diff --git a/android/app/src/main/AndroidManifest.xml b/android/app/src/main/AndroidManifest.xml
index aef434f..32d6a67 100644
--- a/android/app/src/main/AndroidManifest.xml
+++ b/android/app/src/main/AndroidManifest.xml
@@ -37,7 +37,7 @@
         android:allowBackup="true"
         android:icon="@mipmap/ic_launcher"
         android:label="Nova"
-
+        android:largeHeap="true"
         android:supportsRtl="true"
         android:theme="@style/LaunchTheme"
         tools:targetApi="34">
diff --git a/android/app/src/main/kotlin/dev/nova/assistant/MainActivity.kt b/android/app/src/main/kotlin/dev/nova/assistant/MainActivity.kt
index 8557cb2..d1cb610 100644
--- a/android/app/src/main/kotlin/dev/nova/assistant/MainActivity.kt
+++ b/android/app/src/main/kotlin/dev/nova/assistant/MainActivity.kt
@@ -82,6 +82,7 @@ class MainActivity : FlutterActivity() {
                                 "pssKb" to pssKb,
                                 "rssKb" to pssKb,
                                 "availMemMb" to (info.availMem / (1024 * 1024)),
+                                "totalMemMb" to (info.totalMem / (1024 * 1024)),
                             )
                         )
                     }
diff --git a/android/app/src/main/kotlin/dev/nova/assistant/NovaApplication.kt b/android/app/src/main/kotlin/dev/nova/assistant/NovaApplication.kt
index 9393b1d..dfedb73 100644
--- a/android/app/src/main/kotlin/dev/nova/assistant/NovaApplication.kt
+++ b/android/app/src/main/kotlin/dev/nova/assistant/NovaApplication.kt
@@ -1,9 +1,173 @@
 package dev.nova.assistant
 
 import android.app.Application
+import android.util.Log
+import java.io.BufferedInputStream
+import java.io.BufferedOutputStream
+import java.io.File
+import java.io.FileInputStream
+import java.io.FileOutputStream
 
+/**
+ * Scrubs oversized chat-history keys from Flutter SharedPreferences **without**
+ * calling [getSharedPreferences]. Loading a 100MB+ prefs XML OOMs during
+ * KXmlParser (POCO F1 / mid-RAM devices).
+ */
 class NovaApplication : Application() {
     override fun onCreate() {
         super.onCreate()
+        scrubOversizedFlutterPrefs()
     }
-}
\ No newline at end of file
+
+    private fun scrubOversizedFlutterPrefs() {
+        val prefsFile = File(dataDir, "shared_prefs/$FLUTTER_PREFS.xml")
+        val size = if (prefsFile.exists()) prefsFile.length() else 0L
+        // #region agent log
+        Log.i(
+            TAG,
+            "AGENT_DBG session=7abc09 hyp=A scrub start size=$size path=${prefsFile.absolutePath}",
+        )
+        // #endregion
+        if (!prefsFile.exists() || size == 0L) return
+
+        try {
+            var removed = false
+            for (key in listOf(CONVERSATIONS_KEY, OLD_CHAT_KEY)) {
+                if (streamRemoveStringEntry(prefsFile, key)) {
+                    removed = true
+                    // #region agent log
+                    Log.i(TAG, "AGENT_DBG session=7abc09 hyp=A removed key=$key")
+                    // #endregion
+                }
+            }
+
+            val after = prefsFile.length()
+            // #region agent log
+            Log.i(
+                TAG,
+                "AGENT_DBG session=7abc09 hyp=A scrub done removed=$removed sizeAfter=$after",
+            )
+            // #endregion
+
+            if (after > MAX_SAFE_PREFS_BYTES) {
+                // Do not copyTo() a huge file (that also OOMs). Just replace.
+                prefsFile.delete()
+                prefsFile.writeText(EMPTY_PREFS_XML)
+                // #region agent log
+                Log.w(
+                    TAG,
+                    "AGENT_DBG session=7abc09 hyp=A wiped prefs after=$after",
+                )
+                // #endregion
+            }
+        } catch (e: Exception) {
+            Log.e(TAG, "AGENT_DBG session=7abc09 hyp=A scrub failed", e)
+            try {
+                prefsFile.writeText(EMPTY_PREFS_XML)
+            } catch (_: Exception) {
+                // User must clear app data.
+            }
+        }
+    }
+
+    /**
+     * Removes `<string name="KEY">…</string>` via streaming so the huge value
+     * is never materialized as a Java String.
+     */
+    private fun streamRemoveStringEntry(prefsFile: File, key: String): Boolean {
+        val open = """<string name="$key">""".toByteArray(Charsets.UTF_8)
+        val close = "</string>".toByteArray(Charsets.UTF_8)
+        val tmp = File(prefsFile.parent, "${prefsFile.name}.tmp")
+        var found = false
+
+        FileInputStream(prefsFile).use { fis ->
+            BufferedInputStream(fis, 256 * 1024).use { input ->
+                FileOutputStream(tmp).use { fos ->
+                    BufferedOutputStream(fos, 256 * 1024).use { output ->
+                        val buffer = ByteArray(256 * 1024)
+                        val pending = ArrayList<Byte>(open.size)
+                        var openIdx = 0
+                        var skipping = false
+                        var closeIdx = 0
+
+                        while (true) {
+                            val n = input.read(buffer)
+                            if (n < 0) break
+                            var i = 0
+                            while (i < n) {
+                                val byte = buffer[i]
+
+                                if (skipping) {
+                                    if (byte == close[closeIdx]) {
+                                        closeIdx++
+                                        if (closeIdx == close.size) {
+                                            skipping = false
+                                            closeIdx = 0
+                                            found = true
+                                        }
+                                    } else {
+                                        closeIdx = if (byte == close[0]) 1 else 0
+                                    }
+                                    i++
+                                    continue
+                                }
+
+                                if (byte == open[openIdx]) {
+                                    pending.add(byte)
+                                    openIdx++
+                                    if (openIdx == open.size) {
+                                        skipping = true
+                                        openIdx = 0
+                                        pending.clear()
+                                    }
+                                    i++
+                                    continue
+                                }
+
+                                if (openIdx > 0) {
+                                    for (p in pending) {
+                                        output.write(p.toInt())
+                                    }
+                                    pending.clear()
+                                    openIdx = 0
+                                    continue
+                                }
+
+                                output.write(byte.toInt())
+                                i++
+                            }
+                        }
+
+                        if (!skipping) {
+                            for (p in pending) {
+                                output.write(p.toInt())
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        if (found) {
+            if (!tmp.renameTo(prefsFile)) {
+                tmp.copyTo(prefsFile, overwrite = true)
+                tmp.delete()
+            }
+        } else {
+            tmp.delete()
+        }
+
+        return found
+    }
+
+    companion object {
+        private const val TAG = "NovaApplication"
+        private const val FLUTTER_PREFS = "FlutterSharedPreferences"
+        private const val CONVERSATIONS_KEY = "flutter.conversations"
+        private const val OLD_CHAT_KEY = "flutter.chat_history"
+        private const val MAX_SAFE_PREFS_BYTES = 4L * 1024L * 1024L
+        private const val EMPTY_PREFS_XML =
+            "<?xml version='1.0' encoding='utf-8' standalone='yes' ?>\n" +
+                "<map>\n</map>\n"
+    }
+}
diff --git a/docs/models.md b/docs/models.md
index 344059f..bf00699 100644
--- a/docs/models.md
+++ b/docs/models.md
@@ -57,4 +57,24 @@ When a thread gets long, **Compact context** (manual button in chat) or **Auto-c
 | 8 GB+ | Gemma 4 E2B OK; close heavy apps first |
 | Debug all day | Enable battery/idle unload; avoid continuous screen capture |
 
