@mastra/memory 1.6.3-alpha.0 → 1.7.0

package/CHANGELOG.md

@@ -1,5 +1,33 @@
 # @mastra/memory
 
+## 1.7.0
+
+### Minor Changes
+
+- Observational Memory now performs local threshold checks with lower CPU and memory overhead. ([#14178](https://github.com/mastra-ai/mastra/pull/14178))
+
+  This update keeps the same multimodal thresholding behavior for image-aware inputs, so existing Observational Memory configurations continue to work without changes.
+
+### Patch Changes
+
+- Updated dependencies [[`709362d`](https://github.com/mastra-ai/mastra/commit/709362d67b80d8832729bbf9e449cad27640a5d2), [`cddf895`](https://github.com/mastra-ai/mastra/commit/cddf895532b8ee7f9fa814136ec672f53d37a9ba), [`9cede11`](https://github.com/mastra-ai/mastra/commit/9cede110abac9d93072e0521bb3c8bcafb9fdadf), [`a59f126`](https://github.com/mastra-ai/mastra/commit/a59f1269104f54726699c5cdb98c72c93606d2df), [`ed8fd75`](https://github.com/mastra-ai/mastra/commit/ed8fd75cbff03bb5e19971ddb30ab7040fc60447), [`c510833`](https://github.com/mastra-ai/mastra/commit/c5108333e8cbc19dafee5f8bfefbcb5ee935335c), [`c4c7dad`](https://github.com/mastra-ai/mastra/commit/c4c7dadfe2e4584f079f6c24bfabdb8c4981827f), [`787f3ac`](https://github.com/mastra-ai/mastra/commit/787f3ac08b3bb77413645a7ab5c447fa851708fd), [`45c3112`](https://github.com/mastra-ai/mastra/commit/45c31122666a0cc56b94727099fcb1871ed1b3f6), [`7296fcc`](https://github.com/mastra-ai/mastra/commit/7296fcc599c876a68699a71c7054a16d5aaf2337), [`00c27f9`](https://github.com/mastra-ai/mastra/commit/00c27f9080731433230a61be69c44e39a7a7b4c7), [`5e7c287`](https://github.com/mastra-ai/mastra/commit/5e7c28701f2bce795dd5c811e4c3060bf2ea2242), [`7e17d3f`](https://github.com/mastra-ai/mastra/commit/7e17d3f656fdda2aad47c4beb8c491636d70820c), [`ee19c9b`](https://github.com/mastra-ai/mastra/commit/ee19c9ba3ec3ed91feb214ad539bdc766c53bb01)]:
+  - @mastra/schema-compat@1.2.1
+  - @mastra/core@1.12.0
+
+## 1.7.0-alpha.1
+
+### Minor Changes
+
+- Observational Memory now performs local threshold checks with lower CPU and memory overhead. ([#14178](https://github.com/mastra-ai/mastra/pull/14178))
+
+  This update keeps the same multimodal thresholding behavior for image-aware inputs, so existing Observational Memory configurations continue to work without changes.
+
+### Patch Changes
+
+- Updated dependencies [[`709362d`](https://github.com/mastra-ai/mastra/commit/709362d67b80d8832729bbf9e449cad27640a5d2), [`9cede11`](https://github.com/mastra-ai/mastra/commit/9cede110abac9d93072e0521bb3c8bcafb9fdadf), [`a59f126`](https://github.com/mastra-ai/mastra/commit/a59f1269104f54726699c5cdb98c72c93606d2df), [`c510833`](https://github.com/mastra-ai/mastra/commit/c5108333e8cbc19dafee5f8bfefbcb5ee935335c), [`7296fcc`](https://github.com/mastra-ai/mastra/commit/7296fcc599c876a68699a71c7054a16d5aaf2337), [`00c27f9`](https://github.com/mastra-ai/mastra/commit/00c27f9080731433230a61be69c44e39a7a7b4c7), [`ee19c9b`](https://github.com/mastra-ai/mastra/commit/ee19c9ba3ec3ed91feb214ad539bdc766c53bb01)]:
+  - @mastra/schema-compat@1.2.1-alpha.1
+  - @mastra/core@1.12.0-alpha.1
+
 ## 1.6.3-alpha.0
 
 ### Patch Changes

package/dist/{chunk-3CM4XQJO.js → chunk-M7RAJAZ6.js}

@@ -9,8 +9,7 @@ import xxhash from 'xxhash-wasm';
 import { createHash, randomUUID } from 'crypto';
 import { AsyncLocalStorage } from 'async_hooks';
 import imageSize from 'image-size';
-import { Tiktoken } from 'js-tiktoken/lite';
-import o200k_base from 'js-tiktoken/ranks/o200k_base';
+import { estimateTokenCount } from 'tokenx';
 
