@anuma/sdk 1.4.0 → 2.0.0

package/dist/expo/{chunk-QKF7PRL5.mjs → chunk-N4FEY34G.mjs}

@@ -4373,29 +4373,29 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
   const uncachedKeys = [];
   const uncachedIds = [];
   for (const m of memories) {
-    const key = m.content;
-    if (isEncrypted(key)) continue;
-    if (!cache.has(key)) {
+    const content = m.content;
+    if (isEncrypted(content)) continue;
+    if (!cache.has(m.uniqueId)) {
       const modelCompatible = (m.embeddingModel ?? currentModel) === currentModel;
       if (m.embedding && modelCompatible) {
         try {
           const parsed = JSON.parse(m.embedding);
           if (Array.isArray(parsed)) {
-            cache.set(key, parsed);
+            cache.set(m.uniqueId, Float32Array.from(parsed));
             continue;
           }
         } catch {
         }
       }
       uncachedTexts.push(m.content);
-      uncachedKeys.push(key);
+      uncachedKeys.push(m.uniqueId);
       uncachedIds.push(m.uniqueId);
     }
   }
   if (uncachedTexts.length > 0) {
     const embeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let i = 0; i < uncachedKeys.length; i++) {
-      cache.set(uncachedKeys[i], embeddings[i]);
+      cache.set(uncachedKeys[i], Float32Array.from(embeddings[i]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         uncachedIds[i],
@@ -4409,7 +4409,7 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
 async function eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, memoryId) {
   if (isEncrypted(content)) return;
   const embedding = await generateEmbedding(content, embeddingOptions);
-  cache.set(content, embedding);
+  if (memoryId) cache.set(memoryId, Float32Array.from(embedding));
   if (vaultCtx && memoryId) {
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     updateVaultMemoryEmbeddingOp(vaultCtx, memoryId, JSON.stringify(embedding), currentModel).catch(
@@ -4450,16 +4450,17 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   let staleReembedCount = 0;
   for (let i = 0; i < memories.length; i++) {
     const content = memories[i].content;
-    const cached = cache.get(content);
+    const memoryId = memories[i].uniqueId;
+    const cached = cache.get(memoryId);
     if (cached && cached.length === queryEmbedding.length) continue;
-    if (cached) cache.delete(content);
+    if (cached) cache.delete(memoryId);
     const storedModel = memories[i].embeddingModel;
     const modelCompatible = (storedModel ?? currentModel) === currentModel;
     if (memories[i].embedding && modelCompatible) {
       try {
         const parsed = JSON.parse(memories[i].embedding);
         if (Array.isArray(parsed) && parsed.length === queryEmbedding.length) {
-          cache.set(content, parsed);
+          cache.set(memoryId, Float32Array.from(parsed));
           continue;
         }
       } catch {
@@ -4477,7 +4478,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   if (uncachedTexts.length > 0) {
     const newEmbeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let j = 0; j < uncachedTexts.length; j++) {
-      cache.set(memories[uncachedIndices[j]].content, newEmbeddings[j]);
+      cache.set(memories[uncachedIndices[j]].uniqueId, Float32Array.from(newEmbeddings[j]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         memories[uncachedIndices[j]].uniqueId,
@@ -4493,7 +4494,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   const embeddedItems = memories.map((m) => ({
     id: m.uniqueId,
     content: m.content,
-    embedding: cache.get(m.content) ?? [],
+    embedding: cache.get(m.uniqueId) ?? [],
     createdAt: m.createdAt,
     updatedAt: m.updatedAt,
     proofCount: m.proofCount,
@@ -4689,7 +4690,7 @@ function createMemoryVaultSearchTool(vaultCtx, embeddingOptions, cache, searchOp
             legacy.map((r) => ({ id: r.uniqueId, content: r.content, score: r.similarity }))
           );
         }
-        const { recall: recall2 } = await import("./recall-YKNVP2DF.mjs");
+        const { recall: recall2 } = await import("./recall-DHOR5WWW.mjs");
         const result = await recall2(
           query,
           { vaultCtx, embeddingOptions, vaultCache: cache },