+Onboarding uses **total device RAM** (not only free RAM): phones under ~6.5 GB
+default to **SmolLM** so a wiped POCO F1 does not start a 2.4 GB Gemma 4 download
+it cannot load for chat.
+
 Android idle unload is shorter (~2 minutes) to reduce LMK pressure.
+
+## On-device image generation (not in this build)
+
+Repos such as
+[FLUX.2-klein-4B-LiteRT](https://huggingface.co/litert-community/FLUX.2-klein-4B-LiteRT)
+and
+[Z-Image-Turbo-LiteRT](https://huggingface.co/litert-community/Z-Image-Turbo-LiteRT)
+are **LiteRT `CompiledModel` multi-graph diffusion pipelines** (many `.tflite`
+chunks + host tokenizer/scheduler), not `flutter_gemma` / LiteRT-LM chat
+(`.litertlm` / `.task`) models.
+
+Nova’s stack only runs LLM/VLM chat via `flutter_gemma`. Shipping those image
+models would mean a separate native LiteRT GPU pipeline, multi‑GB downloads, and
+phones in the Pixel 8a / 8 GB+ class for usable demos — out of scope for the
+current APK (and unsuitable for POCO F1). Prefer **remote LAN** or a PC host
+for image generation for now.
diff --git a/lib/main.dart b/lib/main.dart
index 88ebd12..768390b 100644
--- a/lib/main.dart
+++ b/lib/main.dart
@@ -27,13 +27,28 @@ import 'package:nova_assistant/services/chat_history_service.dart';
 import 'package:nova_assistant/services/widget_service.dart';
 import 'package:nova_assistant/models/conversation.dart';
 import 'package:nova_assistant/models/user_preferences.dart';
+import 'package:nova_assistant/utils/agent_debug_log.dart';
 import 'package:shared_preferences/shared_preferences.dart';
 
 void main() async {
   WidgetsFlutterBinding.ensureInitialized();
 
   try {
+    // #region agent log
+    await AgentDebugLog.log(
+      hypothesisId: 'A',
+      location: 'main.dart:beforePrefs',
+      message: 'About to call SharedPreferences.getInstance',
+    );
+    // #endregion
     await SharedPreferences.getInstance();
+    // #region agent log
+    await AgentDebugLog.log(
+      hypothesisId: 'A',
+      location: 'main.dart:afterPrefs',
+      message: 'SharedPreferences.getInstance succeeded',
+    );
+    // #endregion
 
     await FlutterGemma.initialize(
       inferenceEngines: const [LiteRtLmEngine(), MediaPipeEngine()],
@@ -55,6 +70,7 @@ void main() async {
     await TtsService.instance.initialize();
     await PromptPresetsService.instance.initialize();
 
+    await ModelOrchestrator.instance.applyRamAwareModelDefaults();
     _prefetchModels();
   } catch (e) {
     debugPrint('Initialization error: $e');
diff --git a/lib/screens/assistant_screen.dart b/lib/screens/assistant_screen.dart
index 996ce0a..9d68383 100644
--- a/lib/screens/assistant_screen.dart
+++ b/lib/screens/assistant_screen.dart
@@ -1943,11 +1943,14 @@ class _AssistantScreenState extends State<AssistantScreen>
             ),
           ),
           const SizedBox(width: 6),
-          Text(
-            _status,
-            style: TextStyle(fontSize: 11, color: Colors.grey[500]),
+          Expanded(
+            child: Text(
+              _status,
+              maxLines: 1,
+              overflow: TextOverflow.ellipsis,
+              style: TextStyle(fontSize: 11, color: Colors.grey[500]),
+            ),
           ),
-          const Spacer(),
           if (_thinkingMode)
             Container(
               padding: const EdgeInsets.symmetric(horizontal: 8, vertical: 2),
diff --git a/lib/screens/onboarding/beginner/model_download_screen.dart b/lib/screens/onboarding/beginner/model_download_screen.dart
index 62d2076..822d6a1 100644
--- a/lib/screens/onboarding/beginner/model_download_screen.dart
+++ b/lib/screens/onboarding/beginner/model_download_screen.dart
@@ -4,7 +4,10 @@ import 'package:file_picker/file_picker.dart';
 import 'package:flutter/material.dart';
 import 'package:nova_assistant/models/model_info.dart';
 import 'package:nova_assistant/models/user_preferences.dart';
+import 'package:nova_assistant/services/memory_diagnostics_service.dart';
 import 'package:nova_assistant/services/model_manager.dart';
+import 'package:nova_assistant/services/model_orchestrator.dart';
+import 'package:nova_assistant/services/platform_adaptation_service.dart';
 import 'package:nova_assistant/services/user_preferences_service.dart';
 import 'package:nova_assistant/utils/agent_debug_log.dart';
 
@@ -18,29 +21,68 @@ class ModelDownloadScreen extends StatefulWidget {
 }
 
 class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
-  static const _targetModel = NovaModel.gemma4E2b;
+  NovaModel _targetModel = NovaModel.smollm;
+  bool _resolvedTarget = false;
 
   double _progress = 0;
-  String _status = 'Preparing...';
+  String _status = 'Checking device memory...';
   bool _isComplete = false;
   bool _hasError = false;
   String _errorMessage = '';
   bool _isBusy = false;
 
+  String get _capabilitiesLabel {
+    final parts = <String>[];
+    if (_targetModel.hasVision) parts.add('Vision');
+    if (_targetModel.hasThinking) parts.add('Thinking');
+    parts.add('Chat');
+    return parts.join(' + ');
+  }
+
   @override
   void initState() {
     super.initState();
-    _checkExisting();
+    _resolveTargetAndCheck();
   }
 
   Future<void> _finishReady() async {
     await UserPreferencesService.instance.setMode(UserMode.beginner);
+    ModelOrchestrator.instance.preferredModelType = _targetModel;
+    ModelOrchestrator.instance.selector.primaryHeavy = _targetModel;
+    ModelOrchestrator.instance.selector.fastModel = NovaModel.smollm;
     await Future<void>.delayed(const Duration(milliseconds: 800));
     if (mounted) {
       widget.onDownloadComplete();
     }
   }
 
