@mastra/memory 1.19.1-alpha.0 → 1.20.0-alpha.2

package/CHANGELOG.md

@@ -1,5 +1,31 @@
 # @mastra/memory
 
+## 1.20.0-alpha.2
+
+### Patch Changes
+
+- Removed zod as a required peer dependency. Internal schemas now use plain JSON Schema objects instead of zod runtime. ([#16726](https://github.com/mastra-ai/mastra/pull/16726))
+
+- Updated dependencies [[`c35b962`](https://github.com/mastra-ai/mastra/commit/c35b9625c7e854fcfdeee226a3338a750d0ff211), [`4084113`](https://github.com/mastra-ai/mastra/commit/408411370fc48a822e8b616b3b63f9409774e0e9)]:
+  - @mastra/core@1.37.0-alpha.8
+
+## 1.20.0-alpha.1
+
+### Minor Changes
+
+- Add per-provider capability files and `auto` mode for `observeAttachments` ([#16922](https://github.com/mastra-ai/mastra/pull/16922))
+  - Generate per-provider capability files (e.g. `capabilities/openai.json`) alongside the model router registry, sourced from models.dev API
+  - Export `modelSupportsAttachments(modelRouterId)` from `@mastra/core/llm` to check whether a model supports image/file attachments
+  - Extend `observeAttachments` config to accept `'auto'` in addition to `boolean | string[]`
+  - When set to `'auto'`, the observer resolves the model (including function-based models) and checks the capability registry before deciding to forward or drop attachment parts
+
+### Patch Changes
+
+- Fixed observational memory so in-progress messages are not removed during observation cleanup. ([#16913](https://github.com/mastra-ai/mastra/pull/16913))
+
+- Updated dependencies [[`ac442a4`](https://github.com/mastra-ai/mastra/commit/ac442a42fda0354ac2bcea772bf6691cb3e9dbb3), [`1e5c067`](https://github.com/mastra-ai/mastra/commit/1e5c067d2e20a781af670578180d1ee249806d41), [`008baaf`](https://github.com/mastra-ai/mastra/commit/008baafd8d851f831407045aebead5a2e3342eff), [`8116436`](https://github.com/mastra-ai/mastra/commit/81164363eb225d774e41ff27da6a5ea611406688), [`c27c4b9`](https://github.com/mastra-ai/mastra/commit/c27c4b9f137df5414fca4e45896aceccff6b0ed5), [`08b3b59`](https://github.com/mastra-ai/mastra/commit/08b3b590dd960dee6c9a6e39272f8927d803db6e)]:
+  - @mastra/core@1.37.0-alpha.3
+
 ## 1.19.1-alpha.0
 
 ### Patch Changes

package/dist/{chunk-D4J4XPGM.cjs → chunk-3Q3LOVQP.cjs}

@@ -1,5 +1,11 @@
 'use strict';
 
+var __defProp = Object.defineProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+
 // src/processors/observational-memory/constants.ts
 var OBSERVATIONAL_MEMORY_DEFAULTS = {
   observation: {
@@ -107,5 +113,6 @@ exports.OBSERVATION_CONTEXT_INSTRUCTIONS = OBSERVATION_CONTEXT_INSTRUCTIONS;
 exports.OBSERVATION_CONTEXT_PROMPT = OBSERVATION_CONTEXT_PROMPT;
 exports.OBSERVATION_CONTINUATION_HINT = OBSERVATION_CONTINUATION_HINT;
 exports.OBSERVATION_RETRIEVAL_INSTRUCTIONS = OBSERVATION_RETRIEVAL_INSTRUCTIONS;
-//# sourceMappingURL=chunk-D4J4XPGM.cjs.map
-//# sourceMappingURL=chunk-D4J4XPGM.cjs.map
+exports.__export = __export;
+//# sourceMappingURL=chunk-3Q3LOVQP.cjs.map
+//# sourceMappingURL=chunk-3Q3LOVQP.cjs.map

