From 85753a76ef9ff93233a4e0bfdb8a46144bb7d261 Mon Sep 17 00:00:00 2001
From: Bo26fhmC5M <88071760+Bo26fhmC5M@users.noreply.github.com>
Date: Sat, 10 May 2025 15:00:35 +0900
Subject: [PATCH] feat: implement preset and parallel requests for HypaV3

---
 package.json                                  |    1 +
 pnpm-lock.yaml                                |   16 +
 src/App.svelte                                |    9 +
 src/lib/ChatScreens/DefaultChatScreen.svelte  |   12 +-
 src/lib/Others/AlertComp.svelte               |    9 +-
 src/lib/Others/HypaV3Modal.svelte             |   70 +-
 src/lib/Others/HypaV3Progress.svelte          |   55 +
 src/lib/Setting/Pages/OtherBotSettings.svelte |  277 ++-
 src/lib/SideBars/CharConfig.svelte            |    8 +-
 src/lib/SideBars/Toggles.svelte               |    8 +-
 src/ts/alert.ts                               |    9 +-
 src/ts/process/memory/hypamemory.ts           |   41 +-
 src/ts/process/memory/hypamemoryv2.ts         |  414 ++++
 src/ts/process/memory/hypav3.ts               | 1820 ++++++++++++-----
 src/ts/process/memory/taskRateLimiter.ts      |  188 ++
 src/ts/process/transformers.ts                |    7 +
 src/ts/process/webllm.ts                      |   60 +
 src/ts/storage/database.svelte.ts             |   43 +-
 src/ts/stores.svelte.ts                       |    7 +
 19 files changed, 2382 insertions(+), 672 deletions(-)
 create mode 100644 src/lib/Others/HypaV3Progress.svelte
 create mode 100644 src/ts/process/memory/hypamemoryv2.ts
 create mode 100644 src/ts/process/memory/taskRateLimiter.ts
 create mode 100644 src/ts/process/webllm.ts

diff --git a/package.json b/package.json
index 5e944acf..2e9ef071 100644
--- a/package.json
+++ b/package.json
@@ -26,6 +26,7 @@
     "@dqbd/tiktoken": "^1.0.7",
     "@huggingface/jinja": "^0.2.2",
     "@huggingface/transformers": "^3.1.1",
+    "@mlc-ai/web-llm": "^0.2.79",
     "@mlc-ai/web-tokenizers": "^0.1.2",
     "@risuai/ccardlib": "^0.4.1",
     "@rollup/rollup-win32-arm64-msvc": "^4.40.1",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 71ece55d..e6612fef 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -38,6 +38,9 @@ importers:
       '@huggingface/transformers':
         specifier: ^3.1.1
         version: 3.1.1
+      '@mlc-ai/web-llm':
+        specifier: ^0.2.79
+        version: 0.2.79
       '@mlc-ai/web-tokenizers':
         specifier: ^0.1.2
         version: 0.1.2
@@ -786,6 +789,9 @@ packages:
     resolution: {integrity: sha512-Yhlar6v9WQgUp/He7BdgzOz8lqMQ8sU+jkCq7Wx8Myc5YFJLbEe7lgui/V7G1qB1DJykHSGwreceSaD60Y0PUQ==}
     hasBin: true
 
+  '@mlc-ai/web-llm@0.2.79':
+    resolution: {integrity: sha512-Hy1ZHQ0o2bZGZoVnGK48+fts/ZSKwLe96xjvqL/6C59Mem9HoHTcFE07NC2E23mRmhd01tL655N6CPeYmwWgwQ==}
+
   '@mlc-ai/web-tokenizers@0.1.2':
     resolution: {integrity: sha512-zpeZjiw8vTuaSjIeoEDYJL1nOrhSL/7eAW7hhqd3CzIodLpuArfyiTZxKwIlew8kkqb0RIxooqeAqtlngh1p8g==}
 
@@ -2508,6 +2514,10 @@ packages:
   lodash@4.17.21:
     resolution: {integrity: sha512-v2kDEe57lecTulaDIuNTPy3Ry4gLGJ6Z1O3vE1krgXZNrsQ+LFTGHVxVjcXPs17LhbZVGedAJv8XZ1tvj5FvSg==}
 
+  loglevel@1.9.2:
+    resolution: {integrity: sha512-HgMmCqIJSAKqo68l0rS2AanEWfkxaZ5wNiEFb5ggm08lDs9Xl2KxBlX3PTcaD2chBM1gXAYf491/M2Rv8Jwayg==}
+    engines: {node: '>= 0.6.0'}
+
   long@5.2.3:
     resolution: {integrity: sha512-lcHwpNoggQTObv5apGNCTdJrO69eHOZMi4BNC+rTLER8iHAqGrUVeLh/irVIM7zTw2bOXA8T6uNPeujwOLg/2Q==}
 
@@ -4366,6 +4376,10 @@ snapshots:
       - supports-color
     optional: true
 
+  '@mlc-ai/web-llm@0.2.79':
+    dependencies:
+      loglevel: 1.9.2
+
   '@mlc-ai/web-tokenizers@0.1.2': {}
 
   '@msgpack/msgpack@2.8.0': {}
@@ -6083,6 +6097,8 @@ snapshots:
 
   lodash@4.17.21: {}
 
+  loglevel@1.9.2: {}
+
   long@5.2.3: {}
 
   lru-cache@10.1.0: {}
diff --git a/src/App.svelte b/src/App.svelte
index 7d21e2d8..be41032c 100644
--- a/src/App.svelte
+++ b/src/App.svelte
@@ -19,6 +19,9 @@
     import CustomGUISettingMenu from './lib/Setting/Pages/CustomGUISettingMenu.svelte';
     import { checkCharOrder } from './ts/globalApi.svelte';
     import { ArrowUpIcon, GlobeIcon, PlusIcon } from 'lucide-svelte';
+    import { hypaV3ModalOpen, hypaV3ProgressStore } from "./ts/stores.svelte";
+    import HypaV3Modal from './lib/Others/HypaV3Modal.svelte';
+    import HypaV3Progress from './lib/Others/HypaV3Progress.svelte';
 
   
     let didFirstSetup: boolean  = $derived(DBState.db?.didFirstSetup)
@@ -167,5 +170,11 @@
     {#if $openPersonaList}
         <ListedPersona close={() => {$openPersonaList = false}} />
     {/if}
+    {#if $hypaV3ModalOpen}
+        <HypaV3Modal />
+    {/if}
     <SavePopupIconComp />
+    {#if $hypaV3ProgressStore.open}
+        <HypaV3Progress />
+    {/if}
 </main>
\ No newline at end of file
diff --git a/src/lib/ChatScreens/DefaultChatScreen.svelte b/src/lib/ChatScreens/DefaultChatScreen.svelte
index f4449e37..6fa1078e 100644
--- a/src/lib/ChatScreens/DefaultChatScreen.svelte
+++ b/src/lib/ChatScreens/DefaultChatScreen.svelte
@@ -3,7 +3,7 @@
     import Suggestion from './Suggestion.svelte';
     import AdvancedChatEditor from './AdvancedChatEditor.svelte';
     import { CameraIcon, DatabaseIcon, DicesIcon, GlobeIcon, ImagePlusIcon, LanguagesIcon, Laugh, MenuIcon, MicOffIcon, PackageIcon, Plus, RefreshCcwIcon, ReplyIcon, Send, StepForwardIcon, XIcon, BrainIcon } from "lucide-svelte";
-    import { selectedCharID, PlaygroundStore, createSimpleCharacter } from "../../ts/stores.svelte";
+    import { selectedCharID, PlaygroundStore, createSimpleCharacter, hypaV3ModalOpen } from "../../ts/stores.svelte";
     import Chat from "./Chat.svelte";
     import { type Message, type character, type groupChat } from "../../ts/storage/database.svelte";
     import { DBState } from 'src/ts/stores.svelte';
@@ -12,7 +12,7 @@
     import { findCharacterbyId, getUserIconProtrait, messageForm, sleep } from "../../ts/util";
     import { language } from "../../lang";
     import { isExpTranslator, translate } from "../../ts/translator/translator";
-    import { alertError, alertNormal, alertWait, showHypaV2Alert, showHypaV3Alert } from "../../ts/alert";
+    import { alertError, alertNormal, alertWait, showHypaV2Alert } from "../../ts/alert";
     import sendSound from '../../etc/send.mp3'
     import { processScript } from "src/ts/process/scripts";
     import CreatorQuote from "./CreatorQuote.svelte";
@@ -829,17 +829,17 @@
                     {/if}
 
                     {#if DBState.db.showMenuHypaMemoryModal}
-                        {#if DBState.db.supaModelType !== 'none' && (DBState.db.hypav2 || DBState.db.hypaV3)}
+                        {#if (DBState.db.supaModelType !== 'none' && DBState.db.hypav2) || DBState.db.hypaV3}
                             <div class="flex items-center cursor-pointer hover:text-green-500 transition-colors" onclick={() => {
                                 if (DBState.db.hypav2) {
                                     DBState.db.characters[$selectedCharID].chats[DBState.db.characters[$selectedCharID].chatPage].hypaV2Data ??= {
                                         lastMainChunkID: 0,
                                         mainChunks: [],
-                                        chunks: [],
+                                        chunks: []
                                     }
-                                    showHypaV2Alert();
+                                    showHypaV2Alert()
                                 } else if (DBState.db.hypaV3) {
-                                    showHypaV3Alert();
+                                    $hypaV3ModalOpen = true
                                 }
 
                                 openMenu = false
diff --git a/src/lib/Others/AlertComp.svelte b/src/lib/Others/AlertComp.svelte
index 7801aee6..f14b789a 100644
--- a/src/lib/Others/AlertComp.svelte
+++ b/src/lib/Others/AlertComp.svelte
@@ -23,9 +23,8 @@
     import Help from "./Help.svelte";
     import { getChatBranches } from "src/ts/gui/branches";
     import { getCurrentCharacter } from "src/ts/storage/database.svelte";
-    import { message } from "@tauri-apps/plugin-dialog";
-    import HypaV3Modal from './HypaV3Modal.svelte';
-  import Googli from "../UI/Googli.svelte";
+    import Googli from "../UI/Googli.svelte";
+
     let btn
     let input = $state('')
     let cardExportType = $state('realm')
@@ -651,10 +650,6 @@
     </div>
 {/if}
 
-{#if $alertStore.type === "hypaV3"}
-    <HypaV3Modal />
-{/if}
-
 <style>
     .break-any{
         word-break: normal;
diff --git a/src/lib/Others/HypaV3Modal.svelte b/src/lib/Others/HypaV3Modal.svelte
index 14975b2a..63ee1be4 100644
--- a/src/lib/Others/HypaV3Modal.svelte
+++ b/src/lib/Others/HypaV3Modal.svelte
@@ -14,23 +14,22 @@
     CheckIcon,
   } from "lucide-svelte";
   import {
-    alertConfirm,
-    alertNormalWait,
-    showHypaV3Alert,
-  } from "../../ts/alert";
+    summarize,
+    getCurrentHypaV3Preset,
+  } from "src/ts/process/memory/hypav3";
+  import { type OpenAIChat } from "src/ts/process/index.svelte";
+  import { processScriptFull, risuChatParser } from "src/ts/process/scripts";
+  import { type Message } from "src/ts/storage/database.svelte";
+  import { translateHTML } from "src/ts/translator/translator";
+  import { alertConfirm, alertNormalWait } from "src/ts/alert";
   import {
     DBState,
-    alertStore,
     selectedCharID,
     settingsOpen,
     SettingsMenuIndex,
-  } from "../../ts/stores.svelte";
-  import { type OpenAIChat } from "../../ts/process/index.svelte";
-  import { processScriptFull, risuChatParser } from "../../ts/process/scripts";
-  import { summarize } from "../../ts/process/memory/hypav3";
-  import { type Message } from "../../ts/storage/database.svelte";
-  import { translateHTML } from "../../ts/translator/translator";
-  import { language } from "../../lang";
+    hypaV3ModalOpen,
+  } from "src/ts/stores.svelte";
+  import { language } from "src/lang";
 
   interface SummaryUI {
     originalRef: HTMLTextAreaElement;
@@ -452,9 +451,7 @@
 
       const summarizeResult = await summarize(toSummarize);
 
-      if (summarizeResult.success) {
-        summaryUIState.rerolledText = summarizeResult.data;
-      }
+      summaryUIState.rerolledText = summarizeResult;
     } catch (error) {
       summaryUIState.rerolledText = "Reroll failed";
     } finally {
@@ -507,7 +504,7 @@
       return null;
     }
 
-    return DBState.db.hypaV3Settings.processRegexScript
+    return getCurrentHypaV3Preset().settings.processRegexScript
       ? await processRegexScript(unprocessed)
       : unprocessed;
   }
@@ -642,7 +639,7 @@
       return null;
     }
 
-    return DBState.db.hypaV3Settings.processRegexScript
+    return getCurrentHypaV3Preset().settings.processRegexScript
       ? await processRegexScript(unprocessed)
       : unprocessed;
   }
@@ -826,7 +823,7 @@
 </script>
 
 <!-- Modal backdrop -->
-<div class="fixed inset-0 z-50 p-1 sm:p-2 bg-black/50">
+<div class="fixed inset-0 z-40 p-1 sm:p-2 bg-black/50">
   <!-- Modal wrapper -->
   <div class="flex justify-center w-full h-full">
     <!-- Modal window -->
@@ -877,13 +874,9 @@
             class="p-2 text-zinc-400 hover:text-zinc-200 transition-colors"
             tabindex="-1"
             onclick={() => {
-              alertStore.set({
-                type: "none",
-                msg: "",
-              });
-
-              settingsOpen.set(true);
-              SettingsMenuIndex.set(2); // Other bot settings
+              $hypaV3ModalOpen = false;
+              $settingsOpen = true;
+              $SettingsMenuIndex = 2; // Other bot settings
             }}
           >
             <SettingsIcon class="w-6 h-6" />
@@ -906,8 +899,6 @@
                   summaries: [],
                   lastSelectedSummaries: [],
                 };
-              } else {
-                showHypaV3Alert();
               }
             }}
           >
@@ -919,10 +910,7 @@
             class="p-2 text-zinc-400 hover:text-zinc-200 transition-colors"
             tabindex="-1"
             onclick={() => {
-              alertStore.set({
-                type: "none",
-                msg: "",
-              });
+              $hypaV3ModalOpen = false;
             }}
           >
             <XIcon class="w-6 h-6" />
@@ -960,8 +948,6 @@
                         )
                       );
                     }
-
-                    showHypaV3Alert();
                   }}
                 >
                   {language.hypaV3Modal.convertButton}
@@ -976,7 +962,7 @@
 
           <!-- Search Bar -->
         {:else if searchUIState}
-          <div class="sticky top-0 z-50 p-2 sm:p-3 bg-zinc-800">
+          <div class="sticky top-0 z-40 p-2 sm:p-3 bg-zinc-800">
             <div class="flex items-center gap-2">
               <div class="relative flex flex-1 items-center">
                 <form
@@ -1093,17 +1079,17 @@
                     <button
                       class="p-2 text-zinc-400 hover:text-rose-300 transition-colors"
                       tabindex="-1"
-                      disabled={!isOrphan(i)}
                       onclick={async () => {
                         if (
-                          await alertConfirm(language.hypaV3Modal.deleteThisConfirmMessage)
+                          await alertConfirm(
+                            language.hypaV3Modal.deleteThisConfirmMessage
+                          )
                         ) {
-                          hypaV3DataState.summaries = hypaV3DataState.summaries.filter(
-                            (_, index) => index !== i
-                          );
+                          hypaV3DataState.summaries =
+                            hypaV3DataState.summaries.filter(
+                              (_, index) => index !== i
+                            );
                         }
-
-                        showHypaV3Alert();
                       }}
                     >
                       <Trash2Icon class="w-4 h-4" />
@@ -1122,8 +1108,6 @@
                         ) {
                           hypaV3DataState.summaries.splice(i + 1);
                         }
-
-                        showHypaV3Alert();
                       }}
                     >
                       <ScissorsLineDashed class="w-4 h-4" />
diff --git a/src/lib/Others/HypaV3Progress.svelte b/src/lib/Others/HypaV3Progress.svelte
new file mode 100644
index 00000000..35d098ab
--- /dev/null
+++ b/src/lib/Others/HypaV3Progress.svelte
@@ -0,0 +1,55 @@
+<script lang="ts">
+  import { hypaV3ProgressStore } from "src/ts/stores.svelte";
+
+  let isExpanded = $state(false);
+
+  const toggleExpand = () => {
+    isExpanded = !isExpanded;
+  };
+</script>
+
+{#if isExpanded}
+  <div
+    class="absolute w-full h-full z-40 flex justify-center items-center pointer-events-none"
+  >
+    <button
+      class="bg-darkbg p-4 break-any rounded-md flex flex-col max-w-3xl max-h-full overflow-y-auto transition-opacity duration-300 pointer-events-auto"
+      type="button"
+      onclick={toggleExpand}
+    >
+      <span class="mb-6 text-left text-gray-500 text-sm"
+        >{$hypaV3ProgressStore.msg || ""}</span
+      >
+      <div
+        class="w-full min-w-64 md:min-w-138 h-2 bg-darkbg border border-darkborderc rounded-md"
+      >
+        <div
+          class="h-full bg-gradient-to-r from-blue-500 to-purple-800 saving-animation transition-[width]"
+        ></div>
+      </div>
+      <span class="w-full mt-6 text-center text-gray-500 text-sm"
+        >{$hypaV3ProgressStore.subMsg || ""}</span
+      >
+    </button>
+  </div>
+{:else}
+  <button
+    class="fixed top-4 right-4 z-40 bg-darkbg p-2 rounded-full shadow-lg transition-all duration-300 flex items-center justify-center"
+    type="button"
+    style="opacity: 0.8;"
+    onclick={toggleExpand}
+    onmouseenter={(e) => (e.currentTarget.style.opacity = "1")}
+    onmouseleave={(e) => (e.currentTarget.style.opacity = "0.8")}
+  >
+    <div class="w-8 h-8 relative">
+      <div
+        class="absolute inset-0 border-t-2 border-red-500 rounded-full animate-spin"
+      ></div>
+      <div
+        class="absolute inset-1 flex items-center justify-center text-xs text-gray-300"
+      >
+        {$hypaV3ProgressStore.miniMsg || ""}
+      </div>
+    </div>
+  </button>
+{/if}
diff --git a/src/lib/Setting/Pages/OtherBotSettings.svelte b/src/lib/Setting/Pages/OtherBotSettings.svelte
index 474eb59b..2698b6a9 100644
--- a/src/lib/Setting/Pages/OtherBotSettings.svelte
+++ b/src/lib/Setting/Pages/OtherBotSettings.svelte
@@ -3,15 +3,13 @@
     import { language } from "src/lang";
     import Help from "src/lib/Others/Help.svelte";
     import { selectSingleFile } from "src/ts/util";
-    
-    import { DBState } from 'src/ts/stores.svelte';
-    import { isTauri, saveAsset } from "src/ts/globalApi.svelte";
+    import { DBState, selectedCharID } from 'src/ts/stores.svelte';
+    import { isTauri, saveAsset, downloadFile } from "src/ts/globalApi.svelte";
     import NumberInput from "src/lib/UI/GUI/NumberInput.svelte";
     import TextInput from "src/lib/UI/GUI/TextInput.svelte";
     import SelectInput from "src/lib/UI/GUI/SelectInput.svelte";
     import OptionInput from "src/lib/UI/GUI/OptionInput.svelte";
     import SliderInput from "src/lib/UI/GUI/SliderInput.svelte";
-    import Button from "src/lib/UI/GUI/Button.svelte";
     import { getCharImage } from "src/ts/characters";
     import Arcodion from "src/lib/UI/Arcodion.svelte";
     import CheckInput from "src/lib/UI/GUI/CheckInput.svelte";
@@ -19,7 +17,9 @@
     import { untrack } from "svelte";
     import { tokenizePreset } from "src/ts/process/prompt";
     import { getCharToken } from "src/ts/tokenizer";
-    import { selectedCharID } from "src/ts/stores.svelte";
+    import { PlusIcon, PencilIcon, TrashIcon, DownloadIcon, FolderUpIcon } from "lucide-svelte";
+    import { alertError, alertInput, alertConfirm, alertNormal } from "src/ts/alert";
+    import { createHypaV3Preset } from "src/ts/process/memory/hypav3";
 
