@anuma/sdk 1.4.1 → 2.0.1

package/dist/expo/{chunk-RMZ4QAQV.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-UAL346D2.mjs");
+        const { recall: recall2 } = await import("./recall-DHOR5WWW.mjs");
         const result = await recall2(
           query,
           { vaultCtx, embeddingOptions, vaultCache: cache },

package/dist/expo/index.cjs

@@ -5,8 +5,13 @@ var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
 var __getOwnPropNames = Object.getOwnPropertyNames;
 var __getProtoOf = Object.getPrototypeOf;
 var __hasOwnProp = Object.prototype.hasOwnProperty;
-var __esm = (fn, res) => function __init() {
-  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+var __esm = (fn, res, err) => function __init() {
+  if (err) throw err[0];
+  try {
+    return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+  } catch (e) {
+    throw err = [e], e;
+  }
 };
 var __export = (target, all) => {
   for (var name in all)
@@ -6110,29 +6115,29 @@ 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 +6151,7 @@ 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 +6192,17 @@ 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 +6220,7 @@ 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 +6236,7 @@ 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 +12641,6 @@ 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 +12661,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -12732,7 +12738,6 @@ 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 +12757,7 @@ 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 +13827,7 @@ 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-RMZ4QAQV.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-UAL346D2.mjs → recall-DHOR5WWW.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-RMZ4QAQV.mjs";
+} from "./chunk-N4FEY34G.mjs";
 import "./chunk-DFVCS3QQ.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {

package/dist/index.cjs

@@ -3,8 +3,13 @@ var __defProp = Object.defineProperty;
 var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
 var __getOwnPropNames = Object.getOwnPropertyNames;
 var __hasOwnProp = Object.prototype.hasOwnProperty;
-var __esm = (fn, res) => function __init() {
-  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+var __esm = (fn, res, err) => function __init() {
+  if (err) throw err[0];
+  try {
+    return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+  } catch (e) {
+    throw err = [e], e;
+  }
 };
 var __export = (target, all) => {
   for (var name in all)

package/dist/react/{chunk-4FNDAPTF.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-L3QROVGM.mjs");
+        const { recall: recall2 } = await import("./recall-PUXEQHAP.mjs");
         const result = await recall2(
           query,
           { vaultCtx, embeddingOptions, vaultCache: cache },

package/dist/react/index.cjs

@@ -5,8 +5,13 @@ var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
 var __getOwnPropNames = Object.getOwnPropertyNames;
 var __getProtoOf = Object.getPrototypeOf;
 var __hasOwnProp = Object.prototype.hasOwnProperty;
-var __esm = (fn, res) => function __init() {
-  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+var __esm = (fn, res, err) => function __init() {
+  if (err) throw err[0];
+  try {
+    return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+  } catch (e) {
+    throw err = [e], e;
+  }
 };
 var __export = (target, all) => {
   for (var name in all)
@@ -6441,29 +6446,29 @@ 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 +6482,7 @@ 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 +6523,17 @@ 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 +6551,7 @@ 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 +6567,7 @@ 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 +13990,6 @@ 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 +14010,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -14081,7 +14087,6 @@ 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 +14106,7 @@ 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 +15176,7 @@ 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 +17799,7 @@ 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 +17820,7 @@ 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.
  */

package/dist/react/index.mjs

@@ -121,7 +121,7 @@ import {
   updateVaultMemoryEmbeddingOp,
   updateVaultMemoryOp,
   useEncryption
-} from "./chunk-4FNDAPTF.mjs";
+} from "./chunk-63IW5BS5.mjs";
 import {
   PiiRedactor,
   createStreamingDeAnonymizer,
@@ -6733,7 +6733,6 @@ 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,
@@ -6754,6 +6753,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -6830,7 +6830,6 @@ 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, {
@@ -6850,6 +6849,7 @@ 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,
@@ -7894,9 +7894,7 @@ 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
               () => {
@@ -10499,7 +10497,7 @@ 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,
@@ -10520,7 +10518,7 @@ 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/{recall-L3QROVGM.mjs → recall-PUXEQHAP.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-4FNDAPTF.mjs";
+} from "./chunk-63IW5BS5.mjs";
 import "./chunk-DFVCS3QQ.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {

package/dist/server/{chunk-FMGZTVXQ.mjs → chunk-3RPGXXPZ.mjs}

@@ -5,7 +5,11 @@ var __getOwnPropNames = Object.getOwnPropertyNames;
 var __getProtoOf = Object.getPrototypeOf;
 var __hasOwnProp = Object.prototype.hasOwnProperty;
 var __commonJS = (cb, mod) => function __require() {
-  return mod || (0, cb[__getOwnPropNames(cb)[0]])((mod = { exports: {} }).exports, mod), mod.exports;
+  try {
+    return mod || (0, cb[__getOwnPropNames(cb)[0]])((mod = { exports: {} }).exports, mod), mod.exports;
+  } catch (e) {
+    throw mod = 0, e;
+  }
 };
 var __copyProps = (to, from, except, desc) => {
   if (from && typeof from === "object" || typeof from === "function") {

package/dist/server/{chunk-SJLYHCMG.mjs → chunk-45GOZ3AB.mjs}

@@ -1,13 +1,13 @@
 import {
   callPortalJsonCompletion,
   getLogger
-} from "./chunk-RCCKRTBI.mjs";
+} from "./chunk-AHJT66EO.mjs";
 import {
   BASE_URL,
   __commonJS,
   __toESM,
   createClientConfig
-} from "./chunk-FMGZTVXQ.mjs";
+} from "./chunk-3RPGXXPZ.mjs";
 