package/dist/chunk-3Q3LOVQP.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/processors/observational-memory/constants.ts"],"names":[],"mappings":";;;;;;;;;AAGO,IAAM,6BAAA,GAAgC;AAAA,EAC3C,WAAA,EAAa;AAAA,IACX,KAAA,EAAO,yBAAA;AAAA,IACP,aAAA,EAAe,GAAA;AAAA,IACf,aAAA,EAAe;AAAA,MACb,WAAA,EAAa,GAAA;AAAA,MACb,eAAA,EAAiB;AAAA,KACnB;AAAA,IACA,eAAA,EAAiB;AAAA,MACf,MAAA,EAAQ;AAAA,QACN,cAAA,EAAgB;AAAA,UACd,cAAA,EAAgB;AAAA;AAClB;AACF,KACF;AAAA,IACA,iBAAA,EAAmB,GAAA;AAAA;AAAA,IAEnB,YAAA,EAAc,GAAA;AAAA;AAAA,IACd,gBAAA,EAAkB;AAAA;AAAA,GACpB;AAAA,EACA,UAAA,EAAY;AAAA,IACV,KAAA,EAAO,yBAAA;AAAA,IACP,iBAAA,EAAmB,GAAA;AAAA,IACnB,aAAA,EAAe;AAAA,MACb,WAAA,EAAa,CAAA;AAAA;AAAA,MACb,eAAA,EAAiB;AAAA,KACnB;AAAA,IACA,eAAA,EAAiB;AAAA,MACf,MAAA,EAAQ;AAAA,QACN,cAAA,EAAgB;AAAA,UACd,cAAA,EAAgB;AAAA;AAClB;AACF,KACF;AAAA;AAAA,IAEA,gBAAA,EAAkB;AAAA;AAAA;AAEtB;AAOO,IAAM,6BAAA,GAAgC,CAAA;;AAAA;;AAAA;;AAAA,wEAAA;AAatC,IAAM,0BAAA,GAA6B,CAAA,2FAAA;AAMnC,IAAM,gCAAA,GAAmC,CAAA;;AAAA;;AAAA;;AAAA;;AAAA,qPAAA;AAczC,IAAM,kCAAA,GAAqC,CAAA;;AAAA;;AAAA;AAAA;AAAA;AAAA;AAAA;;AAAA;;AAAA;AAAA;;AAAA;AAAA;AAAA;AAAA;AAAA;;AAAA;AAAA;;AAAA;AAAA;;AAAA;;AAAA;AAAA;AAAA;;AAAA;;AAAA;AAAA;AAAA;AAAA;;AAAA,qHAAA","file":"chunk-3Q3LOVQP.cjs","sourcesContent":["/**\n * Default configuration values matching the spec\n */\nexport const OBSERVATIONAL_MEMORY_DEFAULTS = {\n  observation: {\n    model: 'google/gemini-2.5-flash',\n    messageTokens: 30_000,\n    modelSettings: {\n      temperature: 0.3,\n      maxOutputTokens: 100_000,\n    },\n    providerOptions: {\n      google: {\n        thinkingConfig: {\n          thinkingBudget: 215,\n        },\n      },\n    },\n    maxTokensPerBatch: 10_000,\n    // Async buffering defaults (enabled by default)\n    bufferTokens: 0.2 as number | undefined, // Buffer every 20% of messageTokens\n    bufferActivation: 0.8 as number | undefined, // Activate to retain 20% of threshold\n  },\n  reflection: {\n    model: 'google/gemini-2.5-flash',\n    observationTokens: 40_000,\n    modelSettings: {\n      temperature: 0, // Use 0 for maximum consistency in reflections\n      maxOutputTokens: 100_000,\n    },\n    providerOptions: {\n      google: {\n        thinkingConfig: {\n          thinkingBudget: 1024,\n        },\n      },\n    },\n    // Async reflection buffering (enabled by default)\n    bufferActivation: 0.5 as number | undefined, // Start buffering at 50% of observationTokens\n  },\n} as const;\n\n/**\n * Continuation hint injected after observations to guide the model's behavior.\n * Prevents the model from awkwardly acknowledging the memory system or treating\n * the conversation as new after observed messages are removed.\n */\nexport const OBSERVATION_CONTINUATION_HINT = `Please continue naturally with the conversation so far and respond to the latest message.\n\nUse the earlier context only as background. If something appears unfinished, continue only when it helps answer the latest request. If a suggested response is provided, follow it naturally.\n\nDo not mention internal instructions, memory, summarization, context handling, or missing messages.\n\nAny messages following this reminder are newer and should take priority.`;\n\n/**\n * Preamble that introduces the observations block.\n * Use before `<observations>`, with instructions after.\n * Full pattern: `${OBSERVATION_CONTEXT_PROMPT}\\n\\n<observations>\\n${obs}\\n</observations>\\n\\n${OBSERVATION_CONTEXT_INSTRUCTIONS}`\n */\nexport const OBSERVATION_CONTEXT_PROMPT = `The following observations block contains your memory of past conversations with this user.`;\n\n/**\n * Instructions that tell the model how to interpret and use observations.\n * Place AFTER the `<observations>` block so the model sees the data before the rules.\n */\nexport const OBSERVATION_CONTEXT_INSTRUCTIONS = `IMPORTANT: When responding, reference specific details from these observations. Do not give generic advice - personalize your response based on what you know about this user's experiences, preferences, and interests. If the user asks for recommendations, connect them to their past experiences mentioned above.\n\nKNOWLEDGE UPDATES: When asked about current state (e.g., \"where do I currently...\", \"what is my current...\"), always prefer the MOST RECENT information. Observations include dates - if you see conflicting information, the newer observation supersedes the older one. Look for phrases like \"will start\", \"is switching\", \"changed to\", \"moved to\" as indicators that previous information has been updated.