     $effect.pre(() => {
         DBState.db.NAIImgConfig ??= {
@@ -67,25 +67,35 @@
 
     // HypaV3
     $effect(() => {
-        const newValue = Math.min(DBState.db.hypaV3Settings.recentMemoryRatio, 1);
+        const settings = DBState.db.hypaV3Presets?.[DBState.db.hypaV3PresetId]?.settings;
+        const currentValue = settings?.similarMemoryRatio;
+
+        if (!currentValue) return;
 
         untrack(() => {
-            DBState.db.hypaV3Settings.recentMemoryRatio = newValue;
-            
-            if (newValue + DBState.db.hypaV3Settings.similarMemoryRatio > 1) {
-                DBState.db.hypaV3Settings.similarMemoryRatio = 1 - newValue;
+            const newValue = Math.min(currentValue, 1);
+
+            settings.similarMemoryRatio = newValue;
+
+            if (newValue + settings.recentMemoryRatio > 1) {
+                settings.recentMemoryRatio = 1 - newValue;
             }
         })
     });
 
     $effect(() => {
-        const newValue = Math.min(DBState.db.hypaV3Settings.similarMemoryRatio, 1);
+        const settings = DBState.db.hypaV3Presets?.[DBState.db.hypaV3PresetId]?.settings;
+        const currentValue = settings?.recentMemoryRatio;
+
+        if (!currentValue) return;
 
         untrack(() => {
-            DBState.db.hypaV3Settings.similarMemoryRatio = newValue;
+            const newValue = Math.min(currentValue, 1);
 
-            if (newValue + DBState.db.hypaV3Settings.recentMemoryRatio > 1) {
-                DBState.db.hypaV3Settings.recentMemoryRatio = 1 - newValue;
+            settings.recentMemoryRatio = newValue;
+
+            if (newValue + settings.similarMemoryRatio > 1) {
+                settings.similarMemoryRatio = 1 - newValue;
             }
         })
     });
@@ -542,20 +552,11 @@
                 DBState.db.hanuraiEnable = false
                 DBState.db.hypaV3 = false
             } else if (value === 'hypaV3') {
-                DBState.db.supaModelType = 'subModel'
                 DBState.db.memoryAlgorithmType = 'hypaMemoryV3'
-                DBState.db.hypav2 = false
+                DBState.db.supaModelType = 'none'
                 DBState.db.hanuraiEnable = false
+                DBState.db.hypav2 = false
                 DBState.db.hypaV3 = true
-                DBState.db.hypaV3Settings.memoryTokensRatio = 0.2
-                DBState.db.hypaV3Settings.extraSummarizationRatio = 0
-                DBState.db.hypaV3Settings.maxChatsPerSummary = 4
-                DBState.db.hypaV3Settings.recentMemoryRatio = 0.4
-                DBState.db.hypaV3Settings.similarMemoryRatio = 0.4
-                DBState.db.hypaV3Settings.enableSimilarityCorrection = false
-                DBState.db.hypaV3Settings.preserveOrphanedMemory = false
-                DBState.db.hypaV3Settings.processRegexScript = false
-                DBState.db.hypaV3Settings.doNotSummarizeUserMessage = false
             } else {
                 DBState.db.supaModelType = 'none'
                 DBState.db.memoryAlgorithmType = 'none'
@@ -597,46 +598,191 @@
             <span class="text-textcolor">{language.hypaAllocatedTokens}</span>
             <NumberInput size="sm" marginBottom bind:value={DBState.db.hypaAllocatedTokens} min={100} />
         {:else if DBState.db.hypaV3}
-            <span class="mb-2 text-textcolor2 text-sm text-wrap break-words max-w-full">{language.hypaV3Settings.descriptionLabel}</span>
-            <span class="text-textcolor mt-4">{language.SuperMemory} {language.model}</span>
-            <SelectInput className="mt-2 mb-2" bind:value={DBState.db.supaModelType}>
-                <OptionInput value="distilbart">distilbart-cnn-6-6 (Free/Local)</OptionInput>
-                <OptionInput value="subModel">{language.submodel}</OptionInput>
-            </SelectInput>
-            <span class="text-textcolor">{language.summarizationPrompt} <Help key="summarizationPrompt"/></span>
-            <div class="mb-2">
-                <TextAreaInput size="sm" placeholder={language.hypaV3Settings.supaMemoryPromptPlaceHolder} bind:value={DBState.db.supaMemoryPrompt} />
-            </div>
-            {#await getMaxMemoryRatio() then maxMemoryRatio}
-            <span class="text-textcolor">{language.hypaV3Settings.maxMemoryTokensRatioLabel}</span>
-            <NumberInput marginBottom disabled size="sm" value={maxMemoryRatio} />
-            {:catch error}
-            <span class="text-red-400">{language.hypaV3Settings.maxMemoryTokensRatioError}</span>
-            {/await}
-            <span class="text-textcolor">{language.hypaV3Settings.memoryTokensRatioLabel}</span>
-            <SliderInput marginBottom min={0} max={1} step={0.01} fixed={2} bind:value={DBState.db.hypaV3Settings.memoryTokensRatio} />
-            <span class="text-textcolor">{language.hypaV3Settings.extraSummarizationRatioLabel}</span>
-            <SliderInput marginBottom min={0} max={1 - DBState.db.hypaV3Settings.memoryTokensRatio} step={0.01} fixed={2} bind:value={DBState.db.hypaV3Settings.extraSummarizationRatio} />
-            <span class="text-textcolor">{language.hypaV3Settings.maxChatsPerSummaryLabel}</span>
-            <NumberInput marginBottom size="sm" min={1} bind:value={DBState.db.hypaV3Settings.maxChatsPerSummary} />
-            <span class="text-textcolor">{language.hypaV3Settings.recentMemoryRatioLabel}</span>
-            <SliderInput marginBottom min={0} max={1} step={0.01} fixed={2} bind:value={DBState.db.hypaV3Settings.recentMemoryRatio} />
-            <span class="text-textcolor">{language.hypaV3Settings.similarMemoryRatioLabel}</span>
-            <SliderInput marginBottom min={0} max={1} step={0.01} fixed={2} bind:value={DBState.db.hypaV3Settings.similarMemoryRatio} />
-            <span class="text-textcolor">{language.hypaV3Settings.randomMemoryRatioLabel}</span>
-            <NumberInput marginBottom disabled size="sm" value={parseFloat((1 - DBState.db.hypaV3Settings.recentMemoryRatio - DBState.db.hypaV3Settings.similarMemoryRatio).toFixed(2))} />
-            <div class="flex mb-2">
-                <Check name={language.hypaV3Settings.enableSimilarityCorrectionLabel} bind:check={DBState.db.hypaV3Settings.enableSimilarityCorrection} />
-            </div>
-            <div class="flex mb-2">
-                <Check name={language.hypaV3Settings.preserveOrphanedMemoryLabel} bind:check={DBState.db.hypaV3Settings.preserveOrphanedMemory} />
-            </div>
-            <div class="flex mb-2">
-                <Check name={language.hypaV3Settings.applyRegexScriptWhenRerollingLabel} bind:check={DBState.db.hypaV3Settings.processRegexScript} />
-            </div>
-            <div class="flex mb-2">
-                <Check name={language.hypaV3Settings.doNotSummarizeUserMessageLabel} bind:check={DBState.db.hypaV3Settings.doNotSummarizeUserMessage} />
+            <span class="max-w-full mb-6 text-sm text-wrap break-words text-textcolor2">{language.hypaV3Settings.descriptionLabel}</span>
+            <span class="text-textcolor">Preset</span>
+            <select class={"border border-darkborderc focus:border-borderc rounded-md shadow-sm text-textcolor bg-transparent focus:ring-borderc focus:ring-2 focus:outline-none transition-colors duration-200 text-md px-4 py-2 mb-1"}
+                bind:value={DBState.db.hypaV3PresetId}
+            >
+                {#each DBState.db.hypaV3Presets as preset, i}
+                    <option class="bg-darkbg appearance-none" value={i}>{preset.name}</option>
+                {/each}
+            </select>
+
+            <div class="flex items-center mb-8">
+                <button class="mr-2 text-textcolor2 hover:text-green-500 cursor-pointer" onclick={() => {
+                    const newPreset = createHypaV3Preset()
+                    const presets = DBState.db.hypaV3Presets
+
+                    presets.push(newPreset)
+                    DBState.db.hypaV3Presets = presets
+                    DBState.db.hypaV3PresetId = DBState.db.hypaV3Presets.length - 1
+                }}>
+                    <PlusIcon size={24}/>
+                </button>
+
+                <button class="mr-2 text-textcolor2 hover:text-green-500 cursor-pointer" onclick={async () => {
+                    const presets = DBState.db.hypaV3Presets
+
+                    if(presets.length === 0){
+                        alertError("There must be least one preset.")
+                        return
+                    }
+
+                    const id = DBState.db.hypaV3PresetId
+                    const preset = presets[id]
+                    const newName = await alertInput(`Enter new name for ${preset.name}`)
+
+                    if (!newName || newName.trim().length === 0) return
+
+                    preset.name = newName
+                    DBState.db.hypaV3Presets = presets
+                }}>
+                    <PencilIcon size={24}/>
+                </button>
+
+                <button class="mr-2 text-textcolor2 hover:text-green-500 cursor-pointer" onclick={async (e) => {
+                    const presets = DBState.db.hypaV3Presets
+
+                    if(presets.length <= 1){
+                        alertError("There must be least one preset.")
+                        return
+                    }
+
+                    const id = DBState.db.hypaV3PresetId
+                    const preset = presets[id]
+                    const confirmed = await alertConfirm(`${language.removeConfirm}${preset.name}`)
+
+                    if (!confirmed) return
+
+                    DBState.db.hypaV3PresetId = 0
+                    presets.splice(id, 1)
+                    DBState.db.hypaV3Presets = presets
+                }}>
+                    <TrashIcon size={24}/>
+                </button>
+
+                <div class="ml-2 mr-4 w-px h-full bg-darkborderc"></div>
+
+                <button class="mr-2 text-textcolor2 hover:text-green-500 cursor-pointer" onclick={async() => {
+                    try {
+                        const presets = DBState.db.hypaV3Presets
+                        
+                        if(presets.length === 0){
+                            alertError("There must be least one preset.")
+                            return
+                        }
+
+                        const id = DBState.db.hypaV3PresetId
+                        const preset = presets[id]
+                        const bytesExport = Buffer.from(JSON.stringify({
+                            type: 'risu',
+                            ver: 1,
+                            data: preset
+                        }), 'utf-8')
+                        
+                        await downloadFile(`hypaV3_export_${preset.name}.json`, bytesExport)
+                        alertNormal(language.successExport)
+                    } catch (error) {
+                        alertError(`${error}`)
+                    }
+                }}>
+                    <DownloadIcon size={24}/>
+                </button>
+
+                <button class="mr-2 text-textcolor2 hover:text-green-500 cursor-pointer" onclick={async() => {
+                    try {
+                        const bytesImport = (await selectSingleFile(['json'])).data
+
+                        if(!bytesImport) return
+
+                        const objImport = JSON.parse(Buffer.from(bytesImport).toString('utf-8'))
+
+                        if(objImport.type !== 'risu' || !objImport.data) return
+
+                        const newPreset = createHypaV3Preset(
+                            objImport.data.name || "Imported Preset",
+                            objImport.data.settings || {}
+                        );
+                        const presets = DBState.db.hypaV3Presets
+                        
+                        presets.push(newPreset)
+                        DBState.db.hypaV3Presets = presets
+                        DBState.db.hypaV3PresetId = DBState.db.hypaV3Presets.length - 1
+
+                        alertNormal(language.successImport)
+                    } catch (error) {
+                        alertError(`${error}`)
+                    }
+                }}>
+                    <FolderUpIcon size={24}/>
+                </button>
             </div>
+
+            {#if DBState.db.hypaV3Presets?.[DBState.db.hypaV3PresetId]?.settings}
+                {@const settings = DBState.db.hypaV3Presets[DBState.db.hypaV3PresetId].settings}
+
+                <span class="text-textcolor">{language.SuperMemory} {language.model}</span>
+                <SelectInput className="mb-4" bind:value={settings.summarizationModel}>
+                    <OptionInput value="subModel">{language.submodel}</OptionInput>
+                    {#if "gpu" in navigator}
+                        <OptionInput value="Qwen3-1.7B-q4f32_1-MLC">Qwen3 1.7B (GPU)</OptionInput>
+                        <OptionInput value="Qwen3-4B-q4f32_1-MLC">Qwen3 4B (GPU)</OptionInput>
+                        <OptionInput value="Qwen3-8B-q4f32_1-MLC">Qwen3 8B (GPU)</OptionInput>
+                    {/if}
+                </SelectInput>
+                <span class="text-textcolor">{language.summarizationPrompt} <Help key="summarizationPrompt"/></span>
+                <div class="mb-4">
+                    <TextAreaInput size="sm" placeholder={language.hypaV3Settings.supaMemoryPromptPlaceHolder} bind:value={settings.summarizationPrompt} />
+                </div>
+                {#await getMaxMemoryRatio() then maxMemoryRatio}
+                <span class="text-textcolor">{language.hypaV3Settings.maxMemoryTokensRatioLabel}</span>
+                <NumberInput marginBottom disabled size="sm" value={maxMemoryRatio} />
+                {:catch error}
+                <span class="mb-4 text-red-400">{language.hypaV3Settings.maxMemoryTokensRatioError}</span>
+                {/await}
+                <span class="text-textcolor">{language.hypaV3Settings.memoryTokensRatioLabel}</span>
+                <SliderInput marginBottom min={0} max={1} step={0.01} fixed={2} bind:value={settings.memoryTokensRatio} />
+                <span class="text-textcolor">{language.hypaV3Settings.extraSummarizationRatioLabel}</span>
+                <SliderInput marginBottom min={0} max={1 - settings.memoryTokensRatio} step={0.01} fixed={2} bind:value={settings.extraSummarizationRatio} />
+                <span class="text-textcolor">{language.hypaV3Settings.maxChatsPerSummaryLabel}</span>
+                <NumberInput marginBottom size="sm" min={1} bind:value={settings.maxChatsPerSummary} />
+                <span class="text-textcolor">{language.hypaV3Settings.recentMemoryRatioLabel}</span>
+                <SliderInput marginBottom min={0} max={1} step={0.01} fixed={2} bind:value={settings.recentMemoryRatio} />
+                <span class="text-textcolor">{language.hypaV3Settings.similarMemoryRatioLabel}</span>
+                <SliderInput marginBottom min={0} max={1} step={0.01} fixed={2} bind:value={settings.similarMemoryRatio} />
+                <span class="text-textcolor">{language.hypaV3Settings.randomMemoryRatioLabel}</span>
+                <NumberInput marginBottom disabled size="sm" value={parseFloat((1 - settings.recentMemoryRatio - settings.similarMemoryRatio).toFixed(2))} />
+                <div class="mb-2">
+                    <Check name={language.hypaV3Settings.preserveOrphanedMemoryLabel} bind:check={settings.preserveOrphanedMemory} />
+                </div>
+                <div class="mb-2">
+                    <Check name={language.hypaV3Settings.applyRegexScriptWhenRerollingLabel} bind:check={settings.processRegexScript} />
+                </div>
+                <div class="mb-2">
+                    <Check name={language.hypaV3Settings.doNotSummarizeUserMessageLabel} bind:check={settings.doNotSummarizeUserMessage} />
+                </div>
+                <Arcodion name="Advanced Settings" styled>
+                    <div class="mb-2">
+                        <Check name="Use Experimental Implementation" bind:check={settings.useExperimentalImpl} />
+                    </div>
+                    {#if settings.useExperimentalImpl}
+                        <span class="text-textcolor">Summarization Requests Per Minute</span>
+                        <NumberInput marginBottom size="sm" min={1} bind:value={settings.summarizationRequestsPerMinute} />
+                        <span class="text-textcolor">Summarization Max Concurrent</span>
+                        <NumberInput marginBottom size="sm" min={1} max={10} bind:value={settings.summarizationMaxConcurrent} />
+                        <span class="text-textcolor">Embedding Requests Per Minute</span>
+                        <NumberInput marginBottom size="sm" min={1} bind:value={settings.embeddingRequestsPerMinute} />
+                        <span class="text-textcolor">Embedding Max Concurrent</span>
+                        <NumberInput marginBottom size="sm" min={1} max={10} bind:value={settings.embeddingMaxConcurrent} />
+                    {:else}
+                        <div class="mb-2">
+                            <Check name={language.hypaV3Settings.enableSimilarityCorrectionLabel} bind:check={settings.enableSimilarityCorrection} />
+                        </div>
+                    {/if}
+                </Arcodion>
+            {/if}
+
+            <div class="mb-8"></div>
         {:else if (DBState.db.supaModelType !== 'none' && DBState.db.hypav2 === false && DBState.db.hypaV3 === false)}
             <span class="mb-2 text-textcolor2 text-sm text-wrap break-words max-w-full">{language.supaDesc}</span>
             <span class="text-textcolor mt-4">{language.SuperMemory} {language.model}</span>
@@ -663,14 +809,17 @@
         <span class="text-textcolor">{language.embedding}</span>
         <SelectInput className="mt-2 mb-2" bind:value={DBState.db.hypaModel}>
             {#if 'gpu' in navigator}
+                <OptionInput value="MiniLMGPU">MiniLM L6 v2 (GPU)</OptionInput>
                 <OptionInput value="nomicGPU">Nomic Embed Text v1.5 (GPU)</OptionInput>
                 <OptionInput value="bgeSmallEnGPU">BGE Small English (GPU)</OptionInput>
                 <OptionInput value="bgem3GPU">BGE Medium 3 (GPU)</OptionInput>
+                <OptionInput value="multiMiniLMGPU">Multilingual MiniLM L12 v2 (GPU)</OptionInput>
             {/if}
             <OptionInput value="MiniLM">MiniLM L6 v2 (CPU)</OptionInput>
             <OptionInput value="nomic">Nomic Embed Text v1.5 (CPU)</OptionInput>
             <OptionInput value="bgeSmallEn">BGE Small English (CPU)</OptionInput>
             <OptionInput value="bgem3">BGE Medium 3 (CPU)</OptionInput>
+            <OptionInput value="multiMiniLM">Multilingual MiniLM L12 v2 (CPU)</OptionInput>
             <OptionInput value="openai3small">OpenAI text-embedding-3-small</OptionInput>
             <OptionInput value="openai3large">OpenAI text-embedding-3-large</OptionInput>
             <OptionInput value="ada">OpenAI Ada</OptionInput>
diff --git a/src/lib/SideBars/CharConfig.svelte b/src/lib/SideBars/CharConfig.svelte
index c4c90a6e..3fa58c3f 100644
--- a/src/lib/SideBars/CharConfig.svelte
+++ b/src/lib/SideBars/CharConfig.svelte
@@ -3,12 +3,12 @@
     import { tokenizeAccurate } from "../../ts/tokenizer";
     import { saveImage as saveAsset, type Database, type character, type groupChat } from "../../ts/storage/database.svelte";
     import { DBState } from 'src/ts/stores.svelte';
-    import { CharConfigSubMenu, MobileGUI, ShowRealmFrameStore, selectedCharID } from "../../ts/stores.svelte";
+    import { CharConfigSubMenu, MobileGUI, ShowRealmFrameStore, selectedCharID, hypaV3ModalOpen } from "../../ts/stores.svelte";
     import { PlusIcon, SmileIcon, TrashIcon, UserIcon, ActivityIcon, BookIcon, User, CurlyBraces, Volume2Icon, DownloadIcon, FolderUpIcon, Share2Icon } from 'lucide-svelte'
     import Check from "../UI/GUI/CheckInput.svelte";
     import { addCharEmotion, addingEmotion, getCharImage, rmCharEmotion, selectCharImg, makeGroupImage, removeChar, changeCharImage } from "../../ts/characters";
     import LoreBook from "./LoreBook/LoreBookSetting.svelte";
-    import { alertConfirm, alertMd, alertNormal, alertSelectChar, alertTOS, showHypaV2Alert, showHypaV3Alert } from "../../ts/alert";
+    import { alertConfirm, alertMd, alertNormal, alertSelectChar, alertTOS, showHypaV2Alert } from "../../ts/alert";
     import BarIcon from "./BarIcon.svelte";
     import { findCharacterbyId, getAuthorNoteDefaultText, parseKeyValue, selectMultipleFile, selectSingleFile } from "../../ts/util";
     import { onDestroy } from "svelte";
@@ -1103,10 +1103,10 @@
             >
                 {language.hypaMemoryV2Modal}
             </Button>
-        {:else if DBState.db.supaModelType !== 'none' && DBState.db.hypaV3}
+        {:else if DBState.db.hypaV3}
             <Button
                 onclick={() => {
-                    showHypaV3Alert()
+                    $hypaV3ModalOpen = true
                 }}
                 className="mt-4"
             >
diff --git a/src/lib/SideBars/Toggles.svelte b/src/lib/SideBars/Toggles.svelte
index c980a6af..afefa2bd 100644
--- a/src/lib/SideBars/Toggles.svelte
+++ b/src/lib/SideBars/Toggles.svelte
@@ -53,9 +53,9 @@
             <CheckInput bind:check={DBState.db.jailbreakToggle} name={language.jailbreakToggle} reverse />
         </div>
         {@render toggles(true)}
-        {#if DBState.db.supaModelType !== 'none' || DBState.db.hanuraiEnable}
+        {#if DBState.db.supaModelType !== 'none' || DBState.db.hanuraiEnable || DBState.db.hypaV3}
             <div class="flex mt-2 items-center w-full" class:justify-end={$MobileGUI}>
-                <CheckInput bind:check={chara.supaMemory} reverse name={DBState.db.hanuraiEnable ? language.hanuraiMemory : DBState.db.hypaMemory ? language.ToggleHypaMemory : language.ToggleSuperMemory}/>
+                <CheckInput bind:check={chara.supaMemory} reverse name={DBState.db.hypaV3 ? language.ToggleHypaMemory : DBState.db.hanuraiEnable ? language.hanuraiMemory : DBState.db.hypaMemory ? language.ToggleHypaMemory : language.ToggleSuperMemory}/>
             </div>
         {/if}
     </div>
@@ -64,9 +64,9 @@
         <CheckInput bind:check={DBState.db.jailbreakToggle} name={language.jailbreakToggle}/>
     </div>
     {@render toggles()}
-    {#if DBState.db.supaModelType !== 'none' || DBState.db.hanuraiEnable}
+    {#if DBState.db.supaModelType !== 'none' || DBState.db.hanuraiEnable || DBState.db.hypaV3}
         <div class="flex mt-2 items-center">
-            <CheckInput bind:check={chara.supaMemory} name={DBState.db.hanuraiEnable ? language.hanuraiMemory : DBState.db.hypaMemory ? language.ToggleHypaMemory : language.ToggleSuperMemory}/>
+            <CheckInput bind:check={chara.supaMemory} name={DBState.db.hypaV3 ? language.ToggleHypaMemory : DBState.db.hanuraiEnable ? language.hanuraiMemory : DBState.db.hypaMemory ? language.ToggleHypaMemory : language.ToggleSuperMemory}/>
         </div>
     {/if}
 {/if}
\ No newline at end of file
diff --git a/src/ts/alert.ts b/src/ts/alert.ts
index c7c96a88..746f6797 100644
--- a/src/ts/alert.ts
+++ b/src/ts/alert.ts
@@ -10,7 +10,7 @@ export interface alertData{
     type: 'error'|'normal'|'none'|'ask'|'wait'|'selectChar'
             |'input'|'toast'|'wait2'|'markdown'|'select'|'login'
             |'tos'|'cardexport'|'requestdata'|'addchar'|'hypaV2'|'selectModule'
-            |'chatOptions'|'pukmakkurit'|'branches'|'hypaV3'|'progress',
+            |'chatOptions'|'pukmakkurit'|'branches'|'progress',
     msg: string,
     submsg?: string
 }
@@ -319,10 +319,3 @@ export function showHypaV2Alert(){
         'msg': ""
     })
 }
-
-export function showHypaV3Alert(){
-    alertStoreImported.set({
-        'type': 'hypaV3',
-        'msg': ""
-    })
-}
\ No newline at end of file
diff --git a/src/ts/process/memory/hypamemory.ts b/src/ts/process/memory/hypamemory.ts
index 07b808fb..07af73d4 100644
--- a/src/ts/process/memory/hypamemory.ts
+++ b/src/ts/process/memory/hypamemory.ts
@@ -1,27 +1,33 @@
 import localforage from "localforage";
 import { globalFetch } from "src/ts/globalApi.svelte";
 import { runEmbedding } from "../transformers";
-import { alertError } from "src/ts/alert";
 import { appendLastPath } from "src/ts/util";
 import { getDatabase } from "src/ts/storage/database.svelte";
 