package/dist/expo/index.cjs

@@ -6110,29 +6110,29 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
   const uncachedKeys = [];
   const uncachedIds = [];
   for (const m of memories) {
-    const key = m.content;
-    if (isEncrypted(key)) continue;
-    if (!cache.has(key)) {
+    const content = m.content;
+    if (isEncrypted(content)) continue;
+    if (!cache.has(m.uniqueId)) {
       const modelCompatible = (m.embeddingModel ?? currentModel) === currentModel;
       if (m.embedding && modelCompatible) {
         try {
           const parsed = JSON.parse(m.embedding);
           if (Array.isArray(parsed)) {
-            cache.set(key, parsed);
+            cache.set(m.uniqueId, Float32Array.from(parsed));
             continue;
           }
         } catch {
         }
       }
       uncachedTexts.push(m.content);
-      uncachedKeys.push(key);
+      uncachedKeys.push(m.uniqueId);
       uncachedIds.push(m.uniqueId);
     }
   }
   if (uncachedTexts.length > 0) {
     const embeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let i = 0; i < uncachedKeys.length; i++) {
-      cache.set(uncachedKeys[i], embeddings[i]);
+      cache.set(uncachedKeys[i], Float32Array.from(embeddings[i]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         uncachedIds[i],
@@ -6146,7 +6146,7 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
 async function eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, memoryId) {
   if (isEncrypted(content)) return;
   const embedding = await generateEmbedding(content, embeddingOptions);
-  cache.set(content, embedding);
+  if (memoryId) cache.set(memoryId, Float32Array.from(embedding));
   if (vaultCtx && memoryId) {
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     updateVaultMemoryEmbeddingOp(vaultCtx, memoryId, JSON.stringify(embedding), currentModel).catch(
@@ -6187,16 +6187,17 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   let staleReembedCount = 0;
   for (let i = 0; i < memories.length; i++) {
     const content = memories[i].content;
-    const cached = cache.get(content);
+    const memoryId = memories[i].uniqueId;
+    const cached = cache.get(memoryId);
     if (cached && cached.length === queryEmbedding.length) continue;
-    if (cached) cache.delete(content);
+    if (cached) cache.delete(memoryId);
     const storedModel = memories[i].embeddingModel;
     const modelCompatible = (storedModel ?? currentModel) === currentModel;
     if (memories[i].embedding && modelCompatible) {
       try {
         const parsed = JSON.parse(memories[i].embedding);
         if (Array.isArray(parsed) && parsed.length === queryEmbedding.length) {
-          cache.set(content, parsed);
+          cache.set(memoryId, Float32Array.from(parsed));
           continue;
         }
       } catch {
@@ -6214,7 +6215,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   if (uncachedTexts.length > 0) {
     const newEmbeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let j = 0; j < uncachedTexts.length; j++) {
-      cache.set(memories[uncachedIndices[j]].content, newEmbeddings[j]);
+      cache.set(memories[uncachedIndices[j]].uniqueId, Float32Array.from(newEmbeddings[j]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         memories[uncachedIndices[j]].uniqueId,
@@ -6230,7 +6231,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   const embeddedItems = memories.map((m) => ({
     id: m.uniqueId,
     content: m.content,
-    embedding: cache.get(m.content) ?? [],
+    embedding: cache.get(m.uniqueId) ?? [],
     createdAt: m.createdAt,
     updatedAt: m.updatedAt,
     proofCount: m.proofCount,
@@ -12635,7 +12636,6 @@ async function retain(content, ctx, options = {}) {
     }
   }
   const embedding = await generateEmbedding(trimmed, ctx.embeddingOptions);
-  ctx.vaultCache.set(trimmed, embedding);
   const embeddingModel = ctx.embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const created = await createVaultMemoryOp(ctx.vaultCtx, {
     content: trimmed,
@@ -12656,6 +12656,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -12732,7 +12733,6 @@ async function tryConsolidate(trimmed, ctx, options) {
       options.sourceChunkIds ?? []
     );
     const newEmbedding = await generateEmbedding(decision.content, ctx.embeddingOptions);
-    ctx.vaultCache.set(decision.content, newEmbedding);
     const consolidatedModel = ctx.embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     const eventTimeUpdate = pickEventTimeUpdate(existing, options.eventTime);
     const updated = await updateVaultMemoryOp(ctx.vaultCtx, decision.targetId, {
@@ -12752,6 +12752,7 @@ async function tryConsolidate(trimmed, ctx, options) {
       await assertMergeTargetGoneOrThrow(ctx, decision.targetId);
       return null;
     }
+    ctx.vaultCache.set(decision.targetId, Float32Array.from(newEmbedding));
     return {
       action: "update",
       memoryId: decision.targetId,
@@ -13821,9 +13822,7 @@ function createMemoryVaultTool(vaultCtx, options, embeddingOptions, cache) {
             return `Error: Failed to update memory "${id}".`;
           }
           if (embeddingOptions && cache) {
-            if (previousContent) {
-              cache.delete(previousContent);
-            }
+            cache.delete(id);
             eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, id).catch(
               // Silently swallow – SDK must not use console.*; embedding will be retried on next search
               () => {

