@anuma/sdk 7.0.4 → 7.0.5

package/dist/expo/index.cjs

@@ -5496,6 +5496,24 @@ var getApiV1Models = (options) => (options?.client ?? client).get({ url: "/api/v
 // src/lib/memoryEngine/generate.ts
 init_clientConfig();
 init_taskType();
+
+// src/lib/memoryEngine/embeddingCacheState.ts
+var EMBEDDING_CACHE_STATE_KEY = /* @__PURE__ */ Symbol.for("@anuma/sdk/embedding-cache/v1");
+var cacheGlobal = globalThis;
+var fallbackState = {
+  scopedViews: /* @__PURE__ */ new WeakMap(),
+  inFlightByCache: /* @__PURE__ */ new WeakMap(),
+  cachedEmbeddingModels: /* @__PURE__ */ new WeakMap()
+};
+function getEmbeddingCacheState() {
+  try {
+    return cacheGlobal[EMBEDDING_CACHE_STATE_KEY] ?? (cacheGlobal[EMBEDDING_CACHE_STATE_KEY] = fallbackState);
+  } catch {
+    return fallbackState;
+  }
+}
+
+// src/lib/memoryEngine/generate.ts
 var EMBED_MAX_ATTEMPTS = 4;
 var DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS = 15e3;
 var DEFAULT_EMBEDDING_TOKEN_TIMEOUT_MS = 1e4;
@@ -5586,15 +5604,18 @@ async function generateEmbedding(text14, options) {
   if (total === void 0) return embedOne(text14, options, void 0);
   return withDeadline((signal) => embedOne(text14, options, signal), total, "embedding");
 }
-var inFlightByCache = /* @__PURE__ */ new WeakMap();
+var { inFlightByCache, cachedEmbeddingModels } = getEmbeddingCacheState();
 async function embedOne(text14, options, outer) {
   const { cache } = options;
   if (!cache) return Array.from(await requestEmbedding(text14, options, outer));
+  const model = options.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const cached = cache.get(text14);
-  if (cached) return Array.from(cached);
+  const cachedModel = cached && cachedEmbeddingModels.get(cached);
+  if (cached && (cachedModel === void 0 || cachedModel === model)) return Array.from(cached);
   const inFlight = inFlightByCache.get(cache) ?? /* @__PURE__ */ new Map();
   inFlightByCache.set(cache, inFlight);
-  const pending = inFlight.get(text14);
+  const requestKey = JSON.stringify([model, text14]);
+  const pending = inFlight.get(requestKey);
   if (pending) {
     try {
       return Array.from(await pending);
@@ -5603,13 +5624,14 @@ async function embedOne(text14, options, outer) {
     }
   }
   const request = requestEmbedding(text14, options, outer);
-  inFlight.set(text14, request);
+  inFlight.set(requestKey, request);
   try {
     const embedding2 = await request;
+    cachedEmbeddingModels.set(embedding2, model);
     cache.set(text14, embedding2);
     return Array.from(embedding2);
   } finally {
-    if (inFlight.get(text14) === request) inFlight.delete(text14);
+    if (inFlight.get(requestKey) === request) inFlight.delete(requestKey);
   }
 }
 async function requestEmbedding(text14, options, outer) {
@@ -5692,13 +5714,15 @@ async function embedMany(texts, options, outer) {
   const { baseUrl = BASE_URL, model, batchSize, cache } = options;
   const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
   const chunkSize = batchSize ?? DEFAULT_EMBEDDING_BATCH_SIZE;
+  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   const results = new Array(texts.length).fill(null);
   const uncachedIndices = [];
   const uncachedTexts = [];
   for (let i = 0; i < texts.length; i++) {
     if (cache) {
       const cached = cache.get(texts[i]);
-      if (cached) {
+      const cachedModel = cached && cachedEmbeddingModels.get(cached);
+      if (cached && (cachedModel === void 0 || cachedModel === embeddingModel)) {
         results[i] = Array.from(cached);
         continue;
       }
@@ -5714,7 +5738,6 @@ async function embedMany(texts, options, outer) {
     ...taskTypeHeader(options.taskType)
   };
   if (outer?.aborted) throw outer.reason;
-  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   let newEmbeddings;
   if (uncachedTexts.length <= chunkSize) {
     newEmbeddings = await generateEmbeddingsBatch(
@@ -5762,6 +5785,7 @@ async function embedMany(texts, options, outer) {
     const f32Embedding = Float32Array.from(newEmbeddings[i]);