+export type HypaModel = 'custom'|'ada'|'openai3small'|'openai3large'|'MiniLM'|'MiniLMGPU'|'nomic'|'nomicGPU'|'bgeSmallEn'|'bgeSmallEnGPU'|'bgem3'|'bgem3GPU'|'multiMiniLM'|'multiMiniLMGPU'
 
-export type HypaModel = 'ada'|'MiniLM'|'nomic'|'custom'|'nomicGPU'|'bgeSmallEn'|'bgeSmallEnGPU'|'bgem3'|'bgem3GPU'|'openai3small'|'openai3large'
-
-const localModels = {
+// In a typical environment, bge-m3 is a heavy model.
+// If your GPU can't handle this model, you'll see errror below.
+// Failed to execute 'mapAsync' on 'GPUBuffer': [Device] is lost
+export const localModels = {
     models: {
         'MiniLM':'Xenova/all-MiniLM-L6-v2',
+        'MiniLMGPU': "Xenova/all-MiniLM-L6-v2",
         'nomic':'nomic-ai/nomic-embed-text-v1.5',
         'nomicGPU':'nomic-ai/nomic-embed-text-v1.5',
-        'bgeSmallEn': 'BAAI/bge-small-en-v1.5',
-        'bgeSmallEnGPU': 'BAAI/bge-small-en-v1.5',
-        'bgem3': 'BAAI/bge-m3',
-        'bgem3GPU': 'BAAI/bge-m3',
+        'bgeSmallEn': 'Xenova/bge-small-en-v1.5',
+        'bgeSmallEnGPU': 'Xenova/bge-small-en-v1.5',
+        'bgem3': 'Xenova/bge-m3',
+        'bgem3GPU': 'Xenova/bge-m3',
+        'multiMiniLM': 'Xenova/paraphrase-multilingual-MiniLM-L12-v2',
+        'multiMiniLMGPU': 'Xenova/paraphrase-multilingual-MiniLM-L12-v2',
     },
     gpuModels:[
+        'MiniLMGPU',
         'nomicGPU',
         'bgeSmallEnGPU',
-        'bgem3GPU'
+        'bgem3GPU',
+        'multiMiniLMGPU',
     ]
 }
 
@@ -44,7 +50,7 @@ export class HypaProcesser{
         else{
             this.model = model
         }
-        this.customEmbeddingUrl = customEmbeddingUrl || db.hypaCustomSettings.url
+        this.customEmbeddingUrl = customEmbeddingUrl?.trim() || db.hypaCustomSettings?.url?.trim() || ""
     }
 
     async embedDocuments(texts: string[]): Promise<VectorArray[]> {
@@ -80,10 +86,12 @@ export class HypaProcesser{
 
             const db = getDatabase()
             const fetchArgs = {
-                ...(db.hypaCustomSettings.key ? {headers: {"Authorization": "Bearer " + db.hypaCustomSettings.key}} : {}),
+                headers: {
+                    ...(db.hypaCustomSettings?.key?.trim() ? {"Authorization": "Bearer " + db.hypaCustomSettings.key.trim()} : {})
+                },
                 body: {
                     "input": input,
-                    ...(db.hypaCustomSettings.model ? {"model": db.hypaCustomSettings.model} : {})
+                    ...(db.hypaCustomSettings?.model?.trim() ? {"model": db.hypaCustomSettings.model.trim()} : {})
                 }
             };
  
@@ -99,7 +107,7 @@ export class HypaProcesser{
 
             gf = await globalFetch("https://api.openai.com/v1/embeddings", {
                 headers: {
-                    "Authorization": "Bearer " + db.supaMemoryKey || this.oaikey
+                    "Authorization": "Bearer " + (this.oaikey?.trim() || db.supaMemoryKey?.trim())
                 },
                 body: {
                     "input": input,
@@ -134,7 +142,7 @@ export class HypaProcesser{
     
     async addText(texts:string[]) {
         const db = getDatabase()
-        const suffix = (this.model === 'custom' && db.hypaCustomSettings.model) ? `-${db.hypaCustomSettings.model}` : ""
+        const suffix = (this.model === 'custom' && db.hypaCustomSettings?.model?.trim()) ? `-${db.hypaCustomSettings.model.trim()}` : ""
 