package/dist/expo/index.d.mts

@@ -4726,7 +4726,9 @@ interface RecencyOptions {
 
 declare const DEFAULT_VAULT_CACHE_SIZE = 5000;
 /**
- * Create a VaultEmbeddingCache backed by an LRU with a default cap of 1000 entries.
+ * Create a VaultEmbeddingCache backed by an LRU with a default cap of
+ * DEFAULT_VAULT_CACHE_SIZE entries. Values are Float32Array — the model's
+ * native precision — not float64 number[], halving resident RAM losslessly.
  */
 declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCache;
 
@@ -4741,7 +4743,7 @@ declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCach
  * Embedding cache keyed by content string. Stores pre-computed embeddings
  * so that search only needs to embed the query, not the vault entries.
  */
-type VaultEmbeddingCache = Map<string, number[]>;
+type VaultEmbeddingCache = Map<string, Float32Array>;
 /**
  * Options for the vault search tool.
  */

package/dist/expo/index.d.ts

@@ -4726,7 +4726,9 @@ interface RecencyOptions {
 
 declare const DEFAULT_VAULT_CACHE_SIZE = 5000;
 /**
- * Create a VaultEmbeddingCache backed by an LRU with a default cap of 1000 entries.
+ * Create a VaultEmbeddingCache backed by an LRU with a default cap of
+ * DEFAULT_VAULT_CACHE_SIZE entries. Values are Float32Array — the model's
+ * native precision — not float64 number[], halving resident RAM losslessly.
  */
 declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCache;
 
@@ -4741,7 +4743,7 @@ declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCach
  * Embedding cache keyed by content string. Stores pre-computed embeddings
  * so that search only needs to embed the query, not the vault entries.
  */
-type VaultEmbeddingCache = Map<string, number[]>;
+type VaultEmbeddingCache = Map<string, Float32Array>;
 /**
  * Options for the vault search tool.
  */