     results[uncachedIndices[i]] = Array.from(f32Embedding);
     if (cache) {
+      cachedEmbeddingModels.set(f32Embedding, embeddingModel);
       cache.set(uncachedTexts[i], f32Embedding);
     }
   }
@@ -11203,7 +11227,7 @@ var MaskScopedEmbeddingCache = class extends Map {
     return this;
   }
 };
-var scopedViews = /* @__PURE__ */ new WeakMap();
+var { scopedViews } = getEmbeddingCacheState();
 function maskScopedEmbeddingCache(cache, masked) {
   const views = scopedViews.get(cache) ?? {};
   scopedViews.set(cache, views);
@@ -19401,44 +19425,48 @@ async function recall(query, ctx, options = {}) {
       embeddingOptions,
       ctx.vaultCache,
       searchOptions
-    )))(query, {
-      limit: types.includes("chunk") ? Math.max(limit * 2, 16) : limit,
-      minSimilarity: vaultMinScore,
-      useFusion: true,
-      rerank: flags.rerank,
-      ...options.rerankTopN !== void 0 && { rerankTopN: options.rerankTopN },
-      ...options.ceWeight !== void 0 && { ceWeight: options.ceWeight },
-      ...options.rerankLoadTimeoutMs !== void 0 && {
-        rerankLoadTimeoutMs: options.rerankLoadTimeoutMs
-      },
-      ...options.recencyAlpha !== void 0 && { recencyAlpha: options.recencyAlpha },
-      ...options.recency && { recency: options.recency },
-      ...options.mmr !== void 0 && { mmr: options.mmr },
-      ...options.supersessionBoost !== void 0 && {
-        supersessionBoost: options.supersessionBoost
-      },
-      ...options.supersessionWindow !== void 0 && {
-        supersessionWindow: options.supersessionWindow
-      },
-      ...options.proofCountAlpha !== void 0 && {
-        proofCountAlpha: options.proofCountAlpha
-      },
-      ...options.bm25AdmissionDivisor !== void 0 && {
-        bm25AdmissionDivisor: options.bm25AdmissionDivisor
+    )))(
+      query,
+      {
+        limit: types.includes("chunk") ? Math.max(limit * 2, 16) : limit,
+        minSimilarity: vaultMinScore,
+        useFusion: true,
+        rerank: flags.rerank,
+        ...options.rerankTopN !== void 0 && { rerankTopN: options.rerankTopN },
+        ...options.ceWeight !== void 0 && { ceWeight: options.ceWeight },
+        ...options.rerankLoadTimeoutMs !== void 0 && {
+          rerankLoadTimeoutMs: options.rerankLoadTimeoutMs
+        },
+        ...options.recencyAlpha !== void 0 && { recencyAlpha: options.recencyAlpha },
+        ...options.recency && { recency: options.recency },
+        ...options.mmr !== void 0 && { mmr: options.mmr },
+        ...options.supersessionBoost !== void 0 && {
+          supersessionBoost: options.supersessionBoost
+        },
+        ...options.supersessionWindow !== void 0 && {
+          supersessionWindow: options.supersessionWindow
+        },
+        ...options.proofCountAlpha !== void 0 && {
+          proofCountAlpha: options.proofCountAlpha
+        },
+        ...options.bm25AdmissionDivisor !== void 0 && {
+          bm25AdmissionDivisor: options.bm25AdmissionDivisor
+        },
+        ...options.rrfK !== void 0 && { rrfK: options.rrfK },
+        ...options.decryptLast !== void 0 && { decryptLast: options.decryptLast },
+        ...subQueries && { subQueries },
+        ...options.scopes && { scopes: options.scopes },
+        ...options.folderId !== void 0 && { folderId: options.folderId },
+        ...options.factTypes?.length && { factTypes: options.factTypes },
+        ...options.memoryIds !== void 0 && { memoryIds: options.memoryIds },
+        ...options.factTypeWeights && { factTypeWeights: options.factTypeWeights },
+        ...entityRanking.length > 0 && { entityRanking },
+        ...temporalRanking.length > 0 && { temporalRanking },
+        ...needsChunkEmbedding && { queryEmbedding: queryEmbedding ?? [] },
+        queryEmbedTotalTimeoutMs
       },
-      ...options.rrfK !== void 0 && { rrfK: options.rrfK },
-      ...options.decryptLast !== void 0 && { decryptLast: options.decryptLast },
-      ...subQueries && { subQueries },
-      ...options.scopes && { scopes: options.scopes },