         for(let i=0;i<texts.length;i++){
             const itm:memoryVector = await this.forage.getItem(texts[i] + '|' + this.model + suffix)
@@ -205,7 +213,8 @@ export class HypaProcesser{
         return similarity(query1, query2)
     }
 }
-function similarity(a:VectorArray, b:VectorArray) {    
+
+export function similarity(a:VectorArray, b:VectorArray) {    
     let dot = 0;
     for(let i=0;i<a.length;i++){
         dot += a[i] * b[i]
@@ -227,4 +236,4 @@ const chunkArray = <T>(arr: T[], chunkSize: number) =>
         const chunk = chunks[chunkIndex] || [];
         chunks[chunkIndex] = chunk.concat([elem]);
         return chunks;
-}, [] as T[][]);
\ No newline at end of file
+}, [] as T[][]);
diff --git a/src/ts/process/memory/hypamemoryv2.ts b/src/ts/process/memory/hypamemoryv2.ts
new file mode 100644
index 00000000..14e1a08a
--- /dev/null
+++ b/src/ts/process/memory/hypamemoryv2.ts
@@ -0,0 +1,414 @@
+import localforage from "localforage";
+import { type HypaModel, localModels } from "./hypamemory";
+import { TaskRateLimiter, TaskCanceledError } from "./taskRateLimiter";
+import { runEmbedding } from "../transformers";
+import { globalFetch } from "src/ts/globalApi.svelte";
+import { getDatabase } from "src/ts/storage/database.svelte";
+import { appendLastPath } from "src/ts/util";
+
+export interface HypaProcessorV2Options {
+  model?: HypaModel;
+  customEmbeddingUrl?: string;
+  oaiKey?: string;
+  rateLimiter?: TaskRateLimiter;
+}
+
+export interface EmbeddingText<TMetadata> {
+  content: string;
+  metadata?: TMetadata;
+}
+
+export interface EmbeddingResult<TMetadata> extends EmbeddingText<TMetadata> {
+  embedding: EmbeddingVector;
+}
+
+export type EmbeddingVector = number[] | Float32Array;
+
+export class HypaProcessorV2<TMetadata> {
+  private static readonly LOG_PREFIX = "[HypaProcessorV2]";
+  public readonly options: HypaProcessorV2Options;
+  public progressCallback: (queuedCount: number) => void = null;
+  private vectors: Map<string, EmbeddingResult<TMetadata>> = new Map();
+  private forage: LocalForage = localforage.createInstance({
+    name: "hypaVector",
+  });
+
+  public constructor(options?: HypaProcessorV2Options) {
+    const db = getDatabase();
+
+    this.options = {
+      model: db.hypaModel || "MiniLM",
+      customEmbeddingUrl: db.hypaCustomSettings?.url?.trim() || "",
+      oaiKey: db.supaMemoryKey?.trim() || "",
+      rateLimiter: new TaskRateLimiter(),
+      ...options,
+    };
+  }
+
+  public async addTexts(ebdTexts: EmbeddingText<TMetadata>[]): Promise<void> {
+    await this.getEmbeds(ebdTexts, true);
+  }
+
+  public async similaritySearchScored(
+    query: string
+  ): Promise<[EmbeddingResult<TMetadata>, number][]> {
+    const results = await this.similaritySearchScoredBatch([query]);
+    return results[0];
+  }
+
+  public async similaritySearchScoredBatch(
+    queries: string[]
+  ): Promise<[EmbeddingResult<TMetadata>, number][][]> {
+    if (queries.length === 0) {
+      return [];
+    }
+
+    // Remove duplicate queries
+    const uniqueQueries = [...new Set(queries)];
+
+    // Convert queries to EmbeddingText array
+    const ebdTexts: EmbeddingText<TMetadata>[] = uniqueQueries.map((query) => ({
+      content: query,
+    }));
+
+    // Get query embeddings (don't save to memory)
+    const ebdResults = await this.getEmbeds(ebdTexts, false);
+
+    const scoredResultsMap = new Map<
+      string,
+      [EmbeddingResult<TMetadata>, number][]
+    >();
+
+    // Calculate similarity for each unique query
+    for (let i = 0; i < uniqueQueries.length; i++) {
+      const ebdResult = ebdResults[i];
+
+      const scoredVectors = Array.from(this.vectors.values())
+        .map((vector): [EmbeddingResult<TMetadata>, number] => [
+          vector,
+          this.similarity(ebdResult.embedding, vector.embedding),
+        ])
+        .sort((a, b) => b[1] - a[1]);
+
+      scoredResultsMap.set(uniqueQueries[i], scoredVectors);
+    }
+
+    return queries.map((query) => scoredResultsMap.get(query));
+  }
+
+  private async getEmbeds(
+    ebdTexts: EmbeddingText<TMetadata>[],
+    saveToMemory: boolean = true
+  ): Promise<EmbeddingResult<TMetadata>[]> {
+    if (ebdTexts.length === 0) {
+      return [];
+    }
+
+    const resultMap: Map<string, EmbeddingResult<TMetadata>> = new Map();
+    const toEmbed: EmbeddingText<TMetadata>[] = [];
+
+    // Load cache
+    const loadPromises = ebdTexts.map(async (item, index) => {
+      const { content, metadata } = item;
+
+      // Use if already in memory
+      if (this.vectors.has(content)) {
+        resultMap.set(content, this.vectors.get(content));
+        return;
+      }
+
+      try {
+        const cached = await this.forage.getItem<EmbeddingResult<TMetadata>>(
+          this.getCacheKey(content)
+        );
+
+        if (cached) {
+          // Debug log for cache hit
+          console.debug(
+            HypaProcessorV2.LOG_PREFIX,
+            `Cache hit for getting embedding ${index} with model ${this.options.model}`
+          );
+
+          // Add metadata
+          cached.metadata = metadata;
+
+          // Save to memory
+          if (saveToMemory) {
+            this.vectors.set(content, cached);
+          }
+
+          resultMap.set(content, cached);
+        } else {
+          toEmbed.push(item);
+        }
+      } catch (error) {
+        toEmbed.push(item);
+      }
+    });
+
+    await Promise.all(loadPromises);
+
+    if (toEmbed.length === 0) {
+      return ebdTexts.map((item) => resultMap.get(item.content));
+    }
+
+    // Chunking array
+    const chunkSize = await this.getOptimalChunkSize();
+
+    // Debug log for optimal chunk size
+    console.debug(
+      HypaProcessorV2.LOG_PREFIX,
+      `Optimal chunk size for ${this.options.model}: ${chunkSize}`
+    );
+
+    const chunks = this.chunkArray(toEmbed, chunkSize);
+
+    if (this.isLocalModel()) {
+      // Local model: Sequential processing
+      for (let i = 0; i < chunks.length; i++) {
+        // Progress callback
+        this.progressCallback?.(chunks.length - i - 1);
+
+        const chunk = chunks[i];
+        const embeddings = await this.getLocalEmbeds(
+          chunk.map((item) => item.content)
+        );
+
+        const savePromises = embeddings.map(async (embedding, j) => {
+          const { content, metadata } = chunk[j];
+
+          const ebdResult: EmbeddingResult<TMetadata> = {
+            content,
+            embedding,
+            metadata,
+          };
+
+          // Save to DB
+          await this.forage.setItem(this.getCacheKey(content), {
+            content,
+            embedding,
+          });
+
+          // Save to memory
+          if (saveToMemory) {
+            this.vectors.set(content, ebdResult);
+          }
+
+          resultMap.set(content, ebdResult);
+        });
+
+        await Promise.all(savePromises);
+      }
+    } else {
+      // API model: Parallel processing
+      const embeddingTasks = chunks.map((chunk) => {
+        const contents = chunk.map((item) => item.content);
+
+        return () => this.getAPIEmbeds(contents);
+      });
+
+      // Progress callback
+      this.options.rateLimiter.taskQueueChangeCallback = this.progressCallback;
+
+      const batchResult = await this.options.rateLimiter.executeBatch<
+        EmbeddingVector[]
+      >(embeddingTasks);
+      const errors: Error[] = [];
+
+      const chunksSavePromises = batchResult.results.map(async (result, i) => {
+        if (!result.success) {
+          errors.push(result.error);
+          return;
+        }
+
+        if (!result.data) {
+          errors.push(new Error("No embeddings found in the response."));
+          return;
+        }
+
+        const chunk = chunks[i];
+        const savePromises = result.data.map(async (embedding, j) => {
+          const { content, metadata } = chunk[j];
+
+          const ebdResult: EmbeddingResult<TMetadata> = {
+            content,
+            embedding,
+            metadata,
+          };
+
+          // Save to DB
+          await this.forage.setItem(this.getCacheKey(content), {
+            content,
+            embedding,
+          });
+
+          // Save to memory
+          if (saveToMemory) {
+            this.vectors.set(content, ebdResult);
+          }
+
+          resultMap.set(content, ebdResult);
+        });
+
+        await Promise.all(savePromises);
+      });
+
+      await Promise.all(chunksSavePromises);
+
+      // Throw major error if there are errors
+      if (errors.length > 0) {
+        const majorError =
+          errors.find((error) => !(error instanceof TaskCanceledError)) ||
+          errors[0];
+
+        throw majorError;
+      }
+    }
+
+    return ebdTexts.map((item) => resultMap.get(item.content));
+  }
+
+  private similarity(a: EmbeddingVector, b: EmbeddingVector): number {
+    let dot = 0;
+    let magA = 0;
+    let magB = 0;
+
+    for (let i = 0; i < a.length; i++) {
+      dot += a[i] * b[i];
+      magA += a[i] * a[i];
+      magB += b[i] * b[i];
+    }
+
+    return dot / (Math.sqrt(magA) * Math.sqrt(magB));
+  }
+
+  private getCacheKey(content: string): string {
+    const db = getDatabase();
+    const suffix =
+      this.options.model === "custom" && db.hypaCustomSettings?.model?.trim()
+        ? `-${db.hypaCustomSettings.model.trim()}`
+        : "";
+
+    return `${content}|${this.options.model}${suffix}`;
+  }
+
+  private async getOptimalChunkSize(): Promise<number> {
+    // API
+    if (!this.isLocalModel()) {
+      return 50;
+    }
+
+    const isMobile = /Android|iPhone|iPad|iPod|webOS/i.test(
+      navigator.userAgent
+    );
+
+    // WebGPU
+    if ("gpu" in navigator) {
+      return isMobile ? 5 : 10;
+    }
+
+    // WASM
+    const cpuCores = navigator.hardwareConcurrency || 4;
+    const baseChunkSize = isMobile ? Math.floor(cpuCores / 2) : cpuCores;
+
+    return Math.min(baseChunkSize, 10);
+  }
+
+  private isLocalModel(): boolean {
+    return Object.keys(localModels.models).includes(this.options.model);
+  }
+
+  private chunkArray<T>(array: T[], size: number): T[][] {
+    const chunks: T[][] = [];
+
+    for (let i = 0; i < array.length; i += size) {
+      chunks.push(array.slice(i, i + size));
+    }
+
+    return chunks;
+  }
+
+  private async getLocalEmbeds(contents: string[]): Promise<EmbeddingVector[]> {
+    const results: Float32Array[] = await runEmbedding(
+      contents,
+      localModels.models[this.options.model],
+      localModels.gpuModels.includes(this.options.model) ? "webgpu" : "wasm"
+    );
+
+    return results;
+  }
+
+  private async getAPIEmbeds(contents: string[]): Promise<EmbeddingVector[]> {
+    const db = getDatabase();
+    let response = null;
+
+    if (this.options.model === "custom") {
+      if (!this.options.customEmbeddingUrl) {
+        throw new Error("Custom model requires a Custom Server URL");
+      }
+
+      const replaceUrl = this.options.customEmbeddingUrl.endsWith("/embeddings")
+        ? this.options.customEmbeddingUrl
+        : appendLastPath(this.options.customEmbeddingUrl, "embeddings");
+
+      const fetchArgs = {
+        headers: {
+          ...(db.hypaCustomSettings?.key?.trim()
+            ? { Authorization: "Bearer " + db.hypaCustomSettings.key.trim() }
+            : {}),
+        },
+        body: {
+          input: contents,
+          ...(db.hypaCustomSettings?.model?.trim()
+            ? { model: db.hypaCustomSettings.model.trim() }
+            : {}),
+        },
+      };
+
+      response = await globalFetch(replaceUrl, fetchArgs);
+    } else if (
+      ["ada", "openai3small", "openai3large"].includes(this.options.model)
+    ) {
+      const models = {
+        ada: "text-embedding-ada-002",
+        openai3small: "text-embedding-3-small",
+        openai3large: "text-embedding-3-large",
+      };
+
+      const fetchArgs = {
+        headers: {
+          Authorization:
+            "Bearer " +
+            (this.options.oaiKey?.trim() || db.supaMemoryKey?.trim()),
+        },
+        body: {
+          input: contents,
+          model: models[this.options.model],
+        },
+      };
+
+      response = await globalFetch(
+        "https://api.openai.com/v1/embeddings",
+        fetchArgs
+      );
+    } else {
+      throw new Error(`Unsupported model: ${this.options.model}`);
+    }
+
+    if (!response.ok || !response.data.data) {
+      throw new Error(JSON.stringify(response.data));
+    }
+
+    const embeddings: EmbeddingVector[] = response.data.data.map(
+      (item: { embedding: EmbeddingVector }) => {
+        if (!item.embedding) {
+          throw new Error("No embeddings found in the response.");
+        }
+
+        return item.embedding;
+      }
+    );
+
+    return embeddings;
+  }
+}
diff --git a/src/ts/process/memory/hypav3.ts b/src/ts/process/memory/hypav3.ts
index f1e9f8a1..7985cf53 100644
--- a/src/ts/process/memory/hypav3.ts
+++ b/src/ts/process/memory/hypav3.ts
@@ -1,8 +1,6 @@
-import {
-  type VectorArray,
-  type memoryVector,
-  HypaProcesser,
-} from "./hypamemory";
+import { type memoryVector, HypaProcesser, similarity } from "./hypamemory";
+import { TaskRateLimiter } from "./taskRateLimiter";
+import { type EmbeddingText, HypaProcessorV2 } from "./hypamemoryv2";
 import {
   type Chat,
   type character,
@@ -11,15 +9,34 @@ import {
 } from "src/ts/storage/database.svelte";
 import { type OpenAIChat } from "../index.svelte";
 import { requestChatData } from "../request";
-import { runSummarizer } from "../transformers";
-import { globalFetch } from "src/ts/globalApi.svelte";
+import { chatCompletion, unloadEngine } from "../webllm";
 import { parseChatML } from "src/ts/parser.svelte";
+import { hypaV3ProgressStore } from "src/ts/stores.svelte";
 import { type ChatTokenizer } from "src/ts/tokenizer";
 
-interface Summary {
-  text: string;
-  chatMemos: Set<string>;
-  isImportant: boolean;
+export interface HypaV3Preset {
+  name: string;
+  settings: HypaV3Settings;
+}
+
+export interface HypaV3Settings {
+  summarizationModel: string;
+  summarizationPrompt: string;
+  memoryTokensRatio: number;
+  extraSummarizationRatio: number;
+  maxChatsPerSummary: number;
+  recentMemoryRatio: number;
+  similarMemoryRatio: number;
+  enableSimilarityCorrection: boolean;
+  preserveOrphanedMemory: boolean;
+  processRegexScript: boolean;
+  doNotSummarizeUserMessage: boolean;
+  // Experimental
+  useExperimentalImpl: boolean;
+  summarizationRequestsPerMinute: number;
+  summarizationMaxConcurrent: number;
+  embeddingRequestsPerMinute: number;
+  embeddingMaxConcurrent: number;
 }
 
 interface HypaV3Data {
@@ -36,228 +53,29 @@ export interface SerializableHypaV3Data {
   lastSelectedSummaries?: number[];
 }
 
+interface Summary {
+  text: string;
+  chatMemos: Set<string>;
+  isImportant: boolean;
+}
+
 interface SummaryChunk {
   text: string;
   summary: Summary;
 }
 
+export interface HypaV3Result {
+  currentTokens: number;
+  chats: OpenAIChat[];
+  error?: string;
+  memory?: SerializableHypaV3Data;
+}
+
+const logPrefix = "[HypaV3]";
+const memoryPromptTag = "Past Events Summary";
 const minChatsForSimilarity = 3;
-const maxSummarizationFailures = 3;
 const summarySeparator = "\n\n";
 
-// Helper function to check if one Set is a subset of another
-function isSubset(subset: Set<string>, superset: Set<string>): boolean {
-  for (const elem of subset) {
-    if (!superset.has(elem)) {
-      return false;
-    }
-  }
-  return true;
-}
-
-function toSerializableHypaV3Data(data: HypaV3Data): SerializableHypaV3Data {
-  return {
-    ...data,
-    summaries: data.summaries.map((summary) => ({
-      ...summary,
-      chatMemos: [...summary.chatMemos],
-    })),
-  };
-}
-
-function toHypaV3Data(serialData: SerializableHypaV3Data): HypaV3Data {
-  return {
-    ...serialData,
-    summaries: serialData.summaries.map((summary) => ({
-      ...summary,
-      // Convert null back to undefined (JSON serialization converts undefined to null)
-      chatMemos: new Set(
-        summary.chatMemos.map((memo) => (memo === null ? undefined : memo))
-      ),
-    })),
-  };
-}
-
-function encapsulateMemoryPrompt(memoryPrompt: string): string {
-  return `<Past Events Summary>${memoryPrompt}</Past Events Summary>`;
-}
-
-function cleanOrphanedSummary(chats: OpenAIChat[], data: HypaV3Data): void {
-  // Collect all memos from current chats
-  const currentChatMemos = new Set(chats.map((chat) => chat.memo));
-  const originalLength = data.summaries.length;
-
-  // Filter summaries - keep only those whose chatMemos are subset of current chat memos
-  data.summaries = data.summaries.filter((summary) => {
-    return isSubset(summary.chatMemos, currentChatMemos);
-  });
-
-  const removedCount = originalLength - data.summaries.length;
-
-  if (removedCount > 0) {
-    console.log(`[HypaV3] Cleaned ${removedCount} orphaned summaries.`);
-  }
-}
-
-export async function summarize(
-  oaiChats: OpenAIChat[]
-): Promise<{ success: boolean; data: string }> {
-  const db = getDatabase();
-  const stringifiedChats = oaiChats
-    .map((chat) => `${chat.role}: ${chat.content}`)
-    .join("\n");
-
-  if (db.supaModelType === "distilbart") {
-    try {
-      const summaryText = (await runSummarizer(stringifiedChats)).trim();
-      return { success: true, data: summaryText };
-    } catch (error) {
-      return {
-        success: false,
-        data: error,
-      };
-    }
-  }
-
-  const summarizePrompt =
-    db.supaMemoryPrompt === ""
-      ? "[Summarize the ongoing role story, It must also remove redundancy and unnecessary text and content from the output.]"
-      : db.supaMemoryPrompt;
-
-  switch (db.supaModelType) {
-    case "instruct35": {
-      console.log(
-        "[HypaV3] Using openAI gpt-3.5-turbo-instruct for summarization."
-      );
-
-      const requestPrompt = `${stringifiedChats}\n\n${summarizePrompt}\n\nOutput:`;
-      const response = await globalFetch(
-        "https://api.openai.com/v1/completions",
-        {
-          method: "POST",
-          headers: {
-            "Content-Type": "application/json",
-            Authorization: "Bearer " + db.supaMemoryKey,
-          },
-          body: {
-            model: "gpt-3.5-turbo-instruct",
-            prompt: requestPrompt,
-            max_tokens: db.maxResponse,
-            temperature: 0,
-          },
-        }
-      );
-
-      try {
-        if (!response.ok) {
-          return {
-            success: false,
-            data: JSON.stringify(response),
-          };
-        }
-
-        const summaryText =
-          response.data?.choices?.[0]?.message?.content?.trim();
-
-        if (!summaryText) {
-          return {
-            success: false,
-            data: JSON.stringify(response),
-          };
-        }
-
-        return { success: true, data: summaryText };
-      } catch (error) {
-        return {
-          success: false,
-          data: error,
-        };
-      }
-    }
-
-    case "subModel": {
-      console.log(`[HypaV3] Using ax model ${db.subModel} for summarization.`);
-
-      const requestMessages: OpenAIChat[] = parseChatML(
-        summarizePrompt.replaceAll("{{slot}}", stringifiedChats)
-      ) ?? [
-        {
-          role: "user",
-          content: stringifiedChats,
-        },
-        {
-          role: "system",
-          content: summarizePrompt,
-        },
-      ];
-
-      const response = await requestChatData(
-        {
-          formated: requestMessages,
-          bias: {},
-          useStreaming: false,
-          noMultiGen: true,
-        },
-        "memory"
-      );
-
-      if (response.type === "streaming" || response.type === "multiline") {
-        return {
-          success: false,
-          data: "unexpected response type",
-        };
-      }
-
-      if (response.type === "fail") {
-        return {
-          success: false,
-          data: response.result,
-        };
-      }
-
-      return { success: true, data: response.result.trim() };
-    }
-
-    default: {
-      return {
-        success: false,
-        data: `unsupported model ${db.supaModelType} for summarization`,
-      };
-    }
-  }
-}
-
-async function retryableSummarize(
-  oaiChats: OpenAIChat[]
-): Promise<{ success: boolean; data: string }> {
-  let summarizationFailures = 0;
-
-  while (summarizationFailures < maxSummarizationFailures) {
-    console.log(
-      "[HypaV3] Attempting summarization:",
-      "\nAttempt:",
-      summarizationFailures + 1,
-      "\nTarget:",
-      oaiChats
-    );
-
-    const summarizeResult = await summarize(oaiChats);
-
-    if (!summarizeResult.success) {
-      console.log("[HypaV3] Summarization failed:", summarizeResult.data);
-      summarizationFailures++;
-
-      if (summarizationFailures >= maxSummarizationFailures) {
-        return summarizeResult;
-      }
-
-      continue;
-    }
-
-    return summarizeResult;
-  }
-}
-
 export async function hypaMemoryV3(
   chats: OpenAIChat[],
   currentTokens: number,
@@ -265,24 +83,74 @@ export async function hypaMemoryV3(
   room: Chat,
   char: character | groupChat,
   tokenizer: ChatTokenizer
-): Promise<{
-  currentTokens: number;
-  chats: OpenAIChat[];
-  error?: string;
-  memory?: SerializableHypaV3Data;
-}> {
+): Promise<HypaV3Result> {
+  const settings = getCurrentHypaV3Preset().settings;
+
+  try {
+    if (settings.useExperimentalImpl) {
+      console.log(logPrefix, "Using experimental implementation.");
+
+      return await hypaMemoryV3MainExp(
+        chats,
+        currentTokens,
+        maxContextTokens,
+        room,
+        char,
+        tokenizer
+      );
+    }
+
+    return await hypaMemoryV3Main(
+      chats,
+      currentTokens,
+      maxContextTokens,
+      room,
+      char,
+      tokenizer
+    );
+  } catch (error) {
+    if (error instanceof Error) {
+      // Standard Error instance
+      error.message = `${logPrefix} ${error.message}`;
+      throw error;
+    }
+
+    // Fallback for non-Error object
+    let errorMessage: string;
+
+    try {
+      errorMessage = JSON.stringify(error);
+    } catch {
+      errorMessage = String(error);
+    }
+
+    throw new Error(`${logPrefix} ${errorMessage}`);
+  } finally {
+    if (settings.summarizationModel !== "subModel") {
+      try {
+        unloadEngine();
+      } catch {}
+    }
+  }
+}
+
+async function hypaMemoryV3MainExp(
+  chats: OpenAIChat[],
+  currentTokens: number,
+  maxContextTokens: number,
+  room: Chat,
+  char: character | groupChat,
+  tokenizer: ChatTokenizer
+): Promise<HypaV3Result> {
   const db = getDatabase();
+  const settings = getCurrentHypaV3Preset().settings;
 