package/dist/expo/index.mjs

@@ -79,7 +79,7 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   useEncryption
-} from "./chunk-QKF7PRL5.mjs";
+} from "./chunk-N4FEY34G.mjs";
 import {
   PII_PATTERNS,
   PiiRedactor,
@@ -5911,7 +5911,6 @@ async function retain(content, ctx, options = {}) {
     }
   }
   const embedding = await generateEmbedding(trimmed, ctx.embeddingOptions);
-  ctx.vaultCache.set(trimmed, embedding);
   const embeddingModel = ctx.embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const created = await createVaultMemoryOp(ctx.vaultCtx, {
     content: trimmed,
@@ -5932,6 +5931,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -6008,7 +6008,6 @@ async function tryConsolidate(trimmed, ctx, options) {
       options.sourceChunkIds ?? []
     );
     const newEmbedding = await generateEmbedding(decision.content, ctx.embeddingOptions);
-    ctx.vaultCache.set(decision.content, newEmbedding);
     const consolidatedModel = ctx.embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     const eventTimeUpdate = pickEventTimeUpdate(existing, options.eventTime);
     const updated = await updateVaultMemoryOp(ctx.vaultCtx, decision.targetId, {
@@ -6028,6 +6027,7 @@ async function tryConsolidate(trimmed, ctx, options) {
       await assertMergeTargetGoneOrThrow(ctx, decision.targetId);
       return null;
     }
+    ctx.vaultCache.set(decision.targetId, Float32Array.from(newEmbedding));
     return {
       action: "update",
       memoryId: decision.targetId,
@@ -7072,9 +7072,7 @@ function createMemoryVaultTool(vaultCtx, options, embeddingOptions, cache) {
             return `Error: Failed to update memory "${id}".`;
           }
           if (embeddingOptions && cache) {
-            if (previousContent) {
-              cache.delete(previousContent);
-            }
+            cache.delete(id);
             eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, id).catch(
               // Silently swallow – SDK must not use console.*; embedding will be retried on next search
               () => {

package/dist/expo/{recall-YKNVP2DF.mjs → recall-DHOR5WWW.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-QKF7PRL5.mjs";
+} from "./chunk-N4FEY34G.mjs";
 import "./chunk-DFVCS3QQ.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {

package/dist/index.d.mts

@@ -1253,6 +1253,7 @@ type HandlersConfigResponse = {
      * Apps is the list of active apps
      */
     apps?: Array<HandlersAppConfig>;
+    mobile_app?: HandlersMobileAppVersions;
     /**
      * PhoneCallsEnabled indicates whether Bland phone calling is available
      */
@@ -1640,6 +1641,28 @@ type HandlersListUsersResponse = {
     pagination: HandlersPaginationResponse;
     users: Array<HandlersDeveloperUserResponse>;
 };
+/**
+ * MobileApp is the mobile app version info; absent when not configured
+ */
+type HandlersMobileAppVersions = {
+    android?: HandlersMobilePlatformVersions;
+    ios?: HandlersMobilePlatformVersions;
+};
+/**
+ * Android versions for the Play Store build
+ */
+type HandlersMobilePlatformVersions = {
+    /**
+     * LatestVersion is the newest version live in the platform's store;
+     * older clients show a dismissible update nudge
+     */
+    latest_version?: string;
+    /**
+     * MinSupportedVersion is the oldest version still supported; older
+     * clients are blocked behind an update-required wall
+     */
+    min_supported_version?: string;
+};
 type HandlersModelToolUsageItem = {
     call_count: number;
     cost_usd: number;

package/dist/index.d.ts

@@ -1253,6 +1253,7 @@ type HandlersConfigResponse = {
      * Apps is the list of active apps
      */
     apps?: Array<HandlersAppConfig>;
+    mobile_app?: HandlersMobileAppVersions;
     /**
      * PhoneCallsEnabled indicates whether Bland phone calling is available
      */
@@ -1640,6 +1641,28 @@ type HandlersListUsersResponse = {
     pagination: HandlersPaginationResponse;
     users: Array<HandlersDeveloperUserResponse>;
 };
+/**
+ * MobileApp is the mobile app version info; absent when not configured
+ */
+type HandlersMobileAppVersions = {
+    android?: HandlersMobilePlatformVersions;
+    ios?: HandlersMobilePlatformVersions;
+};
+/**
+ * Android versions for the Play Store build
+ */
+type HandlersMobilePlatformVersions = {
+    /**
+     * LatestVersion is the newest version live in the platform's store;
+     * older clients show a dismissible update nudge
+     */
+    latest_version?: string;
+    /**
+     * MinSupportedVersion is the oldest version still supported; older
+     * clients are blocked behind an update-required wall
+     */
+    min_supported_version?: string;
+};
 type HandlersModelToolUsageItem = {
     call_count: number;
     cost_usd: number;