 // node_modules/.pnpm/@babel+runtime@7.26.0/node_modules/@babel/runtime/helpers/interopRequireDefault.js
 var require_interopRequireDefault = __commonJS({
@@ -18531,29 +18531,29 @@ 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],
@@ -18567,7 +18567,7 @@ 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(
@@ -18608,16 +18608,17 @@ 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 {
@@ -18635,7 +18636,7 @@ 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,
@@ -18651,7 +18652,7 @@ 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,
@@ -18847,7 +18848,7 @@ function createMemoryVaultSearchTool(vaultCtx, embeddingOptions, cache, searchOp
             legacy.map((r) => ({ id: r.uniqueId, content: r.content, score: r.similarity }))
           );
         }
-        const { recall: recall2 } = await import("./recall-UHJN6H5V.mjs");
+        const { recall: recall2 } = await import("./recall-RMBNCEG6.mjs");
         const result = await recall2(
           query,
           { vaultCtx, embeddingOptions, vaultCache: cache },

package/dist/server/{chunk-RCCKRTBI.mjs → chunk-AHJT66EO.mjs}

@@ -1,6 +1,6 @@
 import {
   BASE_URL
-} from "./chunk-FMGZTVXQ.mjs";
+} from "./chunk-3RPGXXPZ.mjs";
 
 // src/lib/logger.ts
 var consoleLogger = {

package/dist/server/{clientConfig-GMADOBPK.mjs → clientConfig-VIMT6Q32.mjs}

@@ -3,7 +3,7 @@ import {
   MCP_R2_DOMAIN,
   createClientConfig,
   resolveBaseUrl
-} from "./chunk-FMGZTVXQ.mjs";
+} from "./chunk-3RPGXXPZ.mjs";
 export {
   BASE_URL,
   MCP_R2_DOMAIN,

package/dist/server/{consolidate-A2ZIX5WF.mjs → consolidate-TDT4HRHG.mjs}

@@ -4,8 +4,8 @@ import {
 import {
   callPortalJsonCompletion,
   getLogger
-} from "./chunk-RCCKRTBI.mjs";
-import "./chunk-FMGZTVXQ.mjs";
+} from "./chunk-AHJT66EO.mjs";
+import "./chunk-3RPGXXPZ.mjs";
 
 // src/lib/memory/consolidate.ts
 var DEFAULT_MODEL = "inclusionai/ling-2.6-flash";

package/dist/server/index.cjs

@@ -5,11 +5,20 @@ var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
 var __getOwnPropNames = Object.getOwnPropertyNames;
 var __getProtoOf = Object.getPrototypeOf;
 var __hasOwnProp = Object.prototype.hasOwnProperty;
-var __esm = (fn, res) => function __init() {
-  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+var __esm = (fn, res, err) => function __init() {
+  if (err) throw err[0];
+  try {
+    return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+  } catch (e) {
+    throw err = [e], e;
+  }
 };
 var __commonJS = (cb, mod) => function __require() {
-  return mod || (0, cb[__getOwnPropNames(cb)[0]])((mod = { exports: {} }).exports, mod), mod.exports;
+  try {
+    return mod || (0, cb[__getOwnPropNames(cb)[0]])((mod = { exports: {} }).exports, mod), mod.exports;
+  } catch (e) {
+    throw mod = 0, e;
+  }
 };
 var __export = (target, all) => {
   for (var name in all)
@@ -20246,29 +20255,29 @@ 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],
@@ -20282,7 +20291,7 @@ 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(
@@ -20323,16 +20332,17 @@ 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 {
@@ -20350,7 +20360,7 @@ 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,
@@ -20366,7 +20376,7 @@ 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,
@@ -25527,9 +25537,7 @@ 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
               () => {
@@ -25640,7 +25648,6 @@ 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,
@@ -25661,6 +25668,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -25737,7 +25745,6 @@ 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, {
@@ -25757,6 +25764,7 @@ 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,

package/dist/server/index.d.mts

@@ -5311,7 +5311,9 @@ 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;
 
@@ -5326,7 +5328,7 @@ 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/server/index.d.ts

@@ -5311,7 +5311,9 @@ 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;
 
@@ -5326,7 +5328,7 @@ 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/server/index.mjs

@@ -95,7 +95,7 @@ import {
   updateVaultMemoryEmbeddingOp,
   updateVaultMemoryOp,
   upsertMessageOp
-} from "./chunk-SJLYHCMG.mjs";
+} from "./chunk-45GOZ3AB.mjs";
 import {
   createStreamingDeAnonymizer,
   resolvePiiRedactor
@@ -106,13 +106,13 @@ import {
   getLogger,
   resolvePortalAuthHeaders,
   supportsResponseFormat
-} from "./chunk-RCCKRTBI.mjs";
+} from "./chunk-AHJT66EO.mjs";
 import {
   BASE_URL,
   __commonJS,
   __decorateClass,
   __toESM
-} from "./chunk-FMGZTVXQ.mjs";
+} from "./chunk-3RPGXXPZ.mjs";
 