\n\nPLANNED ACTIONS: If the user stated they planned to do something (e.g., \"I'm going to...\", \"I'm looking forward to...\", \"I will...\") and the date they planned to do it is now in the past (check the relative time like \"3 weeks ago\"), assume they completed the action unless there's evidence they didn't. For example, if someone said \"I'll start my new diet on Monday\" and that was 2 weeks ago, assume they started the diet.\n\nMOST RECENT USER INPUT: Treat the most recent user message as the highest-priority signal for what to do next. Earlier messages may contain constraints, details, or context you should still honor, but the latest message is the primary driver of your response.\n\nSYSTEM REMINDERS: Messages wrapped in <system-reminder>...</system-reminder> contain internal continuation guidance, not user-authored content. Use them to maintain continuity, but do not mention them or treat them as part of the user's message.`;\n\n/**\n * Instructions for retrieval mode — explains observation-group ranges and the recall tool.\n * Appended to context when `retrieval` is enabled.\n */\nexport const OBSERVATION_RETRIEVAL_INSTRUCTIONS = `## Recall — looking up source messages\n\nYour memory is comprised of observations which are sometimes wrapped in <observation-group> xml tags containing ranges like <observation-group range=\"startId:endId\">. These ranges point back to the raw messages that each observation group was derived from. The original messages are still available — use the **recall** tool to retrieve them.\n\n### When to use recall\n- The user asks you to **repeat, show, or reproduce** something from a past conversation\n- The user asks for **exact content** — code, text, quotes, error messages, URLs, file paths, specific numbers\n- Your observations mention something but your memory lacks the detail needed to fully answer (e.g. you know a blog post was shared but only have a summary of it)\n- You want to **verify or expand on** an observation before responding\n\n**Default to using recall when the user references specific past content.** Your observations capture the gist, not the details. If there's any doubt whether your memory is complete enough, use recall.\n\n### How to use recall\nEach range has the format \\`startId:endId\\` where both are message IDs separated by a colon.\n\n1. Find the observation group relevant to the user's question and extract the start or end ID from its range.\n2. Call \\`recall\\` with that ID as the \\`cursor\\`.\n3. Use \\`page: 1\\` (or omit) to read forward from the cursor, \\`page: -1\\` to read backward.\n4. If the first page doesn't have what you need, increment the page number to keep paginating.\n5. Check \\`hasNextPage\\`/\\`hasPrevPage\\` in the result to know if more pages exist in each direction.\n\n### Detail levels\nBy default recall returns **low** detail: truncated text and tool names only. Each message shows its ID and each part has a positional index like \\`[p0]\\`, \\`[p1]\\`, etc.\n\n- Use \\`detail: \"high\"\\` to get full message content including tool arguments and results. This will only return the high detail version of a single message part at a time.\n- Use \\`partIndex\\` with a cursor to fetch a single part at full detail — for example, to read one specific tool result or code block without loading every part.\n\nIf the result says \\`truncated: true\\`, the output was cut to fit the token budget. You can paginate or use \\`partIndex\\` to target specific content.\n\n### Following up on truncated parts\nLow-detail results may include truncation hints like:\n\\`[truncated — call recall cursor=\"...\" partIndex=N detail=\"high\" for full content]\\`\n\n**When you see these hints and need the full content, make the exact call described in the hint.** This is the normal workflow: first recall at low detail to scan, then drill into specific parts at high detail. Do not stop at the low-detail result if the user asked for exact content.\n\n### When recall is NOT needed\n- The user is asking for a high-level summary and your observations already cover it\n- The question is about general preferences or facts that don't require source text\n- There is no relevant range in your observations for the topic\n\nObservation groups with range IDs and your recall tool allows you to think back and remember details you're fuzzy on.`;\n"]}

