@anuma/sdk 1.4.0 → 2.0.0

package/dist/react/index.mjs

@@ -121,7 +121,7 @@ import {
   updateVaultMemoryEmbeddingOp,
   updateVaultMemoryOp,
   useEncryption
-} from "./chunk-SEZCFG24.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-UWH5HU6V.mjs → recall-PUXEQHAP.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-SEZCFG24.mjs";
+} from "./chunk-63IW5BS5.mjs";
 import "./chunk-DFVCS3QQ.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {

package/dist/server/{chunk-Z7CJGOTG.mjs → chunk-P7KA2HAM.mjs}

@@ -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-24ZIA2VL.mjs");
+        const { recall: recall2 } = await import("./recall-JJ3E74XY.mjs");
         const result = await recall2(
           query,
           { vaultCtx, embeddingOptions, vaultCache: cache },

package/dist/server/index.cjs

@@ -20246,29 +20246,29 @@ 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 +20282,7 @@ 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 +20323,17 @@ 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 +20351,7 @@ 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 +20367,7 @@ 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 +25528,7 @@ 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 +25639,6 @@ 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 +25659,7 @@ async function retain(content, ctx, options = {}) {
       }
     }
   });
+  ctx.vaultCache.set(created.uniqueId, Float32Array.from(embedding));
   return {
     action: "create",
     memoryId: created.uniqueId,
@@ -25737,7 +25736,6 @@ 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 +25755,7 @@ 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-Z7CJGOTG.mjs";
+} from "./chunk-P7KA2HAM.mjs";
 import {
   createStreamingDeAnonymizer,
   resolvePiiRedactor
@@ -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,
@@ -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,

package/dist/server/{recall-24ZIA2VL.mjs → recall-JJ3E74XY.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-Z7CJGOTG.mjs";
+} from "./chunk-P7KA2HAM.mjs";
 import "./chunk-RCCKRTBI.mjs";
 import "./chunk-FMGZTVXQ.mjs";
 export {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "1.4.0",
+  "version": "2.0.0",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",
@@ -145,7 +145,7 @@
   "homepage": "https://github.com/anuma-ai/sdk#readme",
   "packageManager": "pnpm@10.23.0",
   "dependencies": {
-    "@anuma/portal": "1.0.0-next.20260709030454",
+    "@anuma/portal": "1.0.0-next.20260715224031",
     "@babel/parser": "^7.29.2",
     "@babel/types": "^7.29.0",
     "ai": "6.0.202",