokens-for-chat-completions-api-calls
   // Every message follows <|start|>{role/name}\n{content}<|end|>
-  public TOKENS_PER_MESSAGE = 3; // tokens added for each message (start & end tokens)
-  public TOKENS_PER_TOOL = 2; // empirical adjustment for tool calls
+  public TOKENS_PER_MESSAGE = 3.8; // tokens added for each message (start & end tokens)
+  public TOKENS_PER_TOOL = 2.2; // empirical adjustment for tool calls
   public TOKENS_PER_CONVERSATION = 25; // fixed overhead for the conversation
 
   /**