+  Future<void> _resolveTargetAndCheck() async {
+    final recommended = await PlatformAdaptationService.instance
+        .recommendModelForDevice();
+    // #region agent log
+    await AgentDebugLog.log(
+      hypothesisId: 'E',
+      location: 'model_download_screen.dart:_resolveTargetAndCheck',
+      message: 'Onboarding model recommendation',
+      data: {
+        'recommended': recommended.displayName,
+        'sizeMB': recommended.sizeMB,
+        'availMemMb': MemoryDiagnosticsService.instance.lastAvailMemMb,
+        'totalMemMb': MemoryDiagnosticsService.instance.lastTotalMemMb,
+      },
+      runId: 'post-fix',
+    );
+    // #endregion
+
+    if (!mounted) return;
+    setState(() {
+      _targetModel = recommended;
+      _resolvedTarget = true;
+      _status = 'Choose how to install ${recommended.displayName}';
+    });
+    await _checkExisting();
+  }
+
   Future<void> _checkExisting() async {
     final fileName = ModelHuggingFaceURLs.fileNameFor(_targetModel);
     final path = await ModelManager.instance.findModelPath(fileName);
@@ -69,7 +111,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
 
     if (mounted) {
       setState(() {
-        _status = 'Choose how to install Gemma 4 E2B';
+        _status = 'Choose how to install ${_targetModel.displayName}';
         _progress = 0;
       });
     }
@@ -80,7 +122,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
       _isBusy = true;
       _progress = 0;
       _hasError = false;
-      _status = 'Downloading Gemma 4 E2B...';
+      _status = 'Downloading ${_targetModel.displayName}...';
     });
 
     try {
@@ -157,7 +199,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
         if (mounted) {
           setState(() {
             _isBusy = false;
-            _status = 'Choose how to install Gemma 4 E2B';
+            _status = 'Choose how to install ${_targetModel.displayName}';
           });
         }
         return;
@@ -188,8 +230,8 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
           _isBusy = false;
           _hasError = true;
           _errorMessage =
-              'Import failed. Select a .litertlm file '
-              '(e.g. gemma-4-E2B-it.litertlm).';
+              'Import failed. Select a matching model file '
+              '(e.g. ${ModelHuggingFaceURLs.fileNameFor(_targetModel)}).';
         });
       }
     } catch (e) {
@@ -205,6 +247,12 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
 
   @override
   Widget build(BuildContext context) {
+    final sizeLabel = '~${_targetModel.sizeMB} MB';
+    final subtitle = _resolvedTarget
+        ? '${_targetModel.displayName} — download ($sizeLabel) or import a\n'
+              'model file you already have. Chosen for this device\'s free RAM.'
+        : 'Checking free RAM to pick a model that can actually load...';
+
     return Padding(
       padding: const EdgeInsets.all(32),
       child: Column(
@@ -239,8 +287,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
           ),
           const SizedBox(height: 12),
           Text(
-            'Gemma 4 E2B — download (~2.4 GB) or import a\n'
-            '.litertlm file you already have.',
+            subtitle,
             textAlign: TextAlign.center,
             style: TextStyle(
               color: Colors.grey[400],
@@ -249,7 +296,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
             ),
           ),
           const SizedBox(height: 48),
-          _buildModelInfo(),
+          _buildModelInfo(sizeLabel),
           const SizedBox(height: 32),
           if (_hasError) ...[
             Container(
@@ -276,7 +323,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
             SizedBox(
               width: double.infinity,
               child: FilledButton(
-                onPressed: _isBusy ? null : _downloadModel,
+                onPressed: _isBusy || !_resolvedTarget ? null : _downloadModel,
                 style: FilledButton.styleFrom(
                   backgroundColor: const Color(0xFF6C63FF),
                   padding: const EdgeInsets.symmetric(vertical: 14),
@@ -291,14 +338,16 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
               ),
             ),
             TextButton(
-              onPressed: _isBusy ? null : _importFromStorage,
+              onPressed: _isBusy || !_resolvedTarget
+                  ? null
+                  : _importFromStorage,
               child: const Text('Import from storage instead'),
             ),
           ] else if (!_isComplete && !_isBusy) ...[
             SizedBox(
               width: double.infinity,
               child: FilledButton(
-                onPressed: _downloadModel,
+                onPressed: !_resolvedTarget ? null : _downloadModel,
                 style: FilledButton.styleFrom(
                   backgroundColor: const Color(0xFF6C63FF),
                   padding: const EdgeInsets.symmetric(vertical: 14),
@@ -316,7 +365,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
             SizedBox(
               width: double.infinity,
               child: OutlinedButton(
-                onPressed: _importFromStorage,
+                onPressed: !_resolvedTarget ? null : _importFromStorage,
                 style: OutlinedButton.styleFrom(
                   foregroundColor: Colors.white,
                   side: BorderSide(color: Colors.white.withValues(alpha: 0.3)),
@@ -334,13 +383,16 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
             Row(
               mainAxisAlignment: MainAxisAlignment.spaceBetween,
               children: [
-                Text(
-                  _status,
-                  style: TextStyle(
-                    color: _isComplete
-                        ? const Color(0xFF4CAF50)
-                        : Colors.grey[400],
-                    fontSize: 13,
+                Flexible(
+                  child: Text(
+                    _status,
+                    overflow: TextOverflow.ellipsis,
+                    style: TextStyle(
+                      color: _isComplete
+                          ? const Color(0xFF4CAF50)
+                          : Colors.grey[400],
+                      fontSize: 13,
+                    ),
                   ),
                 ),
                 Text(
@@ -389,7 +441,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
             )
           else if (!_hasError && !_isBusy)
             Text(
-              'Tip: If you already have the .litertlm file, use Import.',
+              'Tip: If you already have the model file, use Import.',
               textAlign: TextAlign.center,
               style: TextStyle(color: Colors.grey[600], fontSize: 12),
             ),
@@ -399,7 +451,7 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
     );
   }
 
-  Widget _buildModelInfo() {
+  Widget _buildModelInfo(String sizeLabel) {
     return Container(
       padding: const EdgeInsets.all(16),
       decoration: BoxDecoration(
@@ -409,13 +461,17 @@ class _ModelDownloadScreenState extends State<ModelDownloadScreen> {
       ),
       child: Column(
         children: [
-          _buildInfoRow('Model', 'Gemma 4 E2B', Icons.model_training_outlined),
+          _buildInfoRow(
+            'Model',
+            _targetModel.displayName,
+            Icons.model_training_outlined,
+          ),
           const SizedBox(height: 12),
-          _buildInfoRow('Size', '~2400 MB', Icons.storage_outlined),
+          _buildInfoRow('Size', sizeLabel, Icons.storage_outlined),
           const SizedBox(height: 12),
           _buildInfoRow(
             'Capabilities',
-            'Vision + Thinking + Function Calling',
+            _capabilitiesLabel,
             Icons.auto_awesome_outlined,
           ),
         ],
diff --git a/lib/services/chat_history_service.dart b/lib/services/chat_history_service.dart
index abe1ac1..a807802 100644
--- a/lib/services/chat_history_service.dart
+++ b/lib/services/chat_history_service.dart
@@ -1,40 +1,133 @@
 import 'dart:convert';
+import 'dart:io';
 
 import 'package:flutter/foundation.dart';
+import 'package:path_provider/path_provider.dart';
 import 'package:shared_preferences/shared_preferences.dart';
 import 'package:nova_assistant/models/chat_message.dart';
 import 'package:nova_assistant/models/conversation.dart';
 