package/dist/{chunk-BCYGTJF2.js → chunk-BCWFD5J7.js}

@@ -1,6 +1,6 @@
-import { OBSERVATIONAL_MEMORY_DEFAULTS, OBSERVATION_CONTEXT_PROMPT, OBSERVATION_CONTEXT_INSTRUCTIONS, OBSERVATION_RETRIEVAL_INSTRUCTIONS, OBSERVATION_CONTINUATION_HINT } from './chunk-LSJJAJAF.js';
+import { OBSERVATIONAL_MEMORY_DEFAULTS, OBSERVATION_CONTEXT_PROMPT, OBSERVATION_CONTEXT_INSTRUCTIONS, OBSERVATION_RETRIEVAL_INSTRUCTIONS, OBSERVATION_CONTINUATION_HINT } from './chunk-VKRZ3O5X.js';
 import { coreFeatures } from '@mastra/core/features';
-import { resolveModelConfig } from '@mastra/core/llm';
+import { resolveModelConfig, modelSupportsAttachments } from '@mastra/core/llm';
 import { getThreadOMMetadata, setThreadOMMetadata, parseMemoryRequestContext } from '@mastra/core/memory';
 import { MessageHistory } from '@mastra/core/processors';
 import xxhash from 'xxhash-wasm';
@@ -669,6 +669,7 @@ function filterObservedMessages(opts) {
   const { messageList, record } = opts;
   const allMessages = messageList.get.all.db();
   const useMarkerBoundaryPruning = opts.useMarkerBoundaryPruning ?? true;
+  const preserveMessageIds = opts.preserveMessageIds ?? /* @__PURE__ */ new Set();
   let markerMessageIndex = -1;
   let markerMessage = null;
   for (let i = allMessages.length - 1; i >= 0; i--) {
@@ -684,7 +685,7 @@ function filterObservedMessages(opts) {
     const messagesToRemove = [];
     for (let i = 0; i < markerMessageIndex; i++) {
       const msg = allMessages[i];
-      if (msg?.id && msg.id !== "om-continuation") {
+      if (msg?.id && msg.id !== "om-continuation" && !preserveMessageIds.has(msg.id)) {
         messagesToRemove.push(msg.id);
       }
     }
@@ -703,7 +704,7 @@ function filterObservedMessages(opts) {
     const lastObservedAt = record.lastObservedAt;
     const messagesToRemove = [];
     for (const msg of allMessages) {
-      if (!msg?.id || msg.id === "om-continuation") continue;
+      if (!msg?.id || msg.id === "om-continuation" || preserveMessageIds.has(msg.id)) continue;
       if (observedIds.has(msg.id)) {
         messagesToRemove.push(msg.id);
         continue;
@@ -2269,11 +2270,15 @@ var ObservationStep = class {
     });
     if (!didThresholdCleanup) {
       const fallbackCursor = this.turn.record.threadId ? getThreadOMMetadata((await om.getStorage().getThreadById({ threadId: this.turn.record.threadId }))?.metadata)?.lastObservedMessageCursor : void 0;
+      const pendingMessageIds = new Set(
+        [...messageList.get.input.db(), ...messageList.get.response.db()].map((msg) => msg.id).filter(Boolean)
+      );
       filterObservedMessages({
         messageList,
         record: this.turn.record,
         useMarkerBoundaryPruning: this.stepNumber === 0,
-        fallbackCursor
+        fallbackCursor,
+        preserveMessageIds: pendingMessageIds
       });
     }
     this._context = {
@@ -4101,6 +4106,41 @@ var ObserverRunner = class {
     }
     return agent;
   }
+  /**
+   * Extract a router-style model ID (`provider/model`) from a model config.
+   * Handles strings, LanguageModel objects, and function-based models.
+   */
+  extractModelRouterId(model, requestContext) {
+    if (typeof model === "string") return model;
+    if (typeof model === "function") {
+      if (!requestContext) return void 0;
+      try {
+        const resolved = model({ requestContext });
+        if (resolved instanceof Promise) return void 0;