package/dist/react/{chunk-SEZCFG24.mjs → chunk-63IW5BS5.mjs}

@@ -4704,29 +4704,29 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
   const uncachedKeys = [];
   const uncachedIds = [];
   for (const m of memories) {
-    const key = m.content;
-    if (isEncrypted(key)) continue;
-    if (!cache.has(key)) {
+    const content = m.content;
+    if (isEncrypted(content)) continue;
+    if (!cache.has(m.uniqueId)) {
       const modelCompatible = (m.embeddingModel ?? currentModel) === currentModel;
       if (m.embedding && modelCompatible) {
         try {
           const parsed = JSON.parse(m.embedding);
           if (Array.isArray(parsed)) {
-            cache.set(key, parsed);
+            cache.set(m.uniqueId, Float32Array.from(parsed));
             continue;
           }
         } catch {
         }
       }
       uncachedTexts.push(m.content);
-      uncachedKeys.push(key);
+      uncachedKeys.push(m.uniqueId);
       uncachedIds.push(m.uniqueId);
     }
   }
   if (uncachedTexts.length > 0) {
     const embeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let i = 0; i < uncachedKeys.length; i++) {
-      cache.set(uncachedKeys[i], embeddings[i]);
+      cache.set(uncachedKeys[i], Float32Array.from(embeddings[i]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         uncachedIds[i],
@@ -4740,7 +4740,7 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
 async function eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, memoryId) {
   if (isEncrypted(content)) return;
   const embedding = await generateEmbedding(content, embeddingOptions);
-  cache.set(content, embedding);
+  if (memoryId) cache.set(memoryId, Float32Array.from(embedding));
   if (vaultCtx && memoryId) {
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     updateVaultMemoryEmbeddingOp(vaultCtx, memoryId, JSON.stringify(embedding), currentModel).catch(
@@ -4781,16 +4781,17 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   let staleReembedCount = 0;
   for (let i = 0; i < memories.length; i++) {
     const content = memories[i].content;
-    const cached = cache.get(content);
+    const memoryId = memories[i].uniqueId;
+    const cached = cache.get(memoryId);
     if (cached && cached.length === queryEmbedding.length) continue;
-    if (cached) cache.delete(content);
+    if (cached) cache.delete(memoryId);
     const storedModel = memories[i].embeddingModel;
     const modelCompatible = (storedModel ?? currentModel) === currentModel;
     if (memories[i].embedding && modelCompatible) {
       try {
         const parsed = JSON.parse(memories[i].embedding);
         if (Array.isArray(parsed) && parsed.length === queryEmbedding.length) {
-          cache.set(content, parsed);
+          cache.set(memoryId, Float32Array.from(parsed));
           continue;
         }
       } catch {
@@ -4808,7 +4809,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   if (uncachedTexts.length > 0) {
     const newEmbeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let j = 0; j < uncachedTexts.length; j++) {
-      cache.set(memories[uncachedIndices[j]].content, newEmbeddings[j]);
+      cache.set(memories[uncachedIndices[j]].uniqueId, Float32Array.from(newEmbeddings[j]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         memories[uncachedIndices[j]].uniqueId,
@@ -4824,7 +4825,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   const embeddedItems = memories.map((m) => ({
     id: m.uniqueId,
     content: m.content,
-    embedding: cache.get(m.content) ?? [],
+    embedding: cache.get(m.uniqueId) ?? [],
     createdAt: m.createdAt,
     updatedAt: m.updatedAt,
     proofCount: m.proofCount,
@@ -5020,7 +5021,7 @@ function createMemoryVaultSearchTool(vaultCtx, embeddingOptions, cache, searchOp
             legacy.map((r) => ({ id: r.uniqueId, content: r.content, score: r.similarity }))
           );
         }
-        const { recall: recall2 } = await import("./recall-UWH5HU6V.mjs");
+        const { recall: recall2 } = await import("./recall-PUXEQHAP.mjs");
         const result = await recall2(
           query,
           { vaultCtx, embeddingOptions, vaultCache: cache },