-      ...options.folderId !== void 0 && { folderId: options.folderId },
-      ...options.factTypes?.length && { factTypes: options.factTypes },
-      ...options.memoryIds !== void 0 && { memoryIds: options.memoryIds },
-      ...options.factTypeWeights && { factTypeWeights: options.factTypeWeights },
-      ...entityRanking.length > 0 && { entityRanking },
-      ...temporalRanking.length > 0 && { temporalRanking },
-      ...needsChunkEmbedding && { queryEmbedding: queryEmbedding ?? [] },
-      queryEmbedTotalTimeoutMs
-    });
+      options.embeddingCache
+    );
     factResults.push(
       ...dedupeBy(
         results,
@@ -20585,10 +20613,19 @@ function createLocalMemoryStore(options) {
       retainOptions
     ),
     factSource: {
-      search: (query, searchOptions) => searchVaultMemoriesWithSize(
+      search: (query, searchOptions, embeddingCache) => searchVaultMemoriesWithSize(
         query,
         vaultCtx,
-        { ...options.embeddingOptions, taskType: "memory_recall_embed" },
+        {
+          ...options.embeddingOptions,
+          taskType: "memory_recall_embed",
+          ...embeddingCache && !options.embeddingOptions.cache && {
+            cache: maskScopedEmbeddingCache(
+              embeddingCache,
+              Boolean(options.embeddingOptions.maskInput)
+            )
+          }
+        },
         vaultCache,
         searchOptions
       ),
@@ -21220,7 +21257,7 @@ function stored(row) {
 function createRemoteMemoryPipeline(options) {
   const { persistence, embeddingOptions } = options;
   const model = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
-  const operation = (mode) => {
+  const operation = (mode, queryOptions = embeddingOptions) => {
     const rejectUndecryptable = (failed) => {
       if (!failed.length) return;
       if (mode === "retain")
@@ -21251,7 +21288,7 @@ function createRemoteMemoryPipeline(options) {
       const start = Date.now();
       const budgetMs = search.queryEmbedTotalTimeoutMs ?? 8e3;
       const embed = (texts) => generateEmbeddings(texts, {
-        ...embeddingOptions,
+        ...queryOptions,
         taskType: mode === "retain" ? "memory_retain_embed" : "memory_recall_embed",
         totalTimeoutMs: budgetMs
       });
@@ -21271,7 +21308,11 @@ function createRemoteMemoryPipeline(options) {
         }
       };
       const [primary, facetEmbeddings] = await Promise.all([
-        search.queryEmbedding ?? embed([query]).then((vectors2) => vectors2[0]),
+        search.queryEmbedding ?? (mode === "recall" && queryOptions.cache ? generateEmbedding(query, {
+          ...queryOptions,
+          taskType: "memory_recall_embed",
+          totalTimeoutMs: budgetMs
+        }) : embed([query]).then((vectors2) => vectors2[0])),
         embedFacets()
       ]);
       const queryEmbedding = storedVector(primary);
@@ -21495,15 +21536,17 @@ function createRemoteMemoryPipeline(options) {
   const factSource = {
     graphRanking: options.graphRanking,
     temporalRanking: options.temporalRanking,
-    search: async (text14, searchOptions) => {
-      const { prepare } = operation("recall");
+    search: async (text14, searchOptions, embeddingCache) => {
+      const queryOptions = {
+        ...embeddingOptions,
+        taskType: "memory_recall_embed",
+        ...embeddingCache && !embeddingOptions.cache && {
+          cache: maskScopedEmbeddingCache(embeddingCache, Boolean(embeddingOptions.maskInput))
+        }
+      };
+      const { prepare } = operation("recall", queryOptions);
       const prepared = await prepare(text14, searchOptions);
-      const ranked = await rankPreparedVaultCandidates(
-        text14,
-        prepared,
-        { ...embeddingOptions, taskType: "memory_recall_embed" },
-        searchOptions
-      );
+      const ranked = await rankPreparedVaultCandidates(text14, prepared, queryOptions, searchOptions);
       return {
         ...ranked,
         rankedOnCosine: prepared.memories.length > 0 && !prepared.embeddingsUnavailable,