+/// Persists conversations as a JSON file — never in SharedPreferences.
+///
+/// Storing chat (especially base64 screenshots) in prefs OOMs Android when
+/// Flutter decodes the prefs map across the platform channel (~256 MB Java
+/// heap). Native [NovaApplication] migrates/drops the old prefs key on boot.
 class ChatHistoryService {
   static const _oldKey = 'chat_history';
   static const _key = 'conversations';
-  static const _migratedKey = 'conversations_migrated';
+  static const _fileName = 'conversations.json';
+  static const _maxConversations = 40;
+  static const _maxMessagesPerConversation = 200;
+
+  /// Soft cap for the on-disk blob; oversized files are trimmed on load.
+  static const _maxFileBytes = 4 * 1024 * 1024;
+
   static SharedPreferences? _prefs;
   static List<Conversation>? _cachedConversations;
+  static File? _file;
 
   static Future<SharedPreferences> get _p async =>
       _prefs ??= await SharedPreferences.getInstance();
 
-  static Future<void> ensureMigrated() async {
-    final p = await _p;
-    final alreadyMigrated = p.getBool(_migratedKey) ?? false;
-    if (alreadyMigrated) return;
+  static Future<File> _conversationsFile() async {
+    if (_file != null) return _file!;
+    final dir = await getApplicationDocumentsDirectory();
+    _file = File('${dir.path}/$_fileName');
+
+    return _file!;
+  }
+
+  /// Strips screenshot bytes — they bloated prefs to 100MB+ and crash startup.
+  static Map<String, dynamic> _messageToPersistJson(ChatMessage message) {
+    final json = message.toJson();
+    json['imageData'] = null;
+
+    return json;
+  }
+
+  static Conversation _sanitizeConversation(Conversation conversation) {
+    final messages = conversation.messages.map((m) {
+      if (m.imageData == null || m.imageData!.isEmpty) return m;
+      final text = m.text.trim().isEmpty ? '[Screenshot]' : m.text;
 
-    final oldJson = p.getString(_oldKey);
-    if (oldJson != null && oldJson.isNotEmpty) {
+      return ChatMessage(
+        id: m.id,
+        text: text,
+        isUser: m.isUser,
+        timestamp: m.timestamp,
+        modelName: m.modelName,
+        isStreaming: m.isStreaming,
+        isError: m.isError,
+        thinking: m.thinking,
+        toolCalls: m.toolCalls,
+        inferenceTimeMs: m.inferenceTimeMs,
+        reactions: m.reactions,
+      );
+    }).toList();
+    final trimmed = messages.length > _maxMessagesPerConversation
+        ? messages.sublist(messages.length - _maxMessagesPerConversation)
+        : messages;
+
+    return conversation.copyWith(messages: trimmed);
+  }
+
+  static List<Conversation> _capConversations(List<Conversation> list) {
+    final sorted = List<Conversation>.from(list)
+      ..sort((a, b) => b.updatedAt.compareTo(a.updatedAt));
+    if (sorted.length <= _maxConversations) {
+      return sorted.map(_sanitizeConversation).toList();
+    }
+
+    return sorted.take(_maxConversations).map(_sanitizeConversation).toList();
+  }
+
+  static Future<void> ensureMigrated() async {
+    final file = await _conversationsFile();
+    if (await file.exists() && await file.length() > 0) {
+      // Drop leftover prefs keys if native migrate already wrote the file.
       try {
-        final list = jsonDecode(oldJson) as List<dynamic>;
-        if (list.isNotEmpty) {
-          final messages = list
-              .map((e) => ChatMessage.fromJson(e as Map<String, dynamic>))
-              .toList();
-          final conversation = Conversation(messages: messages);
-          await _saveConversationsInternal([conversation]);
-          await p.remove(_oldKey);
+        final p = await _p;
+        await p.remove(_key);
+        await p.remove(_oldKey);
+      } on Exception catch (e) {
+        debugPrint('ChatHistoryService: prefs cleanup after file migrate: $e');
+      }
+
+      return;
+    }
+
+    // Legacy path: small prefs payloads only (native drops oversized ones).
+    try {
+      final p = await _p;
+      final json = p.getString(_key);
+      if (json != null && json.isNotEmpty) {
+        await file.writeAsString(json);
+        await p.remove(_key);
+        debugPrint(
+          'ChatHistoryService: migrated conversations prefs → file '
+          '(${json.length} chars)',
+        );
+      }
+
+      final oldJson = p.getString(_oldKey);
+      if (oldJson != null && oldJson.isNotEmpty) {
+        try {
+          final list = jsonDecode(oldJson) as List<dynamic>;
+          if (list.isNotEmpty) {
+            final messages = list
+                .map((e) => ChatMessage.fromJson(e as Map<String, dynamic>))
+                .toList();
+            final conversation = Conversation(messages: messages);
+            await _saveConversationsInternal([conversation]);
+          }
+        } on Exception catch (e) {
+          debugPrint('ChatHistoryService: old chat_history migrate failed: $e');
         }
-      } catch (_) {}
+        await p.remove(_oldKey);
+      }
+    } on Exception catch (e) {
+      debugPrint('ChatHistoryService.ensureMigrated prefs read failed: $e');
     }
-    await p.setBool(_migratedKey, true);
   }
 