package/dist/react/index.cjs

@@ -6441,29 +6441,29 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
   const uncachedKeys = [];
   const uncachedIds = [];
   for (const m of memories) {
-    const key = m.content;
-    if (isEncrypted(key)) continue;
-    if (!cache.has(key)) {
+    const content = m.content;
+    if (isEncrypted(content)) continue;
+    if (!cache.has(m.uniqueId)) {
       const modelCompatible = (m.embeddingModel ?? currentModel) === currentModel;
       if (m.embedding && modelCompatible) {
         try {
           const parsed = JSON.parse(m.embedding);
           if (Array.isArray(parsed)) {
-            cache.set(key, parsed);
+            cache.set(m.uniqueId, Float32Array.from(parsed));
             continue;
           }
         } catch {
         }
       }
       uncachedTexts.push(m.content);
-      uncachedKeys.push(key);
+      uncachedKeys.push(m.uniqueId);
       uncachedIds.push(m.uniqueId);
     }
   }
   if (uncachedTexts.length > 0) {
     const embeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let i = 0; i < uncachedKeys.length; i++) {
-      cache.set(uncachedKeys[i], embeddings[i]);
+      cache.set(uncachedKeys[i], Float32Array.from(embeddings[i]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         uncachedIds[i],
@@ -6477,7 +6477,7 @@ async function preEmbedVaultMemories(vaultCtx, embeddingOptions, cache, options)
 async function eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, memoryId) {
   if (isEncrypted(content)) return;
   const embedding = await generateEmbedding(content, embeddingOptions);
-  cache.set(content, embedding);
+  if (memoryId) cache.set(memoryId, Float32Array.from(embedding));
   if (vaultCtx && memoryId) {
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     updateVaultMemoryEmbeddingOp(vaultCtx, memoryId, JSON.stringify(embedding), currentModel).catch(
@@ -6518,16 +6518,17 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   let staleReembedCount = 0;
   for (let i = 0; i < memories.length; i++) {
     const content = memories[i].content;
-    const cached = cache.get(content);
+    const memoryId = memories[i].uniqueId;
+    const cached = cache.get(memoryId);
     if (cached && cached.length === queryEmbedding.length) continue;
-    if (cached) cache.delete(content);
+    if (cached) cache.delete(memoryId);
     const storedModel = memories[i].embeddingModel;
     const modelCompatible = (storedModel ?? currentModel) === currentModel;
     if (memories[i].embedding && modelCompatible) {
       try {
         const parsed = JSON.parse(memories[i].embedding);
         if (Array.isArray(parsed) && parsed.length === queryEmbedding.length) {
-          cache.set(content, parsed);
+          cache.set(memoryId, Float32Array.from(parsed));
           continue;
         }
       } catch {
@@ -6545,7 +6546,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   if (uncachedTexts.length > 0) {
     const newEmbeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
     for (let j = 0; j < uncachedTexts.length; j++) {
-      cache.set(memories[uncachedIndices[j]].content, newEmbeddings[j]);
+      cache.set(memories[uncachedIndices[j]].uniqueId, Float32Array.from(newEmbeddings[j]));
       updateVaultMemoryEmbeddingOp(
         vaultCtx,
         memories[uncachedIndices[j]].uniqueId,
@@ -6561,7 +6562,7 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
   const embeddedItems = memories.map((m) => ({
     id: m.uniqueId,
     content: m.content,
-    embedding: cache.get(m.content) ?? [],
+    embedding: cache.get(m.uniqueId) ?? [],
     createdAt: m.createdAt,
     updatedAt: m.updatedAt,
     proofCount: m.proofCount,
@@ -13984,7 +13985,6 @@ async function retain(content, ctx, options = {}) {
     }
   }
   const embedding = await generateEmbedding(trimmed, ctx.embeddingOptions);
-  ctx.vaultCache.set(trimmed, embedding);
   const embeddingModel = ctx.embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const created = await createVaultMemoryOp(ctx.vaultCtx, {
     content: trimmed,
@@ -14005,6 +14005,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -14081,7 +14082,6 @@ async function tryConsolidate(trimmed, ctx, options) {
       options.sourceChunkIds ?? []
     );
     const newEmbedding = await generateEmbedding(decision.content, ctx.embeddingOptions);
-    ctx.vaultCache.set(decision.content, newEmbedding);
     const consolidatedModel = ctx.embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     const eventTimeUpdate = pickEventTimeUpdate(existing, options.eventTime);
     const updated = await updateVaultMemoryOp(ctx.vaultCtx, decision.targetId, {
@@ -14101,6 +14101,7 @@ async function tryConsolidate(trimmed, ctx, options) {
       await assertMergeTargetGoneOrThrow(ctx, decision.targetId);
       return null;
     }
+    ctx.vaultCache.set(decision.targetId, Float32Array.from(newEmbedding));
     return {
       action: "update",
       memoryId: decision.targetId,
@@ -15170,9 +15171,7 @@ function createMemoryVaultTool(vaultCtx, options, embeddingOptions, cache) {
             return `Error: Failed to update memory "${id}".`;
           }
           if (embeddingOptions && cache) {
-            if (previousContent) {
-              cache.delete(previousContent);
-            }
+            cache.delete(id);
             eagerEmbedContent(content, embeddingOptions, cache, vaultCtx, id).catch(
               // Silently swallow – SDK must not use console.*; embedding will be retried on next search
               () => {
@@ -17795,7 +17794,7 @@ function useChatStorage(options) {
       const result = await updateVaultMemoryOp(vaultCtx, id, { content, scope, embedding: null });
       if (result && getToken) {
         if (existing) {
-          vaultEmbeddingCacheRef.current.delete(existing.content);
+          vaultEmbeddingCacheRef.current.delete(id);
         }
         eagerEmbedContent(
           content,
@@ -17816,7 +17815,7 @@ function useChatStorage(options) {
       const existing = await getVaultMemoryOp(vaultCtx, id);
       const result = await deleteVaultMemoryOp(vaultCtx, id);
       if (result && existing) {
-        vaultEmbeddingCacheRef.current.delete(existing.content);
+        vaultEmbeddingCacheRef.current.delete(id);
       }
       return result;
     },