   // Validate settings
-  if (
-    db.hypaV3Settings.recentMemoryRatio + db.hypaV3Settings.similarMemoryRatio >
-    1
-  ) {
+  if (settings.recentMemoryRatio + settings.similarMemoryRatio > 1) {
     return {
       currentTokens,
       chats,
-      error:
-        "[HypaV3] The sum of Recent Memory Ratio and Similar Memory Ratio is greater than 1.",
+      error: `${logPrefix} The sum of Recent Memory Ratio and Similar Memory Ratio is greater than 1.`,
     };
   }
 
@@ -300,7 +168,7 @@ export async function hypaMemoryV3(
   }
 
   // Clean orphaned summaries
-  if (!db.hypaV3Settings.preserveOrphanedMemory) {
+  if (!settings.preserveOrphanedMemory) {
     cleanOrphanedSummary(chats, data);
   }
 
@@ -324,33 +192,32 @@ export async function hypaMemoryV3(
     }
   }
 
+  console.log(logPrefix, "Starting index:", startIdx);
+
   // Reserve memory tokens
   const emptyMemoryTokens = await tokenizer.tokenizeChat({
     role: "system",
-    content: encapsulateMemoryPrompt(""),
+    content: wrapWithXml(memoryPromptTag, ""),
   });
   const memoryTokens = Math.floor(
-    maxContextTokens * db.hypaV3Settings.memoryTokensRatio
+    maxContextTokens * settings.memoryTokensRatio
   );
-  const shouldReserveEmptyMemoryTokens =
-    data.summaries.length === 0 &&
-    currentTokens + emptyMemoryTokens <= maxContextTokens;
-  let availableMemoryTokens = shouldReserveEmptyMemoryTokens
-    ? 0
-    : memoryTokens - emptyMemoryTokens;
+  const shouldReserveMemoryTokens =
+    data.summaries.length > 0 || currentTokens > maxContextTokens;
+  let availableMemoryTokens = shouldReserveMemoryTokens
+    ? memoryTokens - emptyMemoryTokens
+    : 0;
 
-  if (shouldReserveEmptyMemoryTokens) {
-    currentTokens += emptyMemoryTokens;
-    console.log("[HypaV3] Reserved empty memory tokens:", emptyMemoryTokens);
-  } else {
+  if (shouldReserveMemoryTokens) {
     currentTokens += memoryTokens;
-    console.log("[HypaV3] Reserved max memory tokens:", memoryTokens);
+    console.log(logPrefix, "Reserved memory tokens:", memoryTokens);
   }
 
   // If summarization is needed
-  let summarizationMode = currentTokens > maxContextTokens;
+  const summarizationMode = currentTokens > maxContextTokens;
   const targetTokens =
-    maxContextTokens * (1 - db.hypaV3Settings.extraSummarizationRatio);
+    maxContextTokens * (1 - settings.extraSummarizationRatio);
+  const toSummarizeArray: OpenAIChat[][] = [];
 
   while (summarizationMode) {
     if (currentTokens <= targetTokens) {
@@ -364,43 +231,41 @@ export async function hypaMemoryV3(
         return {
           currentTokens,
           chats,
-          error: `[HypaV3] Cannot summarize further: input token count (${currentTokens}) exceeds max context size (${maxContextTokens}), but minimum ${minChatsForSimilarity} messages required.`,
+          error: `${logPrefix} Cannot summarize further: input token count (${currentTokens}) exceeds max context size (${maxContextTokens}), but minimum ${minChatsForSimilarity} messages required.`,
           memory: toSerializableHypaV3Data(data),
         };
       }
     }
 
     const toSummarize: OpenAIChat[] = [];
-    const endIdx = Math.min(
-      startIdx + db.hypaV3Settings.maxChatsPerSummary,
-      chats.length - minChatsForSimilarity
-    );
     let toSummarizeTokens = 0;
+    let currentIndex = startIdx;
 
     console.log(
-      "[HypaV3] Evaluating summarization batch:",
+      logPrefix,
+      "Evaluating summarization batch:",
       "\nCurrent Tokens:",
       currentTokens,
       "\nMax Context Tokens:",
       maxContextTokens,
       "\nStart Index:",
       startIdx,
-      "\nEnd Index:",
-      endIdx,
-      "\nChat Count:",
-      endIdx - startIdx,
       "\nMax Chats Per Summary:",
-      db.hypaV3Settings.maxChatsPerSummary
+      settings.maxChatsPerSummary
     );
 
-    for (let i = startIdx; i < endIdx; i++) {
-      const chat = chats[i];
+    while (
+      toSummarize.length < settings.maxChatsPerSummary &&
+      currentIndex < chats.length - minChatsForSimilarity
+    ) {
+      const chat = chats[currentIndex];
       const chatTokens = await tokenizer.tokenizeChat(chat);
 
       console.log(
-        "[HypaV3] Evaluating chat:",
+        logPrefix,
+        "Evaluating chat:",
         "\nIndex:",
-        i,
+        currentIndex,
         "\nRole:",
         chat.role,
         "\nContent:",
@@ -411,23 +276,40 @@ export async function hypaMemoryV3(
 
       toSummarizeTokens += chatTokens;
 
-      if (i === 0 || !chat.content.trim()) {
+      let shouldSummarize = true;
+
+      if (
+        chat.name === "example_user" ||
+        chat.name === "example_assistant" ||
+        chat.memo === "NewChatExample"
+      ) {
         console.log(
-          `[HypaV3] Skipping ${
-            i === 0 ? "[Start a new chat]" : "empty content"
-          } at index ${i}`
+          logPrefix,
+          `Skipping example chat at index ${currentIndex}`
         );
-
-        continue;
+        shouldSummarize = false;
       }
 
-      if (db.hypaV3Settings.doNotSummarizeUserMessage && chat.role === "user") {
-        console.log(`[HypaV3] Skipping user role at index ${i}`);
-
-        continue;
+      if (chat.memo === "NewChat") {
+        console.log(logPrefix, `Skipping new chat at index ${currentIndex}`);
+        shouldSummarize = false;
       }
 
-      toSummarize.push(chat);
+      if (chat.content.trim().length === 0) {
+        console.log(logPrefix, `Skipping empty chat at index ${currentIndex}`);
+        shouldSummarize = false;
+      }
+
+      if (settings.doNotSummarizeUserMessage && chat.role === "user") {
+        console.log(logPrefix, `Skipping user role at index ${currentIndex}`);
+        shouldSummarize = false;
+      }
+
+      if (shouldSummarize) {
+        toSummarize.push(chat);
+      }
+
+      currentIndex++;
     }
 
     // Stop summarization if further reduction would go below target tokens (unless we're over max tokens)
@@ -436,43 +318,119 @@ export async function hypaMemoryV3(
       currentTokens - toSummarizeTokens < targetTokens
     ) {
       console.log(
-        `[HypaV3] Stopping summarization: currentTokens(${currentTokens}) - toSummarizeTokens(${toSummarizeTokens}) < targetTokens(${targetTokens})`
+        logPrefix,
+        "Stopping summarization:",
+        `\ncurrentTokens(${currentTokens}) - toSummarizeTokens(${toSummarizeTokens}) < targetTokens(${targetTokens})`
       );
       break;
     }
 
-    // Attempt summarization
+    // Collect summarization batch
     if (toSummarize.length > 0) {
-      const summarizeResult = await retryableSummarize(toSummarize);
+      console.log(
+        logPrefix,
+        "Collecting summarization batch:",
+        "\nTarget:",
+        toSummarize
+      );
+
+      toSummarizeArray.push([...toSummarize]);
+    }
+
+    currentTokens -= toSummarizeTokens;
+    startIdx = currentIndex;
+  }
+
+  // Process all collected summarization tasks
+  if (toSummarizeArray.length > 0) {
+    // Initialize rate limiter
+    // Local model must be processed sequentially
+    const rateLimiter = new TaskRateLimiter({
+      tasksPerMinute:
+        settings.summarizationModel === "subModel"
+          ? settings.summarizationRequestsPerMinute
+          : 1000,
+      maxConcurrentTasks:
+        settings.summarizationModel === "subModel"
+          ? settings.summarizationMaxConcurrent
+          : 1,
+    });
+
+    rateLimiter.taskQueueChangeCallback = (queuedCount) => {
+      hypaV3ProgressStore.set({
+        open: true,
+        miniMsg: `${rateLimiter.queuedTaskCount}`,
+        msg: `${logPrefix} Summarizing...`,
+        subMsg: `${rateLimiter.queuedTaskCount} queued`,
+      });
+    };
+
+    const summarizationTasks = toSummarizeArray.map(
+      (item) => () => summarize(item)
+    );
+
+    // Start of performance measurement: summarize
+    console.log(
+      logPrefix,
+      `Starting ${toSummarizeArray.length} summarization.`
+    );
+    const summarizeStartTime = performance.now();
+
+    const batchResult = await rateLimiter.executeBatch<string>(
+      summarizationTasks
+    );
+
+    const summarizeEndTime = performance.now();
+    console.debug(
+      `${logPrefix} summarization completed in ${
+        summarizeEndTime - summarizeStartTime
+      }ms`
+    );
+    // End of performance measurement: summarize
+
+    hypaV3ProgressStore.set({
+      open: false,
+      miniMsg: "",
+      msg: "",
+      subMsg: "",
+    });
+
+    // Note:
+    // We can't save some successful summaries to the DB temporarily
+    // because don't know the actual summarization model name.
+    // It is possible that the user can change the summarization model.
+    for (let i = 0; i < batchResult.results.length; i++) {
+      const result = batchResult.results[i];
+
+      // Push consecutive successes
+      if (!result.success || !result.data) {
+        const errorMessage = !result.success
+          ? result.error
+          : "Empty summary returned";
+
+        console.log(logPrefix, "Summarization failed:", `\n${errorMessage}`);
 
-      if (
-        !summarizeResult.success ||
-        !summarizeResult.data ||
-        summarizeResult.data.trim().length === 0
-      ) {
         return {
           currentTokens,
           chats,
-          error: `[HypaV3] Summarization failed after maximum retries: ${summarizeResult.data}`,
+          error: `${logPrefix} Summarization failed: ${errorMessage}`,
           memory: toSerializableHypaV3Data(data),
         };
       }
 
+      const summaryText = result.data;
+
       data.summaries.push({
-        text: summarizeResult.data,
-        chatMemos: new Set(toSummarize.map((chat) => chat.memo)),
+        text: summaryText,
+        chatMemos: new Set(toSummarizeArray[i].map((chat) => chat.memo)),
         isImportant: false,
       });
     }
-
-    currentTokens -= toSummarizeTokens;
-    startIdx = endIdx;
   }
 
   console.log(
-    `[HypaV3] ${
-      summarizationMode ? "Completed" : "Skipped"
-    } summarization phase:`,
+    logPrefix,
+    `${summarizationMode ? "Completed" : "Skipped"} summarization phase:`,
     "\nCurrent Tokens:",
     currentTokens,
     "\nMax Context Tokens:",
@@ -483,20 +441,11 @@ export async function hypaMemoryV3(
 
   // Early return if no summaries
   if (data.summaries.length === 0) {
-    // Generate final memory prompt
-    const memory = encapsulateMemoryPrompt("");
-
-    const newChats: OpenAIChat[] = [
-      {
-        role: "system",
-        content: memory,
-        memo: "supaMemory",
-      },
-      ...chats.slice(startIdx),
-    ];
+    const newChats: OpenAIChat[] = chats.slice(startIdx);
 
     console.log(
-      "[HypaV3] Exiting function:",
+      logPrefix,
+      "Exiting function:",
       "\nCurrent Tokens:",
       currentTokens,
       "\nAll chats, including memory prompt:",
@@ -514,49 +463,50 @@ export async function hypaMemoryV3(
 
   const selectedSummaries: Summary[] = [];
   const randomMemoryRatio =
-    1 -
-    db.hypaV3Settings.recentMemoryRatio -
-    db.hypaV3Settings.similarMemoryRatio;
+    1 - settings.recentMemoryRatio - settings.similarMemoryRatio;
 
   // Select important summaries
-  const selectedImportantSummaries: Summary[] = [];
+  {
+    const selectedImportantSummaries: Summary[] = [];
 
-  for (const summary of data.summaries) {
-    if (summary.isImportant) {
-      const summaryTokens = await tokenizer.tokenizeChat({
-        role: "system",
-        content: summary.text + summarySeparator,
-      });
+    for (const summary of data.summaries) {
+      if (summary.isImportant) {
+        const summaryTokens = await tokenizer.tokenizeChat({
+          role: "system",
+          content: summary.text + summarySeparator,
+        });
 
-      if (summaryTokens > availableMemoryTokens) {
-        break;
+        if (summaryTokens > availableMemoryTokens) {
+          break;
+        }
+
+        selectedImportantSummaries.push(summary);
+
+        availableMemoryTokens -= summaryTokens;
       }
-
-      selectedImportantSummaries.push(summary);
-
-      availableMemoryTokens -= summaryTokens;
     }
+
+    selectedSummaries.push(...selectedImportantSummaries);
+
+    console.log(
+      logPrefix,
+      "After important memory selection:",
+      "\nSummary Count:",
+      selectedImportantSummaries.length,
+      "\nSummaries:",
+      selectedImportantSummaries,
+      "\nAvailable Memory Tokens:",
+      availableMemoryTokens
+    );
   }
 
-  selectedSummaries.push(...selectedImportantSummaries);
-
-  console.log(
-    "[HypaV3] After important memory selection:",
-    "\nSummary Count:",
-    selectedImportantSummaries.length,
-    "\nSummaries:",
-    selectedImportantSummaries,
-    "\nAvailable Memory Tokens:",
-    availableMemoryTokens
-  );
-
   // Select recent summaries
   const reservedRecentMemoryTokens = Math.floor(
-    availableMemoryTokens * db.hypaV3Settings.recentMemoryRatio
+    availableMemoryTokens * settings.recentMemoryRatio
   );
   let consumedRecentMemoryTokens = 0;
 
-  if (db.hypaV3Settings.recentMemoryRatio > 0) {
+  if (settings.recentMemoryRatio > 0) {
     const selectedRecentSummaries: Summary[] = [];
 
     // Target only summaries that haven't been selected yet
@@ -586,7 +536,8 @@ export async function hypaMemoryV3(
     selectedSummaries.push(...selectedRecentSummaries);
 
     console.log(
-      "[HypaV3] After recent memory selection:",
+      logPrefix,
+      "After recent memory selection:",
       "\nSummary Count:",
       selectedRecentSummaries.length,
       "\nSummaries:",
@@ -600,11 +551,11 @@ export async function hypaMemoryV3(
 