package/dist/expo/index.d.mts

@@ -4508,7 +4508,7 @@ interface RecallOptions {
 }
 /** A fact source supplies bounded candidates; recall owns ranking and result handling. */
 interface RecallFactSource {
-    search: (query: string, options?: MemoryVaultSearchOptions) => ReturnType<typeof searchVaultMemoriesWithSize>;
+    search: (query: string, options?: MemoryVaultSearchOptions, embeddingCache?: Map<string, Float32Array>) => ReturnType<typeof searchVaultMemoriesWithSize>;
     graphRanking: (query: string, traverse: boolean, options: RecallOptions) => Promise<string[]>;
     temporalRanking: (query: string, now?: number) => Promise<string[]>;
 }
@@ -7990,13 +7990,6 @@ declare const chatStorageMigrations: Readonly<{
     sortedMigrations: _nozbe_watermelondb_Schema_migrations.Migration[];
 }>;
 
-/**
- * Namespacing for a caller-shared embedding cache.
- *
- * Lives here, next to the `embeddingCache` send-arg it serves, rather than in either platform hook:
- * react and expo both thread the option through their own send paths, and a copy per platform is how
- * the two drift.
- */
 declare function maskScopedEmbeddingCache(cache: Map<string, Float32Array>, masked: boolean): Map<string, Float32Array>;
 
 /**

package/dist/expo/index.d.ts

@@ -4508,7 +4508,7 @@ interface RecallOptions {
 }
 /** A fact source supplies bounded candidates; recall owns ranking and result handling. */
 interface RecallFactSource {
-    search: (query: string, options?: MemoryVaultSearchOptions) => ReturnType<typeof searchVaultMemoriesWithSize>;
+    search: (query: string, options?: MemoryVaultSearchOptions, embeddingCache?: Map<string, Float32Array>) => ReturnType<typeof searchVaultMemoriesWithSize>;
     graphRanking: (query: string, traverse: boolean, options: RecallOptions) => Promise<string[]>;
     temporalRanking: (query: string, now?: number) => Promise<string[]>;
 }
@@ -7990,13 +7990,6 @@ declare const chatStorageMigrations: Readonly<{
     sortedMigrations: _nozbe_watermelondb_Schema_migrations.Migration[];
 }>;
 
-/**
- * Namespacing for a caller-shared embedding cache.
- *
- * Lives here, next to the `embeddingCache` send-arg it serves, rather than in either platform hook:
- * react and expo both thread the option through their own send paths, and a copy per platform is how
- * the two drift.
- */
 declare function maskScopedEmbeddingCache(cache: Map<string, Float32Array>, masked: boolean): Map<string, Float32Array>;
 
 /**

package/dist/expo/index.mjs

@@ -3916,6 +3916,22 @@ var postApiV1Embeddings = (options) => (options.client ?? client).post({
 });
 var getApiV1Models = (options) => (options?.client ?? client).get({ url: "/api/v1/models", ...options });
 