   static Future<List<Conversation>> loadConversations() async {
@@ -43,20 +136,49 @@ class ChatHistoryService {
     }
     await ensureMigrated();
     try {
-      final p = await _p;
-      final json = p.getString(_key);
-      if (json == null || json.isEmpty) {
+      final file = await _conversationsFile();
+      if (!await file.exists()) {
         _cachedConversations = [];
+
         return [];
       }
+
+      final length = await file.length();
+      if (length == 0) {
+        _cachedConversations = [];
+
+        return [];
+      }
+
+      if (length > _maxFileBytes * 4) {
+        debugPrint(
+          'ChatHistoryService: conversations file too large '
+          '($length bytes) — resetting to avoid OOM',
+        );
+        await file.delete();
+        _cachedConversations = [];
+
+        return [];
+      }
+
+      final json = await file.readAsString();
       final list = jsonDecode(json) as List<dynamic>;
-      _cachedConversations = list
+      var conversations = list
           .map((e) => Conversation.fromJson(e as Map<String, dynamic>))
           .toList();
-      _cachedConversations!.sort((a, b) => b.updatedAt.compareTo(a.updatedAt));
+      conversations = _capConversations(conversations);
+      _cachedConversations = conversations;
+
+      // Rewrite stripped/capped form so the blob shrinks on disk.
+      if (length > _maxFileBytes) {
+        await _saveConversationsInternal(conversations);
+      }
+
       return List.unmodifiable(_cachedConversations!);
-    } catch (_) {
+    } on Exception catch (e) {
+      debugPrint('ChatHistoryService.loadConversations error: $e');
       _cachedConversations = [];
+
       return [];
     }
   }
@@ -65,11 +187,31 @@ class ChatHistoryService {
     List<Conversation> conversations,
   ) async {
     try {
-      final p = await _p;
-      final json = jsonEncode(conversations.map((c) => c.toJson()).toList());
-      await p.setString(_key, json);
-      _cachedConversations = conversations;
-    } catch (e) {
+      final capped = _capConversations(conversations);
+      final encoded = jsonEncode(
+        capped
+            .map(
+              (c) => {
+                'id': c.id,
+                'title': c.title,
+                'messages': c.messages.map(_messageToPersistJson).toList(),
+                'createdAt': c.createdAt.toIso8601String(),
+                'updatedAt': c.updatedAt.toIso8601String(),
+                'summary': c.summary,
+              },
+            )
+            .toList(),
+      );
+      final file = await _conversationsFile();
+      await file.writeAsString(encoded);
+      _cachedConversations = capped;
+
+      // Ensure prefs never holds the blob again.
+      try {
+        final p = await _p;
+        if (p.containsKey(_key)) await p.remove(_key);
+      } on Exception catch (_) {}
+    } on Exception catch (e) {
       debugPrint('ChatHistoryService._saveConversationsInternal error: $e');
     }
   }
@@ -85,6 +227,7 @@ class ChatHistoryService {
     final conversation = Conversation(title: title);
     final updated = <Conversation>[conversation, ...conversations];
     await _saveConversationsInternal(updated);
+
     return conversation;
   }
 
@@ -106,6 +249,7 @@ class ChatHistoryService {
 
   static Future<Conversation?> getConversation(String id) async {
     final conversations = await loadConversations();
+
     return conversations.where((c) => c.id == id).firstOrNull;
   }
 
@@ -127,32 +271,11 @@ class ChatHistoryService {
     }
   }
 
-  static Future<void> updateMessage(
-    String conversationId,
-    ChatMessage message,
-  ) async {
-    final conversations = await loadConversations();
-    final convoIndex = conversations.indexWhere((c) => c.id == conversationId);
-    if (convoIndex != -1) {
-      final conversation = conversations[convoIndex];
-      final messages = List<ChatMessage>.from(conversation.messages);
-      final msgIndex = messages.indexWhere((m) => m.id == message.id);
-      if (msgIndex != -1) {
-        messages[msgIndex] = message;
-        final updated = List<Conversation>.from(conversations);
-        updated[convoIndex] = conversation.copyWith(
-          messages: messages,
-          updatedAt: DateTime.now(),
-        );
-        await _saveConversationsInternal(updated);
-      }
-    }
-  }
-
   static Future<List<ChatMessage>> load() async {
     final conversations = await loadConversations();
     if (conversations.isEmpty) return [];
     final first = conversations.first;
+
     return first.messages;
   }
 
@@ -181,8 +304,21 @@ class ChatHistoryService {
 
   static Future<void> clear() async {
     _cachedConversations = [];
-    final p = await _p;
-    await p.remove(_key);
+    try {
+      final file = await _conversationsFile();
+      if (await file.exists()) await file.delete();
+    } on Exception catch (e) {
+      debugPrint('ChatHistoryService.clear file error: $e');
+    } finally {
+      _file = null;
+    }
+    try {
+      final p = await _p;
+      await p.remove(_key);
+      await p.remove(_oldKey);
+    } on Exception catch (e) {
+      debugPrint('ChatHistoryService.clear prefs error: $e');
+    }
   }
 
   /// Builds a plain-text export of all conversations (no file write).
diff --git a/lib/services/memory_diagnostics_service.dart b/lib/services/memory_diagnostics_service.dart
index 7a4b914..d1bc064 100644
--- a/lib/services/memory_diagnostics_service.dart
+++ b/lib/services/memory_diagnostics_service.dart
@@ -15,12 +15,31 @@ class MemoryDiagnosticsService {
   int? _lastPssKb;
   int? _lastRssKb;
   int? _lastAvailMemMb;
+  int? _lastTotalMemMb;
 
   int? get lastPssMb =>
       _lastPssKb != null ? (_lastPssKb! / 1024).round() : null;
   int? get lastRssMb =>
       _lastRssKb != null ? (_lastRssKb! / 1024).round() : null;
   int? get lastAvailMemMb => _lastAvailMemMb;
+  int? get lastTotalMemMb => _lastTotalMemMb;
+
+  void _ingestMemoryMap(Map<Object?, Object?> result) {
+    _lastPssKb = result['pssKb'] as int?;
+    _lastRssKb = result['rssKb'] as int?;
+    final avail = result['availMemMb'];
+    if (avail is int) {
+      _lastAvailMemMb = avail;
+    } else if (avail is num) {
+      _lastAvailMemMb = avail.round();
+    }
+    final total = result['totalMemMb'];
+    if (total is int) {
+      _lastTotalMemMb = total;
+    } else if (total is num) {
+      _lastTotalMemMb = total.round();
+    }
+  }
 
   /// Best-effort free system RAM in MB (Android ActivityManager).
   Future<int?> readAvailableMemMb() async {
@@ -30,19 +49,9 @@ class MemoryDiagnosticsService {
           'getProcessMemory',
         );
         if (result != null) {
-          _lastPssKb = result['pssKb'] as int?;
-          _lastRssKb = result['rssKb'] as int?;
-          final avail = result['availMemMb'];
-          if (avail is int) {
-            _lastAvailMemMb = avail;
+          _ingestMemoryMap(result);
 
-            return avail;
-          }
-          if (avail is num) {
-            _lastAvailMemMb = avail.round();
-
-            return _lastAvailMemMb;
-          }
+          return _lastAvailMemMb;
         }
       } on PlatformException catch (e) {
         debugPrint('MemoryDiagnosticsService availMem: $e');
@@ -52,6 +61,14 @@ class MemoryDiagnosticsService {
     return _lastAvailMemMb;
   }
 