   // Select similar summaries
   let reservedSimilarMemoryTokens = Math.floor(
-    availableMemoryTokens * db.hypaV3Settings.similarMemoryRatio
+    availableMemoryTokens * settings.similarMemoryRatio
   );
   let consumedSimilarMemoryTokens = 0;
 
-  if (db.hypaV3Settings.similarMemoryRatio > 0) {
+  if (settings.similarMemoryRatio > 0) {
     const selectedSimilarSummaries: Summary[] = [];
 
     // Utilize unused token space from recent selection
@@ -614,7 +565,8 @@ export async function hypaMemoryV3(
 
       reservedSimilarMemoryTokens += unusedRecentTokens;
       console.log(
-        "[HypaV3] Additional available token space for similar memory:",
+        logPrefix,
+        "Additional available token space for similar memory:",
         "\nFrom recent:",
         unusedRecentTokens
       );
@@ -625,155 +577,182 @@ export async function hypaMemoryV3(
       (e) => !selectedSummaries.includes(e)
     );
 
-    // Dynamically generate summary chunks
-    const summaryChunks: SummaryChunk[] = [];
+    // Dynamically generate embedding texts
+    const ebdTexts: EmbeddingText<Summary>[] = unusedSummaries.flatMap(
+      (summary) => {
+        const splitted = summary.text
+          .split("\n\n")
+          .filter((e) => e.trim().length > 0);
 
-    unusedSummaries.forEach((summary) => {
-      const splitted = summary.text
-        .split("\n\n")
-        .filter((e) => e.trim().length > 0);
+        return splitted.map((e) => ({
+          content: e.trim(),
+          metadata: summary,
+        }));
+      }
+    );
 
-      summaryChunks.push(
-        ...splitted.map((e) => ({
-          text: e.trim(),
-          summary,
-        }))
-      );
+    // Initialize embedding processor
+    const processor = new HypaProcessorV2<Summary>({
+      rateLimiter: new TaskRateLimiter({
+        tasksPerMinute: settings.embeddingRequestsPerMinute,
+        maxConcurrentTasks: settings.embeddingMaxConcurrent,
+      }),
     });
 
-    // Fetch memory from summaryChunks
-    const processor = new HypaProcesserEx(db.hypaModel);
-    processor.oaikey = db.supaMemoryKey;
+    processor.progressCallback = (queuedCount) => {
+      hypaV3ProgressStore.set({
+        open: true,
+        miniMsg: `${queuedCount}`,
+        msg: `${logPrefix} Similarity searching...`,
+        subMsg: `${queuedCount} queued`,
+      });
+    };
 
-    // Add summaryChunks to processor for similarity search
     try {
-      await processor.addSummaryChunks(summaryChunks);
+      // Start of performance measurement: addTexts
+      console.log(
+        `${logPrefix} Starting addTexts with ${ebdTexts.length} chunks`
+      );
+      const addStartTime = performance.now();
+
+      // Add EmbeddingTexts to processor for similarity search
+      await processor.addTexts(ebdTexts);
+
+      const addEndTime = performance.now();
+      console.debug(
+        `${logPrefix} addTexts completed in ${addEndTime - addStartTime}ms`
+      );
+      // End of performance measurement: addTexts
     } catch (error) {
       return {
         currentTokens,
         chats,
-        error: `[HypaV3] Similarity search failed: ${error}`,
+        error: `${logPrefix} Similarity search failed: ${error}`,
         memory: toSerializableHypaV3Data(data),
       };
-    }
-
-    const scoredSummaries = new Map<Summary, number>();
-
-    // (1) Raw recent chat search
-    for (let i = 0; i < minChatsForSimilarity; i++) {
-      const pop = chats[chats.length - i - 1];
-
-      if (!pop) break;
-
-      try {
-        const searched = await processor.similaritySearchScoredEx(pop.content);
-
-        for (const [chunk, similarity] of searched) {
-          const summary = chunk.summary;
-
-          scoredSummaries.set(
-            summary,
-            (scoredSummaries.get(summary) || 0) + similarity
-          );
-        }
-      } catch (error) {
-        return {
-          currentTokens,
-          chats,
-          error: `[HypaV3] Similarity search failed: ${error}`,
-          memory: toSerializableHypaV3Data(data),
-        };
-      }
-    }
-
-    // (2) Summarized recent chat search
-    if (db.hypaV3Settings.enableSimilarityCorrection) {
-      // Attempt summarization
-      const recentChats = chats.slice(-minChatsForSimilarity);
-      const summarizeResult = await retryableSummarize(recentChats);
-
-      if (
-        !summarizeResult.success ||
-        !summarizeResult.data ||
-        summarizeResult.data.trim().length === 0
-      ) {
-        return {
-          currentTokens,
-          chats,
-          error: `[HypaV3] Summarization failed after maximum retries: ${summarizeResult.data}`,
-          memory: toSerializableHypaV3Data(data),
-        };
-      }
-
-      try {
-        const searched = await processor.similaritySearchScoredEx(
-          summarizeResult.data
-        );
-
-        for (const [chunk, similarity] of searched) {
-          const summary = chunk.summary;
-
-          scoredSummaries.set(
-            summary,
-            (scoredSummaries.get(summary) || 0) + similarity
-          );
-        }
-      } catch (error) {
-        return {
-          currentTokens,
-          chats,
-          error: `[HypaV3] Similarity search failed: ${error}`,
-          memory: toSerializableHypaV3Data(data),
-        };
-      }
-
-      console.log("[HypaV3] Similarity corrected.");
-    }
-
-    // Sort in descending order
-    const scoredArray = [...scoredSummaries.entries()].sort(
-      ([, scoreA], [, scoreB]) => scoreB - scoreA
-    );
-
-    while (scoredArray.length > 0) {
-      const [summary] = scoredArray.shift();
-      const summaryTokens = await tokenizer.tokenizeChat({
-        role: "system",
-        content: summary.text + summarySeparator,
+    } finally {
+      hypaV3ProgressStore.set({
+        open: false,
+        miniMsg: "",
+        msg: "",
+        subMsg: "",
       });
-
-      /*
-      console.log(
-        "[HypaV3] Trying to add similar summary:",
-        "\nSummary Tokens:",
-        summaryTokens,
-        "\nConsumed Similar Memory Tokens:",
-        consumedSimilarMemoryTokens,
-        "\nReserved Tokens:",
-        reservedSimilarMemoryTokens,
-        "\nWould exceed:",
-        summaryTokens + consumedSimilarMemoryTokens > reservedSimilarMemoryTokens
-      );
-      */
-
-      if (
-        summaryTokens + consumedSimilarMemoryTokens >
-        reservedSimilarMemoryTokens
-      ) {
-        console.log(
-          `[HypaV3] Stopping similar memory selection: consumedSimilarMemoryTokens(${consumedSimilarMemoryTokens}) + summaryTokens(${summaryTokens}) > reservedSimilarMemoryTokens(${reservedSimilarMemoryTokens})`
-        );
-        break;
-      }
-
-      selectedSimilarSummaries.push(summary);
-      consumedSimilarMemoryTokens += summaryTokens;
     }
 
-    selectedSummaries.push(...selectedSimilarSummaries);
+    const recentChats = chats
+      .slice(-minChatsForSimilarity)
+      .filter((chat) => chat.content.trim().length > 0);
+    const queries: string[] = recentChats.flatMap((chat) => {
+      return chat.content.split("\n\n").filter((e) => e.trim().length > 0);
+    });
+
+    if (queries.length > 0) {
+      const scoredSummaries = new Map<Summary, number>();
+
+      try {
+        // Start of performance measurement: similarity search
+        console.log(
+          `${logPrefix} Starting similarity search with ${recentChats.length} queries`
+        );
+        const searchStartTime = performance.now();
+
+        const batchScoredResults = await processor.similaritySearchScoredBatch(
+          queries
+        );
+
+        const searchEndTime = performance.now();
+        console.debug(
+          `${logPrefix} Similarity search completed in ${
+            searchEndTime - searchStartTime
+          }ms`
+        );
+        // End of performance measurement: similarity search
+
+        for (const scoredResults of batchScoredResults) {
+          for (const [ebdResult, similarity] of scoredResults) {
+            const summary = ebdResult.metadata;
+
+            scoredSummaries.set(
+              summary,
+              (scoredSummaries.get(summary) || 0) + similarity
+            );
+          }
+        }
+      } catch (error) {
+        return {
+          currentTokens,
+          chats,
+          error: `${logPrefix} Similarity search failed: ${error}`,
+          memory: toSerializableHypaV3Data(data),
+        };
+      } finally {
+        hypaV3ProgressStore.set({
+          open: false,
+          miniMsg: "",
+          msg: "",
+          subMsg: "",
+        });
+      }
+
+      // Normalize scores
+      if (scoredSummaries.size > 0) {
+        const maxScore = Math.max(...scoredSummaries.values());
+
+        for (const [summary, score] of scoredSummaries.entries()) {
+          scoredSummaries.set(summary, score / maxScore);
+        }
+      }
+
+      // Sort in descending order
+      const scoredArray = [...scoredSummaries.entries()].sort(
+        ([, scoreA], [, scoreB]) => scoreB - scoreA
+      );
+
+      while (scoredArray.length > 0) {
+        const [summary] = scoredArray.shift();
+        const summaryTokens = await tokenizer.tokenizeChat({
+          role: "system",
+          content: summary.text + summarySeparator,
+        });
+
+        /*
+        console.log(
+          logPrefix,
+          "Trying to add similar summary:",
+          "\nSummary Tokens:",
+          summaryTokens,
+          "\nConsumed Similar Memory Tokens:",
+          consumedSimilarMemoryTokens,
+          "\nReserved Tokens:",
+          reservedSimilarMemoryTokens,
+          "\nWould exceed:",
+          summaryTokens + consumedSimilarMemoryTokens >
+            reservedSimilarMemoryTokens
+        );
+        */
+
+        if (
+          summaryTokens + consumedSimilarMemoryTokens >
+          reservedSimilarMemoryTokens
+        ) {
+          console.log(
+            logPrefix,
+            "Stopping similar memory selection:",
+            `\nconsumedSimilarMemoryTokens(${consumedSimilarMemoryTokens}) + summaryTokens(${summaryTokens}) > reservedSimilarMemoryTokens(${reservedSimilarMemoryTokens})`
+          );
+          break;
+        }
+
+        selectedSimilarSummaries.push(summary);
+        consumedSimilarMemoryTokens += summaryTokens;
+      }
+
+      selectedSummaries.push(...selectedSimilarSummaries);
+    }
 
     console.log(
-      "[HypaV3] After similar memory selection:",
+      logPrefix,
+      "After similar memory selection:",
       "\nSummary Count:",
       selectedSimilarSummaries.length,
       "\nSummaries:",
@@ -802,7 +781,8 @@ export async function hypaMemoryV3(
 
     reservedRandomMemoryTokens += unusedRecentTokens + unusedSimilarTokens;
     console.log(
-      "[HypaV3] Additional available token space for random memory:",
+      logPrefix,
+      "Additional available token space for random memory:",
       "\nFrom recent:",
       unusedRecentTokens,
       "\nFrom similar:",
@@ -837,7 +817,8 @@ export async function hypaMemoryV3(
     selectedSummaries.push(...selectedRandomSummaries);
 
     console.log(
-      "[HypaV3] After random memory selection:",
+      logPrefix,
+      "After random memory selection:",
       "\nSummary Count:",
       selectedRandomSummaries.length,
       "\nSummaries:",
@@ -855,7 +836,8 @@ export async function hypaMemoryV3(
   );
 
   // Generate final memory prompt
-  const memory = encapsulateMemoryPrompt(
+  const memory = wrapWithXml(
+    memoryPromptTag,
     selectedSummaries.map((e) => e.text).join(summarySeparator)
   );
   const realMemoryTokens = await tokenizer.tokenizeChat({
@@ -864,16 +846,15 @@ export async function hypaMemoryV3(
   });
 
   // Release reserved memory tokens
-  if (shouldReserveEmptyMemoryTokens) {
-    currentTokens -= emptyMemoryTokens;
-  } else {
+  if (shouldReserveMemoryTokens) {
     currentTokens -= memoryTokens;
   }
 
   currentTokens += realMemoryTokens;
 
   console.log(
-    "[HypaV3] Final memory selection:",
+    logPrefix,
+    "Final memory selection:",
     "\nSummary Count:",
     selectedSummaries.length,
     "\nSummaries:",
@@ -886,7 +867,7 @@ export async function hypaMemoryV3(
 
   if (currentTokens > maxContextTokens) {
     throw new Error(
-      `[HypaV3] Unexpected error: input token count (${currentTokens}) exceeds max context size (${maxContextTokens})`
+      `Unexpected error: input token count (${currentTokens}) exceeds max context size (${maxContextTokens})`
     );
   }
 
@@ -905,7 +886,8 @@ export async function hypaMemoryV3(
   ];
 
   console.log(
-    "[HypaV3] Exiting function:",
+    logPrefix,
+    "Exiting function:",
     "\nCurrent Tokens:",
     currentTokens,
     "\nAll chats, including memory prompt:",
@@ -921,26 +903,870 @@ export async function hypaMemoryV3(
   };
 }
 