+// src/lib/memoryEngine/embeddingCacheState.ts
+var EMBEDDING_CACHE_STATE_KEY = /* @__PURE__ */ Symbol.for("@anuma/sdk/embedding-cache/v1");
+var cacheGlobal = globalThis;
+var fallbackState = {
+  scopedViews: /* @__PURE__ */ new WeakMap(),
+  inFlightByCache: /* @__PURE__ */ new WeakMap(),
+  cachedEmbeddingModels: /* @__PURE__ */ new WeakMap()
+};
+function getEmbeddingCacheState() {
+  try {
+    return cacheGlobal[EMBEDDING_CACHE_STATE_KEY] ?? (cacheGlobal[EMBEDDING_CACHE_STATE_KEY] = fallbackState);
+  } catch {
+    return fallbackState;
+  }
+}
+
 // src/lib/memoryEngine/generate.ts
 var EMBED_MAX_ATTEMPTS = 4;
 var DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS = 15e3;
@@ -4007,15 +4023,18 @@ async function generateEmbedding(text14, options) {
   if (total === void 0) return embedOne(text14, options, void 0);
   return withDeadline((signal) => embedOne(text14, options, signal), total, "embedding");
 }
-var inFlightByCache = /* @__PURE__ */ new WeakMap();
+var { inFlightByCache, cachedEmbeddingModels } = getEmbeddingCacheState();
 async function embedOne(text14, options, outer) {
   const { cache } = options;
   if (!cache) return Array.from(await requestEmbedding(text14, options, outer));
+  const model = options.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const cached = cache.get(text14);
-  if (cached) return Array.from(cached);
+  const cachedModel = cached && cachedEmbeddingModels.get(cached);
+  if (cached && (cachedModel === void 0 || cachedModel === model)) return Array.from(cached);
   const inFlight = inFlightByCache.get(cache) ?? /* @__PURE__ */ new Map();
   inFlightByCache.set(cache, inFlight);
-  const pending = inFlight.get(text14);
+  const requestKey = JSON.stringify([model, text14]);
+  const pending = inFlight.get(requestKey);
   if (pending) {
     try {
       return Array.from(await pending);
@@ -4024,13 +4043,14 @@ async function embedOne(text14, options, outer) {
     }
   }
   const request = requestEmbedding(text14, options, outer);
-  inFlight.set(text14, request);
+  inFlight.set(requestKey, request);
   try {
     const embedding2 = await request;
+    cachedEmbeddingModels.set(embedding2, model);
     cache.set(text14, embedding2);
     return Array.from(embedding2);
   } finally {
-    if (inFlight.get(text14) === request) inFlight.delete(text14);
+    if (inFlight.get(requestKey) === request) inFlight.delete(requestKey);
   }
 }
 async function requestEmbedding(text14, options, outer) {
@@ -4113,13 +4133,15 @@ async function embedMany(texts, options, outer) {
   const { baseUrl = BASE_URL, model, batchSize, cache } = options;
   const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
   const chunkSize = batchSize ?? DEFAULT_EMBEDDING_BATCH_SIZE;
+  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   const results = new Array(texts.length).fill(null);
   const uncachedIndices = [];
   const uncachedTexts = [];
   for (let i = 0; i < texts.length; i++) {
     if (cache) {
       const cached = cache.get(texts[i]);
-      if (cached) {
+      const cachedModel = cached && cachedEmbeddingModels.get(cached);
+      if (cached && (cachedModel === void 0 || cachedModel === embeddingModel)) {
         results[i] = Array.from(cached);
         continue;
       }
@@ -4135,7 +4157,6 @@ async function embedMany(texts, options, outer) {
     ...taskTypeHeader(options.taskType)
   };
   if (outer?.aborted) throw outer.reason;
-  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   let newEmbeddings;
   if (uncachedTexts.length <= chunkSize) {
     newEmbeddings = await generateEmbeddingsBatch(
@@ -4183,6 +4204,7 @@ async function embedMany(texts, options, outer) {
     const f32Embedding = Float32Array.from(newEmbeddings[i]);
     results[uncachedIndices[i]] = Array.from(f32Embedding);
     if (cache) {
+      cachedEmbeddingModels.set(f32Embedding, embeddingModel);
       cache.set(uncachedTexts[i], f32Embedding);