+  /// Best-effort total system RAM in MB (Android ActivityManager).
+  Future<int?> readTotalMemMb() async {
+    if (_lastTotalMemMb != null) return _lastTotalMemMb;
+    await readAvailableMemMb();
+
+    return _lastTotalMemMb;
+  }
+
   /// Best-effort process memory in MB (PSS preferred on Android).
   Future<int?> readProcessMemoryMb() async {
     if (!kIsWeb && Platform.isAndroid) {
@@ -60,14 +77,7 @@ class MemoryDiagnosticsService {
           'getProcessMemory',
         );
         if (result != null) {
-          _lastPssKb = result['pssKb'] as int?;
-          _lastRssKb = result['rssKb'] as int?;
-          final avail = result['availMemMb'];
-          if (avail is int) {
-            _lastAvailMemMb = avail;
-          } else if (avail is num) {
-            _lastAvailMemMb = avail.round();
-          }
+          _ingestMemoryMap(result);
           final pss = _lastPssKb;
           if (pss != null) return (pss / 1024).round();
         }
diff --git a/lib/services/model_manager.dart b/lib/services/model_manager.dart
index accff90..faab7fb 100644
--- a/lib/services/model_manager.dart
+++ b/lib/services/model_manager.dart
@@ -53,6 +53,7 @@ class ModelManager {
   ModelManager._();
 
   static const _prefsKey = 'installed_models';
+  Future<void>? _installLock;
   static const _customModelsPrefsKey = 'custom_models';
   static const _hfTokenKey = 'hf_token';
   final List<InstalledModel> _installedModels = [];
@@ -114,6 +115,34 @@ class ModelManager {
     required ModelType modelType,
     required ModelFileType fileType,
     void Function(int progress)? onProgress,
+  }) async {
+    // Serialize installs — concurrent Gemma 4 onboarding + Gemma 3 fallback
+    // races corrupt flutter_gemma's active identity (POCO F1).
+    while (_installLock != null) {
+      await _installLock;
+    }
+    final gate = Completer<void>();
+    _installLock = gate.future;
+    try {
+      return await _installFromNetworkUnlocked(
+        url: url,
+        modelType: modelType,
+        fileType: fileType,
+        onProgress: onProgress,
+      );
+    } finally {
+      gate.complete();
+      if (identical(_installLock, gate.future)) {
+        _installLock = null;
+      }
+    }
+  }
+
+  Future<InstalledModel?> _installFromNetworkUnlocked({
+    required String url,
+    required ModelType modelType,
+    required ModelFileType fileType,
+    void Function(int progress)? onProgress,
   }) async {
     try {
       final uri = Uri.parse(url);
@@ -125,6 +154,7 @@ class ModelManager {
         location: 'model_manager.dart:installFromNetwork:start',
         message: 'installFromNetwork started',
         data: {'url': url, 'fileName': fileName, 'modelType': modelType.name},
+        runId: 'post-fix',
       );
       // #endregion
 
@@ -161,6 +191,7 @@ class ModelManager {
             'canonicalName': canonicalName,
             'fileOnDiskExists': await fileOnDisk.exists(),
           },
+          runId: 'post-fix',
         );
         // #endregion
 
@@ -211,6 +242,21 @@ class ModelManager {
         }
         final response = await request.close();
 
+        // #region agent log
+        await AgentDebugLog.log(
+          hypothesisId: 'F',
+          location: 'model_manager.dart:installFromNetwork:http',
+          message: 'Download HTTP response',
+          data: {
+            'fileName': fileName,
+            'statusCode': response.statusCode,
+            'contentLength': response.contentLength,
+            'hasToken': hfToken != null && hfToken.isNotEmpty,
+          },
+          runId: 'post-fix',
+        );
+        // #endregion
+
         if (response.statusCode == 401 || response.statusCode == 403) {
           _statusController.add(
             'Download failed: HuggingFace auth required. '
@@ -242,6 +288,19 @@ class ModelManager {
           _statusController.add(
             'Download incomplete: expected $totalBytes bytes, got $receivedBytes',
           );
+          // #region agent log
+          await AgentDebugLog.log(
+            hypothesisId: 'F',
+            location: 'model_manager.dart:installFromNetwork:incomplete',
+            message: 'Download incomplete',
+            data: {
+              'fileName': fileName,
+              'expected': totalBytes,
+              'received': receivedBytes,
+            },
+            runId: 'post-fix',
+          );
+          // #endregion
           try {
             await tempFile.delete();
           } catch (e) {
@@ -285,6 +344,7 @@ class ModelManager {
           'prefsCount': _installedModels.length,
           'prefsNames': _installedModels.map((m) => m.fileName).toList(),
         },
+        runId: 'post-fix',
       );
       // #endregion
 
@@ -303,6 +363,15 @@ class ModelManager {
     } catch (e) {
       _statusController.add('Install failed: $e');
       debugPrint('ModelManager: installFromNetwork failed: $e');
+      // #region agent log
+      await AgentDebugLog.log(
+        hypothesisId: 'F',
+        location: 'model_manager.dart:installFromNetwork:error',
+        message: 'installFromNetwork exception',
+        data: {'error': e.toString()},
+        runId: 'post-fix',
+      );
+      // #endregion
       return null;
     }
   }
diff --git a/lib/services/model_orchestrator.dart b/lib/services/model_orchestrator.dart
index 78d8b35..620931d 100644
--- a/lib/services/model_orchestrator.dart
+++ b/lib/services/model_orchestrator.dart
@@ -774,13 +774,55 @@ class ModelOrchestrator {
     }
 
     // Free-RAM hard gate before a cold load of heavy models (Android).
-    final ramBlock = await PlatformAdaptationService.instance.checkCanLoadModel(
-      model,
+    // If blocked, auto-fallback to a smaller installed-capable model instead of
+    // failing the whole turn (POCO F1 / ~1.2GB free cannot load Gemma 4).
+    var modelToLoad = model;
+    var ramBlock = await PlatformAdaptationService.instance.checkCanLoadModel(
+      modelToLoad,
+    );
+    // #region agent log
+    await AgentDebugLog.log(
+      hypothesisId: 'C',
+      location: 'model_orchestrator.dart:_getOrCreateModel:ramGate',
+      message: 'RAM gate decision',
+      data: {
+        'requested': model.displayName,
+        'blocked': ramBlock != null,
+        'blockMsg': ramBlock,
+        'availMemMb': MemoryDiagnosticsService.instance.lastAvailMemMb,
+      },
     );
+    // #endregion
+    if (ramBlock != null) {
+      final fallback = await _pickRamFallback(modelToLoad);
+      if (fallback != null) {
+        // #region agent log
+        await AgentDebugLog.log(
+          hypothesisId: 'C',
+          location: 'model_orchestrator.dart:_getOrCreateModel:fallback',
+          message: 'RAM gate fallback',
+          data: {
+            'from': modelToLoad.displayName,
+            'to': fallback.displayName,
+            'fileName': ModelHuggingFaceURLs.fileNameFor(fallback),
+            'onDisk': await ModelManager.instance.isInstalledOnDisk(
+              ModelHuggingFaceURLs.fileNameFor(fallback),
+            ),
+          },
+        );
+        // #endregion
+        _statusController.add(
+          'Low free RAM — using ${fallback.displayName} instead of '
+          '${modelToLoad.displayName}',
+        );
+        modelToLoad = fallback;
+        ramBlock = null;
+      }
+    }
     if (ramBlock != null) {
       throw ModelException(
         ramBlock,
-        model: model,
+        model: modelToLoad,
         suggestion:
             'Free RAM by closing apps, or use a smaller model in Settings.',
       );
@@ -810,18 +852,18 @@ class ModelOrchestrator {
       }
     }
 
-    final bool supportImage = needsImageSupport;
+    final bool supportImage = modelToLoad.hasVision;
 