-type SummaryChunkVector = {
+async function hypaMemoryV3Main(
+  chats: OpenAIChat[],
+  currentTokens: number,
+  maxContextTokens: number,
+  room: Chat,
+  char: character | groupChat,
+  tokenizer: ChatTokenizer
+): Promise<HypaV3Result> {
+  const db = getDatabase();
+  const settings = getCurrentHypaV3Preset().settings;
+
+  // Validate settings
+  if (settings.recentMemoryRatio + settings.similarMemoryRatio > 1) {
+    return {
+      currentTokens,
+      chats,
+      error: `${logPrefix} The sum of Recent Memory Ratio and Similar Memory Ratio is greater than 1.`,
+    };
+  }
+
+  // Initial token correction
+  currentTokens -= db.maxResponse;
+
+  // Load existing hypa data if available
+  let data: HypaV3Data = {
+    summaries: [],
+    lastSelectedSummaries: [],
+  };
+
+  if (room.hypaV3Data) {
+    data = toHypaV3Data(room.hypaV3Data);
+  }
+
+  // Clean orphaned summaries
+  if (!settings.preserveOrphanedMemory) {
+    cleanOrphanedSummary(chats, data);
+  }
+
+  // Determine starting index
+  let startIdx = 0;
+
+  if (data.summaries.length > 0) {
+    const lastSummary = data.summaries.at(-1);
+    const lastChatIndex = chats.findIndex(
+      (chat) => chat.memo === [...lastSummary.chatMemos].at(-1)
+    );
+
+    if (lastChatIndex !== -1) {
+      startIdx = lastChatIndex + 1;
+
+      // Exclude tokens from summarized chats
+      const summarizedChats = chats.slice(0, lastChatIndex + 1);
+      for (const chat of summarizedChats) {
+        currentTokens -= await tokenizer.tokenizeChat(chat);
+      }
+    }
+  }
+
+  console.log(logPrefix, "Starting index:", startIdx);
+
+  // Reserve memory tokens
+  const emptyMemoryTokens = await tokenizer.tokenizeChat({
+    role: "system",
+    content: wrapWithXml(memoryPromptTag, ""),
+  });
+  const memoryTokens = Math.floor(
+    maxContextTokens * settings.memoryTokensRatio
+  );
+  const shouldReserveEmptyMemoryTokens =
+    data.summaries.length === 0 &&
+    currentTokens + emptyMemoryTokens <= maxContextTokens;
+  let availableMemoryTokens = shouldReserveEmptyMemoryTokens
+    ? 0
+    : memoryTokens - emptyMemoryTokens;
+
+  if (shouldReserveEmptyMemoryTokens) {
+    currentTokens += emptyMemoryTokens;
+    console.log(logPrefix, "Reserved empty memory tokens:", emptyMemoryTokens);
+  } else {
+    currentTokens += memoryTokens;
+    console.log(logPrefix, "Reserved max memory tokens:", memoryTokens);
+  }
+
+  // If summarization is needed
+  const summarizationMode = currentTokens > maxContextTokens;
+  const targetTokens =
+    maxContextTokens * (1 - settings.extraSummarizationRatio);
+
+  while (summarizationMode) {
+    if (currentTokens <= targetTokens) {
+      break;
+    }
+
+    if (chats.length - startIdx <= minChatsForSimilarity) {
+      if (currentTokens <= maxContextTokens) {
+        break;
+      } else {
+        return {
+          currentTokens,
+          chats,
+          error: `${logPrefix} Cannot summarize further: input token count (${currentTokens}) exceeds max context size (${maxContextTokens}), but minimum ${minChatsForSimilarity} messages required.`,
+          memory: toSerializableHypaV3Data(data),
+        };
+      }
+    }
+
+    const toSummarize: OpenAIChat[] = [];
+    const endIdx = Math.min(
+      startIdx + settings.maxChatsPerSummary,
+      chats.length - minChatsForSimilarity
+    );
+    let toSummarizeTokens = 0;
+
+    console.log(
+      logPrefix,
+      "Evaluating summarization batch:",
+      "\nCurrent Tokens:",
+      currentTokens,
+      "\nMax Context Tokens:",
+      maxContextTokens,
+      "\nStart Index:",
+      startIdx,
+      "\nEnd Index:",
+      endIdx,
+      "\nChat Count:",
+      endIdx - startIdx,
+      "\nMax Chats Per Summary:",
+      settings.maxChatsPerSummary
+    );
+
+    for (let i = startIdx; i < endIdx; i++) {
+      const chat = chats[i];
+      const chatTokens = await tokenizer.tokenizeChat(chat);
+
+      console.log(
+        logPrefix,
+        "Evaluating chat:",
+        "\nIndex:",
+        i,
+        "\nRole:",
+        chat.role,
+        "\nContent:",
+        "\n" + chat.content,
+        "\nTokens:",
+        chatTokens
+      );
+
+      toSummarizeTokens += chatTokens;
+
+      if (
+        chat.name === "example_user" ||
+        chat.name === "example_assistant" ||
+        chat.memo === "NewChatExample"
+      ) {
+        console.log(logPrefix, `Skipping example chat at index ${i}`);
+        continue;
+      }
+
+      if (chat.memo === "NewChat") {
+        console.log(logPrefix, `Skipping new chat at index ${i}`);
+        continue;
+      }
+
+      if (chat.content.trim().length === 0) {
+        console.log(logPrefix, `Skipping empty chat at index ${i}`);
+        continue;
+      }
+
+      if (settings.doNotSummarizeUserMessage && chat.role === "user") {
+        console.log(logPrefix, `Skipping user role at index ${i}`);
+        continue;
+      }
+
+      toSummarize.push(chat);
+    }
+
+    // Stop summarization if further reduction would go below target tokens (unless we're over max tokens)
+    if (
+      currentTokens <= maxContextTokens &&
+      currentTokens - toSummarizeTokens < targetTokens
+    ) {
+      console.log(
+        logPrefix,
+        "Stopping summarization:",
+        `\ncurrentTokens(${currentTokens}) - toSummarizeTokens(${toSummarizeTokens}) < targetTokens(${targetTokens})`
+      );
+      break;
+    }
+
+    // Attempt summarization
+    if (toSummarize.length > 0) {
+      console.log(
+        logPrefix,
+        "Attempting summarization:",
+        "\nTarget:",
+        toSummarize
+      );
+
+      try {
+        const summarizeResult = await summarize(toSummarize);
+
+        data.summaries.push({
+          text: summarizeResult,
+          chatMemos: new Set(toSummarize.map((chat) => chat.memo)),
+          isImportant: false,
+        });
+      } catch (error) {
+        console.log(logPrefix, "Summarization failed:", `\n${error}`);
+
+        return {
+          currentTokens,
+          chats,
+          error: `${logPrefix} Summarization failed: ${error}`,
+          memory: toSerializableHypaV3Data(data),
+        };
+      }
+    }
+
+    currentTokens -= toSummarizeTokens;
+    startIdx = endIdx;
+  }
+
+  console.log(
+    logPrefix,
+    `${summarizationMode ? "Completed" : "Skipped"} summarization phase:`,
+    "\nCurrent Tokens:",
+    currentTokens,
+    "\nMax Context Tokens:",
+    maxContextTokens,
+    "\nAvailable Memory Tokens:",
+    availableMemoryTokens
+  );
+
+  // Early return if no summaries
+  if (data.summaries.length === 0) {
+    // Generate final memory prompt
+    const memory = wrapWithXml(memoryPromptTag, "");
+
+    const newChats: OpenAIChat[] = [
+      {
+        role: "system",
+        content: memory,
+        memo: "supaMemory",
+      },
+      ...chats.slice(startIdx),
+    ];
+
+    console.log(
+      logPrefix,
+      "Exiting function:",
+      "\nCurrent Tokens:",
+      currentTokens,
+      "\nAll chats, including memory prompt:",
+      newChats,
+      "\nMemory Data:",
+      data
+    );
+
+    return {
+      currentTokens,
+      chats: newChats,
+      memory: toSerializableHypaV3Data(data),
+    };
+  }
+
+  const selectedSummaries: Summary[] = [];
+  const randomMemoryRatio =
+    1 - settings.recentMemoryRatio - settings.similarMemoryRatio;
+
+  // Select important summaries
+  {
+    const selectedImportantSummaries: Summary[] = [];
+
+    for (const summary of data.summaries) {
+      if (summary.isImportant) {
+        const summaryTokens = await tokenizer.tokenizeChat({
+          role: "system",
+          content: summary.text + summarySeparator,
+        });
+
+        if (summaryTokens > availableMemoryTokens) {
+          break;
+        }
+
+        selectedImportantSummaries.push(summary);
+
+        availableMemoryTokens -= summaryTokens;
+      }
+    }
+
+    selectedSummaries.push(...selectedImportantSummaries);
+
+    console.log(
+      logPrefix,
+      "After important memory selection:",
+      "\nSummary Count:",
+      selectedImportantSummaries.length,
+      "\nSummaries:",
+      selectedImportantSummaries,
+      "\nAvailable Memory Tokens:",
+      availableMemoryTokens
+    );
+  }
+
+  // Select recent summaries
+  const reservedRecentMemoryTokens = Math.floor(
+    availableMemoryTokens * settings.recentMemoryRatio
+  );
+  let consumedRecentMemoryTokens = 0;
+
+  if (settings.recentMemoryRatio > 0) {
+    const selectedRecentSummaries: Summary[] = [];
+
+    // Target only summaries that haven't been selected yet
+    const unusedSummaries = data.summaries.filter(
+      (e) => !selectedSummaries.includes(e)
+    );
+
+    // Add one by one from the end
+    for (let i = unusedSummaries.length - 1; i >= 0; i--) {
+      const summary = unusedSummaries[i];
+      const summaryTokens = await tokenizer.tokenizeChat({
+        role: "system",
+        content: summary.text + summarySeparator,
+      });
+
+      if (
+        summaryTokens + consumedRecentMemoryTokens >
+        reservedRecentMemoryTokens
+      ) {
+        break;
+      }
+
+      selectedRecentSummaries.push(summary);
+      consumedRecentMemoryTokens += summaryTokens;
+    }
+
+    selectedSummaries.push(...selectedRecentSummaries);
+
+    console.log(
+      logPrefix,
+      "After recent memory selection:",
+      "\nSummary Count:",
+      selectedRecentSummaries.length,
+      "\nSummaries:",
+      selectedRecentSummaries,
+      "\nReserved Tokens:",
+      reservedRecentMemoryTokens,
+      "\nConsumed Tokens:",
+      consumedRecentMemoryTokens
+    );
+  }
+
+  // Select similar summaries
+  let reservedSimilarMemoryTokens = Math.floor(
+    availableMemoryTokens * settings.similarMemoryRatio
+  );
+  let consumedSimilarMemoryTokens = 0;
+
+  if (settings.similarMemoryRatio > 0) {
+    const selectedSimilarSummaries: Summary[] = [];
+
+    // Utilize unused token space from recent selection
+    if (randomMemoryRatio <= 0) {
+      const unusedRecentTokens =
+        reservedRecentMemoryTokens - consumedRecentMemoryTokens;
+
+      reservedSimilarMemoryTokens += unusedRecentTokens;
+      console.log(
+        logPrefix,
+        "Additional available token space for similar memory:",
+        "\nFrom recent:",
+        unusedRecentTokens
+      );
+    }
+
+    // Target only summaries that haven't been selected yet
+    const unusedSummaries = data.summaries.filter(
+      (e) => !selectedSummaries.includes(e)
+    );
+
+    // Dynamically generate summary chunks
+    const summaryChunks: SummaryChunk[] = [];
+
+    unusedSummaries.forEach((summary) => {
+      const splitted = summary.text
+        .split("\n\n")
+        .filter((e) => e.trim().length > 0);
+
+      summaryChunks.push(
+        ...splitted.map((e) => ({
+          text: e.trim(),
+          summary,
+        }))
+      );
+    });
+
+    // Initialize embedding processor
+    const processor = new HypaProcesserEx(db.hypaModel);
+    processor.oaikey = db.supaMemoryKey;
+
+    // Add summaryChunks to processor for similarity search
+    try {
+      await processor.addSummaryChunks(summaryChunks);
+    } catch (error) {
+      return {
+        currentTokens,
+        chats,
+        error: `${logPrefix} Similarity search failed: ${error}`,
+        memory: toSerializableHypaV3Data(data),
+      };
+    }
+
+    const scoredSummaries = new Map<Summary, number>();
+    const recentChats = chats
+      .slice(-minChatsForSimilarity)
+      .filter((chat) => chat.content.trim().length > 0);
+
+    if (recentChats.length > 0) {
+      // Raw recent chat search
+      const queries = recentChats.map((chat) => chat.content);
+
+      if (settings.enableSimilarityCorrection && recentChats.length > 1) {
+        // Raw + Summarized recent chat search
+        // Summarizing is meaningful when there are more than 2 recent chats
+
+        // Attempt summarization
+        console.log(
+          logPrefix,
+          "Attempting summarization for similarity search:",
+          "\nTarget:",
+          recentChats
+        );
+
+        try {
+          const summarizeResult = await summarize(recentChats);
+
+          queries.push(summarizeResult);
+        } catch (error) {
+          console.log(logPrefix, "Summarization failed:", `\n${error}`);
+
+          return {
+            currentTokens,
+            chats,
+            error: `${logPrefix} Summarization failed: ${error}`,
+            memory: toSerializableHypaV3Data(data),
+          };
+        }
+      }
+
+      try {
+        for (const query of queries) {
+          const scoredChunks = await processor.similaritySearchScoredEx(query);
+
+          for (const [chunk, similarity] of scoredChunks) {
+            const summary = chunk.summary;
+
+            scoredSummaries.set(
+              summary,
+              (scoredSummaries.get(summary) || 0) + similarity
+            );
+          }
+        }
+      } catch (error) {
+        return {
+          currentTokens,
+          chats,
+          error: `${logPrefix} Similarity search failed: ${error}`,
+          memory: toSerializableHypaV3Data(data),
+        };
+      }
+    }
+
+    // Sort in descending order
+    const scoredArray = [...scoredSummaries.entries()].sort(
+      ([, scoreA], [, scoreB]) => scoreB - scoreA
+    );
+
+    while (scoredArray.length > 0) {
+      const [summary] = scoredArray.shift();
+      const summaryTokens = await tokenizer.tokenizeChat({
+        role: "system",
+        content: summary.text + summarySeparator,
+      });
+
+      /*
+      console.log(
+        logPrefix,
+        "Trying to add similar summary:",
+        "\nSummary Tokens:",
+        summaryTokens,
+        "\nConsumed Similar Memory Tokens:",
+        consumedSimilarMemoryTokens,
+        "\nReserved Tokens:",
+        reservedSimilarMemoryTokens,
+        "\nWould exceed:",
+        summaryTokens + consumedSimilarMemoryTokens >
+          reservedSimilarMemoryTokens
+      );
+      */
+
+      if (
+        summaryTokens + consumedSimilarMemoryTokens >
+        reservedSimilarMemoryTokens
+      ) {
+        console.log(
+          logPrefix,
+          "Stopping similar memory selection:",
+          `\nconsumedSimilarMemoryTokens(${consumedSimilarMemoryTokens}) + summaryTokens(${summaryTokens}) > reservedSimilarMemoryTokens(${reservedSimilarMemoryTokens})`
+        );
+        break;
+      }
+
+      selectedSimilarSummaries.push(summary);
+      consumedSimilarMemoryTokens += summaryTokens;
+    }
+
+    selectedSummaries.push(...selectedSimilarSummaries);
+
+    console.log(
+      logPrefix,
+      "After similar memory selection:",
+      "\nSummary Count:",
+      selectedSimilarSummaries.length,
+      "\nSummaries:",
+      selectedSimilarSummaries,
+      "\nReserved Tokens:",
+      reservedSimilarMemoryTokens,
+      "\nConsumed Tokens:",
+      consumedSimilarMemoryTokens
+    );
+  }
+
+  // Select random summaries
+  let reservedRandomMemoryTokens = Math.floor(
+    availableMemoryTokens * randomMemoryRatio
+  );
+  let consumedRandomMemoryTokens = 0;
+
+  if (randomMemoryRatio > 0) {
+    const selectedRandomSummaries: Summary[] = [];
+
+    // Utilize unused token space from recent and similar selection
+    const unusedRecentTokens =
+      reservedRecentMemoryTokens - consumedRecentMemoryTokens;
+    const unusedSimilarTokens =
+      reservedSimilarMemoryTokens - consumedSimilarMemoryTokens;
+
+    reservedRandomMemoryTokens += unusedRecentTokens + unusedSimilarTokens;
+    console.log(
+      logPrefix,
+      "Additional available token space for random memory:",
+      "\nFrom recent:",
+      unusedRecentTokens,
+      "\nFrom similar:",
+      unusedSimilarTokens,
+      "\nTotal added:",
+      unusedRecentTokens + unusedSimilarTokens
+    );
+
+    // Target only summaries that haven't been selected yet
+    const unusedSummaries = data.summaries
+      .filter((e) => !selectedSummaries.includes(e))
+      .sort(() => Math.random() - 0.5); // Random shuffle
+
+    for (const summary of unusedSummaries) {
+      const summaryTokens = await tokenizer.tokenizeChat({
+        role: "system",
+        content: summary.text + summarySeparator,
+      });
+
+      if (
+        summaryTokens + consumedRandomMemoryTokens >
+        reservedRandomMemoryTokens
+      ) {
+        // Trying to select more random memory
+        continue;
+      }
+
+      selectedRandomSummaries.push(summary);
+      consumedRandomMemoryTokens += summaryTokens;
+    }
+
+    selectedSummaries.push(...selectedRandomSummaries);
+
+    console.log(
+      logPrefix,
+      "After random memory selection:",
+      "\nSummary Count:",
+      selectedRandomSummaries.length,
+      "\nSummaries:",
+      selectedRandomSummaries,
+      "\nReserved Tokens:",
+      reservedRandomMemoryTokens,
+      "\nConsumed Tokens:",
+      consumedRandomMemoryTokens
+    );
+  }
+
+  // Sort selected summaries chronologically (by index)
+  selectedSummaries.sort(
+    (a, b) => data.summaries.indexOf(a) - data.summaries.indexOf(b)
+  );
+
+  // Generate final memory prompt
+  const memory = wrapWithXml(
+    memoryPromptTag,
+    selectedSummaries.map((e) => e.text).join(summarySeparator)
+  );
+  const realMemoryTokens = await tokenizer.tokenizeChat({
+    role: "system",
+    content: memory,
+  });
+
+  // Release reserved memory tokens
+  if (shouldReserveEmptyMemoryTokens) {
+    currentTokens -= emptyMemoryTokens;
+  } else {
+    currentTokens -= memoryTokens;
+  }
+
+  currentTokens += realMemoryTokens;
+
+  console.log(
+    logPrefix,
+    "Final memory selection:",
+    "\nSummary Count:",
+    selectedSummaries.length,
+    "\nSummaries:",
+    selectedSummaries,
+    "\nReal Memory Tokens:",
+    realMemoryTokens,
+    "\nAvailable Memory Tokens:",
+    availableMemoryTokens
+  );
+
+  if (currentTokens > maxContextTokens) {
+    throw new Error(
+      `Unexpected error: input token count (${currentTokens}) exceeds max context size (${maxContextTokens})`
+    );
+  }
+
+  // Save last selected summaries
+  data.lastSelectedSummaries = selectedSummaries.map((selectedSummary) =>
+    data.summaries.findIndex((summary) => summary === selectedSummary)
+  );
+
+  const newChats: OpenAIChat[] = [
+    {
+      role: "system",
+      content: memory,
+      memo: "supaMemory",
+    },
+    ...chats.slice(startIdx),
+  ];
+
+  console.log(
+    logPrefix,
+    "Exiting function:",
+    "\nCurrent Tokens:",
+    currentTokens,
+    "\nAll chats, including memory prompt:",
+    newChats,
+    "\nMemory Data:",
+    data
+  );
+
+  return {
+    currentTokens,
+    chats: newChats,
+    memory: toSerializableHypaV3Data(data),
+  };
+}
+
+function toHypaV3Data(serialData: SerializableHypaV3Data): HypaV3Data {
+  return {
+    ...serialData,
+    summaries: serialData.summaries.map((summary) => ({
+      ...summary,
+      // Convert null back to undefined (JSON serialization converts undefined to null)
+      chatMemos: new Set(
+        summary.chatMemos.map((memo) => (memo === null ? undefined : memo))
+      ),
+    })),
+  };
+}
+
+function toSerializableHypaV3Data(data: HypaV3Data): SerializableHypaV3Data {
+  return {
+    ...data,
+    summaries: data.summaries.map((summary) => ({
+      ...summary,
+      chatMemos: [...summary.chatMemos],
+    })),
+  };
+}
+
+function cleanOrphanedSummary(chats: OpenAIChat[], data: HypaV3Data): void {
+  // Collect all memos from current chats
+  const currentChatMemos = new Set(chats.map((chat) => chat.memo));
+  const originalLength = data.summaries.length;
+
+  // Filter summaries - keep only those whose chatMemos are subset of current chat memos
+  data.summaries = data.summaries.filter((summary) => {
+    return isSubset(summary.chatMemos, currentChatMemos);
+  });
+
+  const removedCount = originalLength - data.summaries.length;
+
+  if (removedCount > 0) {
+    console.log(logPrefix, `Cleaned ${removedCount} orphaned summaries.`);
+  }
+}
+
+function isSubset(subset: Set<string>, superset: Set<string>): boolean {
+  for (const elem of subset) {
+    if (!superset.has(elem)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+function wrapWithXml(tag: string, content: string): string {
+  return `<${tag}>\n${content}\n</${tag}>`;
+}
+
+export async function summarize(oaiMessages: OpenAIChat[]): Promise<string> {
+  const db = getDatabase();
+  const settings = getCurrentHypaV3Preset().settings;
+
+  const strMessages = oaiMessages
+    .map((chat) => `${chat.role}: ${chat.content}`)
+    .join("\n");
+
+  const summarizationPrompt =
+    settings.summarizationPrompt.trim() === ""
+      ? "[Summarize the ongoing role story, It must also remove redundancy and unnecessary text and content from the output.]"
+      : settings.summarizationPrompt;
+
+  const formated: OpenAIChat[] = parseChatML(
+    summarizationPrompt.replaceAll("{{slot}}", strMessages)
+  ) ?? [
+    {
+      role: "user",
+      content: strMessages,
+    },
+    {
+      role: "system",
+      content: summarizationPrompt,
+    },
+  ];
+
+  // API
+  if (settings.summarizationModel === "subModel") {
+    console.log(logPrefix, `Using ax model ${db.subModel} for summarization.`);
+
+    const response = await requestChatData(
+      {
+        formated,
+        bias: {},
+        useStreaming: false,
+        noMultiGen: true,
+      },
+      "memory"
+    );
+
+    if (response.type === "streaming" || response.type === "multiline") {
+      throw new Error("Unexpected response type");
+    }
+
+    if (response.type === "fail") {
+      throw new Error(response.result);
+    }
+
+    if (!response.result || response.result.trim().length === 0) {
+      throw new Error("Empty summary returned");
+    }
+
+    return response.result.trim();
+  }
+
+  // Local
+  const content = await chatCompletion(formated, settings.summarizationModel, {
+    max_tokens: 8192,
+    temperature: 0,
+    extra_body: {
+      enable_thinking: false,
+    },
+  });
+
+  if (!content || content.trim().length === 0) {
+    throw new Error("Empty summary returned");
+  }
+
+  // Remove think content
+  const thinkRegex = /<think>[\s\S]*?<\/think>/g;
+
+  return content.replace(thinkRegex, "").trim();
+}
+
+export function getCurrentHypaV3Preset(): HypaV3Preset {
+  const db = getDatabase();
+  const preset = db.hypaV3Presets?.[db.hypaV3PresetId];
+
+  if (!preset) {
+    throw new Error("Preset not found. Please select a valid preset.");
+  }
+
+  return preset;
+}
+
+export function createHypaV3Preset(
+  name = "New Preset",
+  existingSettings = {}
+): HypaV3Preset {
+  const settings: HypaV3Settings = {
+    summarizationModel: "subModel",
+    summarizationPrompt: "",
+    memoryTokensRatio: 0.2,
+    extraSummarizationRatio: 0,
+    maxChatsPerSummary: 6,
+    recentMemoryRatio: 0.4,
+    similarMemoryRatio: 0.4,
+    enableSimilarityCorrection: false,
+    preserveOrphanedMemory: false,
+    processRegexScript: false,
+    doNotSummarizeUserMessage: false,
+    // Experimental
+    useExperimentalImpl: false,
+    summarizationRequestsPerMinute: 20,
+    summarizationMaxConcurrent: 1,
+    embeddingRequestsPerMinute: 100,
+    embeddingMaxConcurrent: 1,
+  };
+
+  if (
+    existingSettings &&
+    typeof existingSettings === "object" &&
+    !Array.isArray(existingSettings)
+  ) {
+    for (const [key, value] of Object.entries(existingSettings)) {
+      if (key in settings && typeof value === typeof settings[key]) {
+        settings[key] = value;
+      }
+    }
+  }
+
+  return {
+    name,
+    settings,
+  };
+}
+
+interface SummaryChunkVector {
   chunk: SummaryChunk;
   vector: memoryVector;
-};
+}
 