     }
   }
@@ -9610,7 +9632,7 @@ var MaskScopedEmbeddingCache = class extends Map {
     return this;
   }
 };
-var scopedViews = /* @__PURE__ */ new WeakMap();
+var { scopedViews } = getEmbeddingCacheState();
 function maskScopedEmbeddingCache(cache, masked) {
   const views = scopedViews.get(cache) ?? {};
   scopedViews.set(cache, views);
@@ -17774,44 +17796,48 @@ async function recall(query, ctx, options = {}) {
       embeddingOptions,
       ctx.vaultCache,
       searchOptions
-    )))(query, {
-      limit: types.includes("chunk") ? Math.max(limit * 2, 16) : limit,
-      minSimilarity: vaultMinScore,
-      useFusion: true,
-      rerank: flags.rerank,
-      ...options.rerankTopN !== void 0 && { rerankTopN: options.rerankTopN },
-      ...options.ceWeight !== void 0 && { ceWeight: options.ceWeight },
-      ...options.rerankLoadTimeoutMs !== void 0 && {
-        rerankLoadTimeoutMs: options.rerankLoadTimeoutMs
-      },
-      ...options.recencyAlpha !== void 0 && { recencyAlpha: options.recencyAlpha },
-      ...options.recency && { recency: options.recency },
-      ...options.mmr !== void 0 && { mmr: options.mmr },
-      ...options.supersessionBoost !== void 0 && {
-        supersessionBoost: options.supersessionBoost
-      },
-      ...options.supersessionWindow !== void 0 && {
-        supersessionWindow: options.supersessionWindow
-      },
-      ...options.proofCountAlpha !== void 0 && {
-        proofCountAlpha: options.proofCountAlpha
-      },
-      ...options.bm25AdmissionDivisor !== void 0 && {
-        bm25AdmissionDivisor: options.bm25AdmissionDivisor
+    )))(
+      query,
+      {
+        limit: types.includes("chunk") ? Math.max(limit * 2, 16) : limit,
+        minSimilarity: vaultMinScore,
+        useFusion: true,
+        rerank: flags.rerank,
+        ...options.rerankTopN !== void 0 && { rerankTopN: options.rerankTopN },
+        ...options.ceWeight !== void 0 && { ceWeight: options.ceWeight },
+        ...options.rerankLoadTimeoutMs !== void 0 && {
+          rerankLoadTimeoutMs: options.rerankLoadTimeoutMs
+        },
+        ...options.recencyAlpha !== void 0 && { recencyAlpha: options.recencyAlpha },
+        ...options.recency && { recency: options.recency },
+        ...options.mmr !== void 0 && { mmr: options.mmr },
+        ...options.supersessionBoost !== void 0 && {
+          supersessionBoost: options.supersessionBoost
+        },
+        ...options.supersessionWindow !== void 0 && {
+          supersessionWindow: options.supersessionWindow
+        },
+        ...options.proofCountAlpha !== void 0 && {
+          proofCountAlpha: options.proofCountAlpha
+        },
+        ...options.bm25AdmissionDivisor !== void 0 && {
+          bm25AdmissionDivisor: options.bm25AdmissionDivisor
+        },
+        ...options.rrfK !== void 0 && { rrfK: options.rrfK },
+        ...options.decryptLast !== void 0 && { decryptLast: options.decryptLast },
+        ...subQueries && { subQueries },
+        ...options.scopes && { scopes: options.scopes },
+        ...options.folderId !== void 0 && { folderId: options.folderId },
+        ...options.factTypes?.length && { factTypes: options.factTypes },
+        ...options.memoryIds !== void 0 && { memoryIds: options.memoryIds },
+        ...options.factTypeWeights && { factTypeWeights: options.factTypeWeights },
+        ...entityRanking.length > 0 && { entityRanking },
+        ...temporalRanking.length > 0 && { temporalRanking },
+        ...needsChunkEmbedding && { queryEmbedding: queryEmbedding ?? [] },
+        queryEmbedTotalTimeoutMs
       },
-      ...options.rrfK !== void 0 && { rrfK: options.rrfK },
-      ...options.decryptLast !== void 0 && { decryptLast: options.decryptLast },
-      ...subQueries && { subQueries },
-      ...options.scopes && { scopes: options.scopes },
-      ...options.folderId !== void 0 && { folderId: options.folderId },