 // node_modules/.pnpm/@nozbe+watermelondb@0.28.0/node_modules/@nozbe/watermelondb/utils/fp/sortBy/index.js
 var require_sortBy = __commonJS({
@@ -4595,9 +4595,7 @@ 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
               () => {
@@ -4699,7 +4697,6 @@ 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,
@@ -4720,6 +4717,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -4759,7 +4757,7 @@ async function tryConsolidate(trimmed, ctx, options) {
     content: m.content,
     similarity: m.similarity
   }));
-  const { consolidateMemory: doConsolidate } = await import("./consolidate-A2ZIX5WF.mjs");
+  const { consolidateMemory: doConsolidate } = await import("./consolidate-TDT4HRHG.mjs");
   const decision = await doConsolidate(trimmed, candidates, consolidateOptions);
   if (decision.action === "create") return null;
   if (decision.action === "noop" && decision.targetId) {
@@ -4796,7 +4794,6 @@ 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, {
@@ -4816,6 +4813,7 @@ 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,
@@ -5768,7 +5766,7 @@ async function getServerTools(options) {
     return cached.tools;
   }
   try {
-    const { BASE_URL: BASE_URL2 } = await import("./clientConfig-GMADOBPK.mjs");
+    const { BASE_URL: BASE_URL2 } = await import("./clientConfig-VIMT6Q32.mjs");
     const effectiveBaseUrl = baseUrl ?? BASE_URL2;
     if (apiKey) {
       const response = await fetch(`${effectiveBaseUrl}/api/v1/tools`, {

package/dist/server/recall-RMBNCEG6.mjs

@@ -0,0 +1,8 @@
+import {
+  recall
+} from "./chunk-45GOZ3AB.mjs";
+import "./chunk-AHJT66EO.mjs";
+import "./chunk-3RPGXXPZ.mjs";
+export {
+  recall
+};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "1.4.1",
+  "version": "2.0.1",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",
@@ -200,6 +200,7 @@
     "typescript": "^5.9.3",
     "typescript-eslint": "^8.56.0",
     "unist-util-visit": "^5.0.0",
+    "vite": ">=7.3.5 <8.0.0",
     "vitest": "^4.1.8"
   },
   "peerDependencies": {
@@ -235,22 +236,30 @@
     "overrides": {
       "react": ">=19.0.1 <19.1.0 || >=19.1.2 <19.2.0 || >=19.2.1",
       "react-dom": ">=19.0.1 <19.1.0 || >=19.1.2 <19.2.0 || >=19.2.1",
-      "tar": ">=7.5.11",
+      "tar": ">=7.5.16",
       "picomatch": ">=2.3.2 <3.0.0 || >=4.0.4",
       "@xmldom/xmldom": ">=0.8.13 <0.9.0",
       "protobufjs": ">=7.6.1 <8.0.0",
       "tmp": ">=0.2.7",
       "ws@6": ">=6.2.4 <7.0.0",
       "ws@7": ">=7.5.11 <8.0.0",
+      "ws@8": ">=8.21.0 <9.0.0",
       "shell-quote": ">=1.8.4",
       "axios": ">=1.16.0 <2.0.0",
       "hono": ">=4.12.21 <5.0.0",
       "js-cookie": ">=3.0.7 <4.0.0",
       "postcss": ">=8.5.10 <9.0.0",
-      "dompurify": ">=3.4.0 <4.0.0",
+      "dompurify": ">=3.4.11 <4.0.0",
       "lodash": ">=4.18.0 <5.0.0",
       "minimatch@10": ">=10.2.3 <11.0.0",
-      "uuid@<11.1.1": ">=11.1.1 <12.0.0"
+      "uuid@<11.1.1": ">=11.1.1 <12.0.0",
+      "form-data@>=4 <4.0.6": ">=4.0.6",
+      "undici@>=7 <7.28.0": ">=7.28.0 <8.0.0",
+      "vite@>=7 <7.3.5": ">=7.3.5",
+      "esbuild@>=0.27.3 <0.28.1": ">=0.28.1",
+      "js-yaml@>=4 <4.2.0": ">=4.2.0 <5.0.0",
+      "js-yaml@>=3 <3.15.0": ">=3.15.0 <4.0.0",
+      "markdown-it@<14.2.0": ">=14.2.0"
     },
     "patchedDependencies": {
       "tesseract.js@6.0.1": "patches/tesseract.js@6.0.1.patch",

package/dist/server/recall-UHJN6H5V.mjs

@@ -1,8 +0,0 @@
-import {
-  recall
-} from "./chunk-SJLYHCMG.mjs";
-import "./chunk-RCCKRTBI.mjs";
-import "./chunk-FMGZTVXQ.mjs";
-export {
-  recall
-};