     _isLoadingModel = true;
     _statusController.add(
-      supportImage && model == NovaModel.gemma4E2b
-          ? 'Compiling ${model.displayName} on GPU (vision enabled; first load may take 1–2 min)...'
-          : 'Loading ${model.displayName}...',
+      supportImage && modelToLoad == NovaModel.gemma4E2b
+          ? 'Compiling ${modelToLoad.displayName} on GPU (vision enabled; first load may take 1–2 min)...'
+          : 'Loading ${modelToLoad.displayName}...',
     );
 
     try {
       // First, ensure the model is registered with flutter_gemma
-      final fileName = ModelHuggingFaceURLs.fileNameFor(model);
+      final fileName = ModelHuggingFaceURLs.fileNameFor(modelToLoad);
       final existsOnDisk = await ModelManager.instance.isInstalledOnDisk(
         fileName,
       );
@@ -833,7 +875,7 @@ class ModelOrchestrator {
         location: 'model_orchestrator.dart:_getOrCreateModel:check',
         message: 'Model availability check before load',
         data: {
-          'model': model.name,
+          'model': modelToLoad.name,
           'fileName': fileName,
           'existsOnDisk': existsOnDisk,
           'prefsInstalled': prefsInstalled,
@@ -843,7 +885,7 @@ class ModelOrchestrator {
 
       if (existsOnDisk) {
         _statusController.add(
-          'Found ${model.displayName} on disk, registering...',
+          'Found ${modelToLoad.displayName} on disk, registering...',
         );
         try {
           final modelPath = await _findModelPath(fileName);
@@ -853,29 +895,29 @@ class ModelOrchestrator {
             await ModelManager.instance.registerDiskModel(
               filePath: modelPath,
               fileName: fileName,
-              modelType: model.modelType,
-              fileType: model.fileType,
+              modelType: modelToLoad.modelType,
+              fileType: modelToLoad.fileType,
               fileSizeBytes: fileSize,
             );
 
             // Now get the active model
-            _statusController.add('Loading ${model.displayName}...');
+            _statusController.add('Loading ${modelToLoad.displayName}...');
             if (_debugMode) {
               unawaited(
                 MemoryDiagnosticsService.instance.readProcessMemoryMb(),
               );
             }
             _activeModel = await FlutterGemma.getActiveModel(
-              maxTokens: _tokenLimitFor(model),
+              maxTokens: _tokenLimitFor(modelToLoad),
               preferredBackend: PlatformAdaptationService.instance
-                  .preferredBackendFor(model),
+                  .preferredBackendFor(modelToLoad),
               supportImage: supportImage,
               maxNumImages: supportImage ? 1 : null,
-            ).timeout(_loadTimeoutFor(model));
-            _activeModelType = model;
+            ).timeout(_loadTimeoutFor(modelToLoad));
+            _activeModelType = modelToLoad;
             _activeModelSupportsImage = supportImage;
             _isInitialized = true;
-            _statusController.add('${model.displayName} ready');
+            _statusController.add('${modelToLoad.displayName} ready');
             _resetIdleTimer();
             return _activeModel!;
           }
@@ -883,8 +925,8 @@ class ModelOrchestrator {
           debugPrint('Failed to load local model: $e');
           await _teardownPartialLoad();
           throw ModelLoadException(
-            '${model.displayName} is still loading.',
-            model: model,
+            '${modelToLoad.displayName} is still loading.',
+            model: modelToLoad,
             suggestion:
                 'First boot GPU compile can take 1–2 minutes. Wait and try again, '
                 'or use Settings → Debug → Reset inference engine.',
@@ -898,16 +940,16 @@ class ModelOrchestrator {
           if (e is TimeoutException) {
             await _teardownPartialLoad();
             throw ModelLoadException(
-              '${model.displayName} is still loading.',
-              model: model,
+              '${modelToLoad.displayName} is still loading.',
+              model: modelToLoad,
               suggestion: 'First boot GPU compile can take 1–2 minutes. Wait and try again.',
               underlyingError: e,
             );
           }
           await _teardownPartialLoad();
           throw ModelLoadException(
-            'Failed to load ${model.displayName} from disk.',
-            model: model,
+            'Failed to load ${modelToLoad.displayName} from disk.',
+            model: modelToLoad,
             suggestion:
                 'The model file may be corrupted. '
                 'Try re-downloading, reset inference in Settings, or pick a different file.',
@@ -917,13 +959,13 @@ class ModelOrchestrator {
       }
 
       // No active model or load failed — ask user before downloading
-      final url = ModelHuggingFaceURLs.urlFor(model);
-      final choice = await _showDownloadConsent(model: model, url: url);
+      final url = ModelHuggingFaceURLs.urlFor(modelToLoad);
+      final choice = await _showDownloadConsent(model: modelToLoad, url: url);
 
       if (choice == DownloadConsent.pickFile) {
         // User wants to pick a file from device — throw a special exception
         // so the UI can handle the file picker flow.
-        throw ModelNeedsFilePickException(model);
+        throw ModelNeedsFilePickException(modelToLoad);
       }
 
       if (choice == DownloadConsent.cancel) {
@@ -931,16 +973,16 @@ class ModelOrchestrator {
       }
 
       // choice == DownloadConsent.download — proceed with network install
-      _statusController.add('Downloading ${model.displayName}...');
+      _statusController.add('Downloading ${modelToLoad.displayName}...');
       try {
         final installed = await ModelManager.instance
             .installFromNetwork(
               url: url,
-              modelType: model.modelType,
-              fileType: model.fileType,
+              modelType: modelToLoad.modelType,
+              fileType: modelToLoad.fileType,
               onProgress: (progress) {
                 _statusController.add(
-                  'Downloading ${model.displayName}: $progress%',
+                  'Downloading ${modelToLoad.displayName}: $progress%',
                 );
               },
             )
@@ -948,25 +990,25 @@ class ModelOrchestrator {
 
         if (installed == null) {
           throw ModelDownloadException(
-            'Download of ${model.displayName} failed.',
-            model: model,
+            'Download of ${modelToLoad.displayName} failed.',
+            model: modelToLoad,
             suggestion: 'Check your internet connection and try again.',
           );
         }
 