 class HypaProcesserEx extends HypaProcesser {
   // Maintain references to SummaryChunks and their associated memoryVectors
   summaryChunkVectors: SummaryChunkVector[] = [];
 
-  // Calculate dot product similarity between two vectors
-  similarity(a: VectorArray, b: VectorArray): number {
-    let dot = 0;
-
-    for (let i = 0; i < a.length; i++) {
-      dot += a[i] * b[i];
-    }
-
-    return dot;
-  }
-
   async addSummaryChunks(chunks: SummaryChunk[]): Promise<void> {
     // Maintain the superclass's caching structure by adding texts
     const texts = chunks.map((chunk) => chunk.text);
@@ -977,7 +1803,7 @@ class HypaProcesserEx extends HypaProcesser {
     return this.summaryChunkVectors
       .map((scv) => ({
         chunk: scv.chunk,
-        similarity: this.similarity(queryVector, scv.vector.embedding),
+        similarity: similarity(queryVector, scv.vector.embedding),
       }))
       .sort((a, b) => (a.similarity > b.similarity ? -1 : 0))
       .map((result) => [result.chunk, result.similarity]);
diff --git a/src/ts/process/memory/taskRateLimiter.ts b/src/ts/process/memory/taskRateLimiter.ts
new file mode 100644
index 00000000..0d6cfae3
--- /dev/null
+++ b/src/ts/process/memory/taskRateLimiter.ts
@@ -0,0 +1,188 @@
+export interface TaskRateLimiterOptions {
+  tasksPerMinute?: number;
+  maxConcurrentTasks?: number;
+  failFast?: boolean;
+}
+
+export interface BatchResult<TData> {
+  results: TaskResult<TData>[];
+  successCount: number;
+  failureCount: number;
+  allSucceeded: boolean;
+}
+
+export interface TaskResult<TData> {
+  success: boolean;
+  data?: TData;
+  error?: Error;
+}
+
+export class TaskRateLimiter {
+  private static readonly LOG_PREFIX = "[TaskRateLimiter]";
+  public readonly options: TaskRateLimiterOptions;
+  public taskQueueChangeCallback: (queuedCount: number) => void = null;
+  private timestamps: number[] = [];
+  private active: number = 0;
+  private queue: Array<{
+    task: () => Promise<TaskResult<any>>;
+    resolve: (result: TaskResult<any>) => void;
+  }> = [];
+
+  public constructor(options?: TaskRateLimiterOptions) {
+    this.options = {
+      tasksPerMinute: 20,
+      maxConcurrentTasks: 5,
+      failFast: true,
+      ...options,
+    };
+
+    if (this.options.maxConcurrentTasks > this.options.tasksPerMinute) {
+      throw new Error("maxConcurrentTasks must be less than tasksPerMinute");
+    }
+  }
+
+  public async executeTask<TData>(
+    task: () => Promise<TData>
+  ): Promise<TaskResult<TData>> {
+    return new Promise<TaskResult<TData>>((resolve) => {
+      this.queue.push({
+        task: async () => {
+          try {
+            const data = await task();
+            return { success: true, data };
+          } catch (error) {
+            return { success: false, error };
+          }
+        },
+        resolve,
+      });
+
+      this.taskQueueChangeCallback?.(this.queue.length);
+      this.processNextFromQueue();
+    });
+  }
+
+  public async executeBatch<TData>(
+    tasks: Array<() => Promise<TData>>
+  ): Promise<BatchResult<TData>> {
+    const taskResults = await Promise.all(
+      tasks.map((task) => this.executeTask(task))
+    );
+    const successCount = taskResults.filter((r) => r.success).length;
+    const failureCount = taskResults.length - successCount;
+
+    return {
+      results: taskResults,
+      successCount,
+      failureCount,
+      allSucceeded: failureCount === 0,
+    };
+  }
+
+  public cancelPendingTasks(reason: string): void {
+    const error = new TaskCanceledError(reason);
+
+    while (this.queue.length > 0) {
+      const { resolve } = this.queue.shift();
+      resolve({ success: false, error });
+    }
+
+    this.taskQueueChangeCallback?.(this.queue.length);
+  }
+
+  public get queuedTaskCount(): number {
+    return this.queue.length;
+  }
+
+  private processNextFromQueue(): void {
+    if (this.queue.length === 0) return;
+
+    if (this.active >= this.options.maxConcurrentTasks) {
+      // Debug log for concurrency limit hit
+      console.debug(
+        TaskRateLimiter.LOG_PREFIX,
+        "Concurrency limit hit:",
+        "\nTasks in last minute:",
+        this.timestamps.length + "/" + this.options.tasksPerMinute,
+        "\nActive tasks:",
+        this.active + "/" + this.options.maxConcurrentTasks,
+        "\nWaiting tasks in queue:",
+        this.queue.length
+      );
+
+      return;
+    }
+
+    this.timestamps = this.timestamps.filter(
+      (ts) => Date.now() - ts <= 60 * 1000
+    );
+
+    if (this.timestamps.length >= this.options.tasksPerMinute) {
+      const oldestTimestamp = Math.min(...this.timestamps);
+      const timeUntilExpiry = Math.max(
+        100,
+        60 * 1000 - (Date.now() - oldestTimestamp)
+      );
+
+      // Debug log for rate limit hit
+      console.debug(
+        TaskRateLimiter.LOG_PREFIX,
+        "Rate limit hit:",
+        "\nTasks in last minute:",
+        this.timestamps.length + "/" + this.options.tasksPerMinute,
+        "\nActive tasks:",
+        this.active + "/" + this.options.maxConcurrentTasks,
+        "\nWaiting tasks in queue:",
+        this.queue.length,
+        "\nWill retry in:",
+        timeUntilExpiry + "ms"
+      );
+
+      // Wait until rate limit window advances before retrying
+      setTimeout(() => this.processNextFromQueue(), timeUntilExpiry);
+      return;
+    }
+
+    const { task, resolve } = this.queue.shift();
+
+    this.active++;
+    this.taskQueueChangeCallback?.(this.queue.length);
+    this.timestamps.push(Date.now());
+
+    // Debug log for task start
+    console.debug(
+      TaskRateLimiter.LOG_PREFIX,
+      "Task started:",
+      "\nTasks in last minute:",
+      this.timestamps.length + "/" + this.options.tasksPerMinute,
+      "\nActive tasks:",
+      this.active + "/" + this.options.maxConcurrentTasks,
+      "\nWaiting tasks in queue:",
+      this.queue.length
+    );
+
+    task()
+      .then((result) => {
+        resolve(result);
+
+        if (!result.success && this.options.failFast) {
+          this.cancelPendingTasks("Task canceled due to previous failure");
+        }
+      })
+      .finally(() => {
+        this.active--;
+
+        // Prevents call stack overflow while maintaining concurrency limits
+        queueMicrotask(() => this.processNextFromQueue());
+      });
+  }
+}
+
+export class TaskCanceledError extends Error {
+  public readonly name: string;
+
+  public constructor(message: string) {
+    super(message);
+    this.name = "TaskCanceledError";
+  }
+}
diff --git a/src/ts/process/transformers.ts b/src/ts/process/transformers.ts
index 4e238fa2..ea3a32d6 100644
--- a/src/ts/process/transformers.ts
+++ b/src/ts/process/transformers.ts
@@ -60,12 +60,19 @@ export const runEmbedding = async (texts: string[], model:EmbeddingModel = 'Xeno
     console.log('running embedding')
     let embeddingModelQuery = model + device
     if(!extractor || embeddingModelQuery !== lastEmbeddingModelQuery){
+        // Dispose old extractor
+        if(extractor) {
+            await extractor.dispose()
+        }
         extractor = await pipeline('feature-extraction', model, {
+            // Default dtype for webgpu is fp32, so we can use q8, which is the default dtype in wasm.
+            ...(device === 'webgpu' ? { dtype: "q8" } : {}),
             device: device,
             progress_callback: (progress) => {
                 console.log(progress)
             }
         });
+        lastEmbeddingModelQuery = embeddingModelQuery
         console.log('extractor loaded')
     }
     let result = await extractor(texts, { pooling: 'mean', normalize: true });
diff --git a/src/ts/process/webllm.ts b/src/ts/process/webllm.ts
new file mode 100644
index 00000000..cb1c1485
--- /dev/null
+++ b/src/ts/process/webllm.ts
@@ -0,0 +1,60 @@
+import {
+  type ChatCompletionMessageParam,
+  type ChatCompletionRequestNonStreaming,
+  MLCEngine,
+  CreateMLCEngine,
+} from "@mlc-ai/web-llm";
+
+let engine: MLCEngine = null;
+let lastModel: string = null;
+
+export async function chatCompletion(
+  messages: { role: string; content: string }[],
+  model: string,
+  config: Record<string, any>
+): Promise<string> {
+  try {
+    if (!engine || lastModel !== model) {
+      if (engine) engine.unload();
+
+      const initProgressCallback = (progress) => {
+        console.log("[WebLLM]", progress);
+      };
+
+      engine = await CreateMLCEngine(
+        model,
+        {
+          initProgressCallback,
+        },
+        { context_window_size: 16384 }
+      );
+
+      lastModel = model;
+    }
+
+    const request: ChatCompletionRequestNonStreaming = {
+      messages: messages as ChatCompletionMessageParam[],
+      temperature: 0,
+      max_tokens: 4096,
+      ...config,
+    };
+    const completion = await engine.chat.completions.create(request);
+    const content = completion.choices[0].message.content;
+
+    return content;
+  } catch (error) {
+    if (error instanceof Error) {
+      throw error;
+    }
+
+    throw new Error(JSON.stringify(error));
+  }
+}
+
+export async function unloadEngine(): Promise<void> {
+  if (!engine) return;
+
+  await engine.unload();
+  engine = null;
+  lastModel = null;
+}
diff --git a/src/ts/storage/database.svelte.ts b/src/ts/storage/database.svelte.ts
index 7a5fb478..60c9bf93 100644
--- a/src/ts/storage/database.svelte.ts
+++ b/src/ts/storage/database.svelte.ts
@@ -11,6 +11,7 @@ import { prebuiltNAIpresets, prebuiltPresets } from '../process/templates/templa
 import { defaultColorScheme, type ColorScheme } from '../gui/colorscheme';
 import type { PromptItem, PromptSettings } from '../process/prompt';
 import type { OobaChatCompletionRequestParams } from '../model/ooba';
+import { type HypaV3Settings, type HypaV3Preset, createHypaV3Preset } from '../process/memory/hypav3'
 
 export let appVer = "159.0.0"
 export let webAppSubVer = ''
@@ -515,17 +516,21 @@ export function setDatabase(data:Database){
     data.checkCorruption ??= true
     data.OaiCompAPIKeys ??= {}
     data.reasoningEffort ??= 0
-    data.hypaV3Settings = {
-        memoryTokensRatio: data.hypaV3Settings?.memoryTokensRatio ?? 0.2,
-        extraSummarizationRatio: data.hypaV3Settings?.extraSummarizationRatio ?? 0,
-        maxChatsPerSummary: data.hypaV3Settings?.maxChatsPerSummary ?? 4,
-        recentMemoryRatio: data.hypaV3Settings?.recentMemoryRatio ?? 0.4,
-        similarMemoryRatio: data.hypaV3Settings?.similarMemoryRatio ?? 0.4,
-        enableSimilarityCorrection: data.hypaV3Settings?.enableSimilarityCorrection ?? false,
-        preserveOrphanedMemory: data.hypaV3Settings?.preserveOrphanedMemory ?? false,
-        processRegexScript: data.hypaV3Settings?.processRegexScript ?? false,
-        doNotSummarizeUserMessage: data.hypaV3Settings?.doNotSummarizeUserMessage ?? false
+    data.hypaV3Presets ??= [
+        createHypaV3Preset("Default", {
+            summarizationPrompt: data.supaMemoryPrompt ? data.supaMemoryPrompt : "",
+            ...data.hypaV3Settings
+        })
+    ]
+    if (data.hypaV3Presets.length > 0) {
+        data.hypaV3Presets = data.hypaV3Presets.map((preset, i) =>
+            createHypaV3Preset(
+                preset.name || `Preset ${i + 1}`,
+                preset.settings || {}
+            )
+        )
     }
+    data.hypaV3PresetId ??= 0
     data.returnCSSError ??= true
     data.useExperimentalGoogleTranslator ??= false
     if(data.antiClaudeOverload){ //migration
@@ -535,7 +540,7 @@ export function setDatabase(data:Database){
     data.hypaCustomSettings = {
         url: data.hypaCustomSettings?.url ?? "",
         key: data.hypaCustomSettings?.key ?? "",
-        model: data.hypaCustomSettings?.model ?? "",       
+        model: data.hypaCustomSettings?.model ?? ""     
     }
     data.doNotChangeSeperateModels ??= false
     data.modelTools ??= []
@@ -960,17 +965,10 @@ export interface Database{
     showPromptComparison:boolean
     checkCorruption:boolean
     hypaV3:boolean
-    hypaV3Settings: {
-        memoryTokensRatio: number
-        extraSummarizationRatio: number
-        maxChatsPerSummary: number
-        recentMemoryRatio: number
-        similarMemoryRatio: number
-        enableSimilarityCorrection: boolean
-        preserveOrphanedMemory: boolean
-        processRegexScript: boolean
-        doNotSummarizeUserMessage: boolean
-    }
+    hypaV3Settings: HypaV3Settings // legacy
+    hypaV3Presets: HypaV3Preset[]
+    hypaV3PresetId: number
+    showMenuHypaMemoryModal:boolean
     OaiCompAPIKeys: {[key:string]:string}
     inlayErrorResponse:boolean
     reasoningEffort:number
@@ -1025,7 +1023,6 @@ export interface Database{
     }[]
     igpPrompt:string
     useTokenizerCaching:boolean
-    showMenuHypaMemoryModal:boolean
 }
 
 interface SeparateParameters{
diff --git a/src/ts/stores.svelte.ts b/src/ts/stores.svelte.ts
index ba714e6c..316ca189 100644
--- a/src/ts/stores.svelte.ts
+++ b/src/ts/stores.svelte.ts
@@ -50,6 +50,13 @@ export const alertStore = writable({
     type: 'none',
     msg: 'n',
 } as alertData)
+export const hypaV3ModalOpen = writable(false)
+export const hypaV3ProgressStore = writable({
+    open: false,
+    miniMsg: '',
+    msg: '',
+    subMsg: '',
+})
 export const selIdState = $state({
     selId: -1
 })