-      ...options.factTypes?.length && { factTypes: options.factTypes },
-      ...options.memoryIds !== void 0 && { memoryIds: options.memoryIds },
-      ...options.factTypeWeights && { factTypeWeights: options.factTypeWeights },
-      ...entityRanking.length > 0 && { entityRanking },
-      ...temporalRanking.length > 0 && { temporalRanking },
-      ...needsChunkEmbedding && { queryEmbedding: queryEmbedding ?? [] },
-      queryEmbedTotalTimeoutMs
-    });
+      options.embeddingCache
+    );
     factResults.push(
       ...dedupeBy(
         results,
@@ -18950,10 +18976,19 @@ function createLocalMemoryStore(options) {
       retainOptions
     ),
     factSource: {
-      search: (query, searchOptions) => searchVaultMemoriesWithSize(
+      search: (query, searchOptions, embeddingCache) => searchVaultMemoriesWithSize(
         query,
         vaultCtx,
-        { ...options.embeddingOptions, taskType: "memory_recall_embed" },
+        {
+          ...options.embeddingOptions,
+          taskType: "memory_recall_embed",
+          ...embeddingCache && !options.embeddingOptions.cache && {
+            cache: maskScopedEmbeddingCache(
+              embeddingCache,
+              Boolean(options.embeddingOptions.maskInput)
+            )
+          }
+        },
         vaultCache,
         searchOptions
       ),
@@ -19583,7 +19618,7 @@ function stored(row) {
 function createRemoteMemoryPipeline(options) {
   const { persistence, embeddingOptions } = options;
   const model = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
-  const operation = (mode) => {
+  const operation = (mode, queryOptions = embeddingOptions) => {
     const rejectUndecryptable = (failed) => {
       if (!failed.length) return;
       if (mode === "retain")
@@ -19614,7 +19649,7 @@ function createRemoteMemoryPipeline(options) {
       const start = Date.now();
       const budgetMs = search.queryEmbedTotalTimeoutMs ?? 8e3;
       const embed = (texts) => generateEmbeddings(texts, {
-        ...embeddingOptions,
+        ...queryOptions,
         taskType: mode === "retain" ? "memory_retain_embed" : "memory_recall_embed",
         totalTimeoutMs: budgetMs
       });
@@ -19634,7 +19669,11 @@ function createRemoteMemoryPipeline(options) {
         }
       };
       const [primary, facetEmbeddings] = await Promise.all([
-        search.queryEmbedding ?? embed([query]).then((vectors2) => vectors2[0]),
+        search.queryEmbedding ?? (mode === "recall" && queryOptions.cache ? generateEmbedding(query, {
+          ...queryOptions,
+          taskType: "memory_recall_embed",
+          totalTimeoutMs: budgetMs
+        }) : embed([query]).then((vectors2) => vectors2[0])),
         embedFacets()
       ]);
       const queryEmbedding = storedVector(primary);
@@ -19858,15 +19897,17 @@ function createRemoteMemoryPipeline(options) {
   const factSource = {
     graphRanking: options.graphRanking,
     temporalRanking: options.temporalRanking,
-    search: async (text14, searchOptions) => {
-      const { prepare } = operation("recall");
+    search: async (text14, searchOptions, embeddingCache) => {
+      const queryOptions = {
+        ...embeddingOptions,
+        taskType: "memory_recall_embed",
+        ...embeddingCache && !embeddingOptions.cache && {
+          cache: maskScopedEmbeddingCache(embeddingCache, Boolean(embeddingOptions.maskInput))
+        }
+      };
+      const { prepare } = operation("recall", queryOptions);
       const prepared = await prepare(text14, searchOptions);
-      const ranked = await rankPreparedVaultCandidates(
-        text14,
-        prepared,
-        { ...embeddingOptions, taskType: "memory_recall_embed" },
-        searchOptions
-      );
+      const ranked = await rankPreparedVaultCandidates(text14, prepared, queryOptions, searchOptions);
       return {
         ...ranked,
         rankedOnCosine: prepared.memories.length > 0 && !prepared.embeddingsUnavailable,