 // src/processors/observational-memory/observational-memory.ts
 
@@ -1590,14 +1589,6 @@ function calculateProjectedMessageRemoval(chunks, bufferActivation, messageToken
   }
   return bestBoundaryMessageTokens;
 }
-var GLOBAL_TIKTOKEN_KEY = "__mastraTiktoken";
-function getDefaultEncoder() {
-  const cached = globalThis[GLOBAL_TIKTOKEN_KEY];
-  if (cached) return cached;
-  const encoder = new Tiktoken(o200k_base);
-  globalThis[GLOBAL_TIKTOKEN_KEY] = encoder;
-  return encoder;
-}
 var IMAGE_FILE_EXTENSIONS = /* @__PURE__ */ new Set([
   "png",
   "jpg",
@@ -1611,7 +1602,7 @@ var IMAGE_FILE_EXTENSIONS = /* @__PURE__ */ new Set([
   "heif",
   "avif"
 ]);
-var TOKEN_ESTIMATE_CACHE_VERSION = 5;
+var TOKEN_ESTIMATE_CACHE_VERSION = 6;
 var DEFAULT_IMAGE_ESTIMATOR = {
   baseTokens: 85,
   tileTokens: 170,
@@ -1678,13 +1669,8 @@ function buildEstimateKey(kind, text) {
   const payloadHash = createHash("sha1").update(text).digest("hex");
   return `${kind}:${payloadHash}`;
 }
-function resolveEncodingId(encoding) {
-  if (!encoding) return "o200k_base";
-  try {
-    return `custom:${createHash("sha1").update(JSON.stringify(encoding)).digest("hex")}`;
-  } catch {
-    return "custom:unknown";
-  }
+function resolveEstimatorId() {
+  return "tokenx";
 }
 function isTokenEstimateEntry(value) {
   if (!value || typeof value !== "object") return false;
@@ -2383,20 +2369,17 @@ async function fetchGoogleAttachmentTokenEstimate(modelId, part) {
   }
 }
 var TokenCounter = class _TokenCounter {
-  encoder;
   cacheSource;
   defaultModelContext;
   modelContextStorage = new AsyncLocalStorage();
   inFlightAttachmentCounts = /* @__PURE__ */ new Map();
   // Per-message overhead: accounts for role tokens, message framing, and separators.
-  // Empirically derived from OpenAI's token counting guide (3 tokens per message base +
-  // fractional overhead from name/role encoding). 3.8 is a practical average across models.
+  // 3.8 remains a practical average across providers for OM thresholding.
   static TOKENS_PER_MESSAGE = 3.8;
   // Conversation-level overhead: system prompt framing, reply priming tokens, etc.
   static TOKENS_PER_CONVERSATION = 24;
-  constructor(encoding, options) {
-    this.encoder = encoding ? new Tiktoken(encoding) : getDefaultEncoder();
-    this.cacheSource = `v${TOKEN_ESTIMATE_CACHE_VERSION}:${resolveEncodingId(encoding)}`;
+  constructor(options) {
+    this.cacheSource = `v${TOKEN_ESTIMATE_CACHE_VERSION}:${resolveEstimatorId()}`;
     this.defaultModelContext = parseModelContext(options?.model);
   }
   runWithModelContext(model, fn) {
@@ -2410,7 +2393,7 @@ var TokenCounter = class _TokenCounter {
    */
   countString(text) {
     if (!text) return 0;
-    return this.encoder.encode(text, "all").length;
+    return estimateTokenCount(text);
   }
   readOrPersistPartEstimate(part, kind, payload) {
     const key = buildEstimateKey(kind, payload);
@@ -3306,7 +3289,7 @@ Async buffering is enabled by default \u2014 this opt-out is only needed when us
       ),
       instruction: config.reflection?.instruction
     };
-    this.tokenCounter = new TokenCounter(void 0, {
+    this.tokenCounter = new TokenCounter({
       model: typeof observationModel === "string" ? observationModel : void 0
     });
     this.onDebugEvent = config.onDebugEvent;
@@ -6718,5 +6701,5 @@ ${unreflectedContent}` : freshRecord.bufferedReflection;
 };
 
 export { OBSERVATIONAL_MEMORY_DEFAULTS, OBSERVATION_CONTEXT_INSTRUCTIONS, OBSERVATION_CONTEXT_PROMPT, OBSERVATION_CONTINUATION_HINT, OBSERVER_SYSTEM_PROMPT, ObservationalMemory, TokenCounter, buildObserverPrompt, buildObserverSystemPrompt, extractCurrentTask, formatMessagesForObserver, hasCurrentTaskSection, optimizeObservationsForContext, parseObserverOutput };
-//# sourceMappingURL=chunk-3CM4XQJO.js.map
-//# sourceMappingURL=chunk-3CM4XQJO.js.map
+//# sourceMappingURL=chunk-M7RAJAZ6.js.map
+//# sourceMappingURL=chunk-M7RAJAZ6.js.map