package/dist/react/index.d.mts

@@ -5857,7 +5857,9 @@ interface RecencyOptions {
 
 declare const DEFAULT_VAULT_CACHE_SIZE = 5000;
 /**
- * Create a VaultEmbeddingCache backed by an LRU with a default cap of 1000 entries.
+ * Create a VaultEmbeddingCache backed by an LRU with a default cap of
+ * DEFAULT_VAULT_CACHE_SIZE entries. Values are Float32Array — the model's
+ * native precision — not float64 number[], halving resident RAM losslessly.
  */
 declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCache;
 
@@ -5872,7 +5874,7 @@ declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCach
  * Embedding cache keyed by content string. Stores pre-computed embeddings
  * so that search only needs to embed the query, not the vault entries.
  */
-type VaultEmbeddingCache = Map<string, number[]>;
+type VaultEmbeddingCache = Map<string, Float32Array>;
 /**
  * Options for the vault search tool.
  */

package/dist/react/index.d.ts

@@ -5857,7 +5857,9 @@ interface RecencyOptions {
 
 declare const DEFAULT_VAULT_CACHE_SIZE = 5000;
 /**
- * Create a VaultEmbeddingCache backed by an LRU with a default cap of 1000 entries.
+ * Create a VaultEmbeddingCache backed by an LRU with a default cap of
+ * DEFAULT_VAULT_CACHE_SIZE entries. Values are Float32Array — the model's
+ * native precision — not float64 number[], halving resident RAM losslessly.
  */
 declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCache;
 
@@ -5872,7 +5874,7 @@ declare function createVaultEmbeddingCache(maxSize?: number): VaultEmbeddingCach
  * Embedding cache keyed by content string. Stores pre-computed embeddings
  * so that search only needs to embed the query, not the vault entries.
  */
-type VaultEmbeddingCache = Map<string, number[]>;
+type VaultEmbeddingCache = Map<string, Float32Array>;
 /**
  * Options for the vault search tool.
  */