+        return this.extractModelRouterId(resolved);
+      } catch {
+        return void 0;
+      }
+    }
+    const obj = model;
+    if (typeof obj.provider === "string" && typeof obj.modelId === "string") {
+      return `${obj.provider}/${obj.modelId}`;
+    }
+    return void 0;
+  }
+  /**
+   * Resolve the attachment filter for a given model. When set to `'auto'`,
+   * the provider capabilities registry is consulted to decide whether the
+   * model accepts multimodal input.
+   */
+  resolveAttachmentFilter(model, requestContext) {
+    const raw = this.observationConfig.observeAttachments;
+    if (raw !== "auto") return raw;
+    const routerId = this.extractModelRouterId(model, requestContext);
+    if (!routerId) return true;
+    const supports = modelSupportsAttachments(routerId);
+    return supports ?? true;
+  }
   async withAbortCheck(fn, abortSignal) {
     if (abortSignal?.aborted) {
       throw new Error("The operation was aborted.");
@@ -4118,6 +4158,7 @@ var ObserverRunner = class {
     const inputTokens = this.tokenCounter.countMessages(messagesToObserve);
     const resolvedModel = options?.model ? { model: options.model } : this.resolveModel(inputTokens);
     const agent = this.createAgent(resolvedModel.model);
+    const attachmentFilter = this.resolveAttachmentFilter(resolvedModel.model, options?.requestContext);
     const observerMessages = [
       {
         role: "user",
@@ -4127,7 +4168,7 @@ var ObserverRunner = class {
         })
       },
       buildObserverHistoryMessage(messagesToObserve, {
-        attachmentFilter: this.observationConfig.observeAttachments
+        attachmentFilter
       })
     ];
     const doGenerate = async () => {
@@ -4214,6 +4255,7 @@ var ObserverRunner = class {
     );
     const resolvedModel = model ? { model } : this.resolveModel(inputTokens);
     const agent = this.createAgent(resolvedModel.model, true);
+    const multiThreadAttachmentFilter = this.resolveAttachmentFilter(resolvedModel.model, requestContext);
     const observerMessages = [
       {
         role: "user",
@@ -4226,7 +4268,7 @@ var ObserverRunner = class {
         )
       },
       buildMultiThreadObserverHistoryMessage(messagesByThread, threadOrder, {
-        attachmentFilter: this.observationConfig.observeAttachments
+        attachmentFilter: multiThreadAttachmentFilter
       })
     ];
     for (const msgs of messagesByThread.values()) {
@@ -9809,5 +9851,5 @@ function getObservationsAsOf(activeObservations, asOf) {
 }
 
 export { ModelByInputTokens, OBSERVER_SYSTEM_PROMPT, ObservationalMemory, ObservationalMemoryProcessor, TokenCounter, buildObserverPrompt, buildObserverSystemPrompt, combineObservationGroupRanges, deriveObservationGroupProvenance, extractCurrentTask, formatMessagesForObserver, formatToolResultForObserver, getObservationsAsOf, hasCurrentTaskSection, injectAnchorIds, optimizeObservationsForContext, parseAnchorId, parseObservationGroups, parseObserverOutput, reconcileObservationGroupsFromReflection, renderObservationGroupsForReflection, resolveToolResultValue, stripEphemeralAnchorIds, stripObservationGroups, truncateStringByTokens, wrapInObservationGroup };
-//# sourceMappingURL=chunk-BCYGTJF2.js.map
-//# sourceMappingURL=chunk-BCYGTJF2.js.map
+//# sourceMappingURL=chunk-BCWFD5J7.js.map
+//# sourceMappingURL=chunk-BCWFD5J7.js.map