         // Now get the model after install
-        _statusController.add('Loading ${model.displayName}...');
+        _statusController.add('Loading ${modelToLoad.displayName}...');
         _activeModel = await FlutterGemma.getActiveModel(
-          maxTokens: _tokenLimitFor(model),
+          maxTokens: _tokenLimitFor(modelToLoad),
           preferredBackend: PlatformAdaptationService.instance
-              .preferredBackendFor(model),
+              .preferredBackendFor(modelToLoad),
           supportImage: supportImage,
           maxNumImages: supportImage ? 1 : null,
-        ).timeout(_loadTimeoutFor(model));
-        _activeModelType = model;
+        ).timeout(_loadTimeoutFor(modelToLoad));
+        _activeModelType = modelToLoad;
         _activeModelSupportsImage = supportImage;
         _isInitialized = true;
-        _statusController.add('${model.displayName} ready');
+        _statusController.add('${modelToLoad.displayName} ready');
         _resetIdleTimer();
         return _activeModel!;
       } on ModelException {
@@ -974,21 +1016,21 @@ class ModelOrchestrator {
       } on TimeoutException catch (e) {
         await _teardownPartialLoad();
         throw ModelLoadException(
-          '${model.displayName} is still loading.',
-          model: model,
+          '${modelToLoad.displayName} is still loading.',
+          model: modelToLoad,
           suggestion:
               'First boot GPU compile can take 1–2 minutes. Wait and try again, '
               'or use Settings → Debug → Reset inference engine.',
           underlyingError: e,
         );
       } catch (e) {
-        _statusController.add('Failed to load ${model.displayName}: $e');
+        _statusController.add('Failed to load ${modelToLoad.displayName}: $e');
         // Classify the error
         final msg = e.toString().toLowerCase();
         if (msg.contains('storage') || msg.contains('no space')) {
           throw ModelStorageException(
-            'Not enough storage for ${model.displayName}.',
-            model: model,
+            'Not enough storage for ${modelToLoad.displayName}.',
+            model: modelToLoad,
             suggestion: 'Free up storage space and try again.',
             underlyingError: e,
           );
@@ -997,8 +1039,8 @@ class ModelOrchestrator {
             msg.contains('invalid') ||
             msg.contains('unsupported')) {
           throw ModelLoadException(
-            '${model.displayName} file is corrupted or incompatible.',
-            model: model,
+            '${modelToLoad.displayName} file is corrupted or incompatible.',
+            model: modelToLoad,
             suggestion: 'Re-download the model or pick a different file.',
             underlyingError: e,
           );
@@ -1007,15 +1049,15 @@ class ModelOrchestrator {
             msg.contains('connection') ||
             msg.contains('socket')) {
           throw ModelDownloadException(
-            'Network error downloading ${model.displayName}.',
-            model: model,
+            'Network error downloading ${modelToLoad.displayName}.',
+            model: modelToLoad,
             suggestion: 'Check your internet connection and try again.',
             underlyingError: e,
           );
         }
         throw ModelLoadException(
-          'Failed to load ${model.displayName}.',
-          model: model,
+          'Failed to load ${modelToLoad.displayName}.',
+          model: modelToLoad,
           suggestion: 'Try again or pick a file from your device.',
           underlyingError: e,
         );
@@ -1025,6 +1067,54 @@ class ModelOrchestrator {
     }
   }
 
+  /// Pick a smaller model that fits free RAM.
+  ///
+  /// Prefers models already on disk (no download), then SmolLM (tiny download)
+  /// before Gemma 3 1B. Avoids falling back to a missing mid-size model while
+  /// the device only has Gemma 4 installed (POCO F1).
+  Future<NovaModel?> _pickRamFallback(NovaModel blocked) async {
+    const candidates = <NovaModel>[NovaModel.smollm, NovaModel.gemma3_1b];
+    NovaModel? firstDownloadable;
+
+    for (final candidate in candidates) {
+      if (candidate == blocked) continue;
+      final block = await PlatformAdaptationService.instance.checkCanLoadModel(
+        candidate,
+      );
+      if (block != null) continue;
+
+      final fileName = ModelHuggingFaceURLs.fileNameFor(candidate);
+      final onDisk = await ModelManager.instance.isInstalledOnDisk(fileName);
+      if (onDisk) {
+        return candidate;
+      }
+      firstDownloadable ??= candidate;
+    }
+
+    return firstDownloadable;
+  }
+
+  /// Apply Auto-mode defaults for low free-RAM devices (call at startup).
+  Future<void> applyRamAwareModelDefaults() async {
+    final recommended = await PlatformAdaptationService.instance
+        .recommendModelForDevice();
+    if (selector.primaryHeavy != recommended) {
+      selector.primaryHeavy = recommended;
+      // #region agent log
+      await AgentDebugLog.log(
+        hypothesisId: 'E',
+        location: 'model_orchestrator.dart:applyRamAwareModelDefaults',
+        message: 'Adjusted Auto primaryHeavy for free RAM',
+        data: {
+          'primaryHeavy': recommended.displayName,
+          'availMemMb': MemoryDiagnosticsService.instance.lastAvailMemMb,
+        },
+        runId: 'post-fix',
+      );
+      // #endregion
+    }
+  }
+
   /// Process a message using a custom (user-imported) model.
   Stream<InferenceResult> _processWithCustomModel({
     required String query,
diff --git a/lib/services/platform_adaptation_service.dart b/lib/services/platform_adaptation_service.dart
index eac77ca..42f8bc4 100644
--- a/lib/services/platform_adaptation_service.dart
+++ b/lib/services/platform_adaptation_service.dart
@@ -297,12 +297,42 @@ class PlatformAdaptationService {
     if (kIsWeb) return null;
     if (defaultTargetPlatform != TargetPlatform.android) return null;
 
-    if (model.sizeMB >= 2000) return 3500;
-    if (model.sizeMB >= 400) return 1200;
+    // Softened for 6GB devices (POCO F1 etc.): hard-block only when truly
+    // starved. Prefer auto-fallback in the orchestrator over refusing chat.
+    if (model.sizeMB >= 2000) return 1800;
+    if (model.sizeMB >= 400) return 700;
 
     return null;
   }
 
+  /// Onboarding / default model for device RAM class.
+  ///
+  /// Free RAM alone is unreliable: a wiped POCO F1 (6 GB total) can report
+  /// >1800 MB free at boot and wrongly recommend Gemma 4, which then fails to
+  /// load once the system fills. Prefer **total RAM** for onboarding.
+  Future<NovaModel> recommendModelForDevice() async {
+    if (kIsWeb) return NovaModel.smollm;
+    if (defaultTargetPlatform != TargetPlatform.android) {
+      return NovaModel.gemma4E2b;
+    }
+
+    final mem = MemoryDiagnosticsService.instance;
+    final avail = await mem.readAvailableMemMb();
+    final total = await mem.readTotalMemMb();
+
+    // ≤6.5 GB class (POCO F1 6 GB, etc.): never onboard Gemma 4.
+    // Fresh reboot can show plenty of free RAM and still OOM on load.
+    if (total != null && total < 6656) {
+      return NovaModel.smollm;
+    }
+
+    if (avail == null) return NovaModel.gemma3_1b;
+    if (avail < 900) return NovaModel.smollm;
+    if (avail < 1800) return NovaModel.gemma3_1b;
+
+    return NovaModel.gemma4E2b;
+  }
+
   /// Pure gate check: returns an error message when [availMemMb] is too low.
   String? freeRamGateMessage({
     required NovaModel model,
diff --git a/lib/utils/agent_debug_log.dart b/lib/utils/agent_debug_log.dart
index 13ed091..04e2f7a 100644
--- a/lib/utils/agent_debug_log.dart
+++ b/lib/utils/agent_debug_log.dart
@@ -3,12 +3,12 @@ import 'dart:io';
 
 import 'package:flutter/foundation.dart';
 
-/// Debug-mode NDJSON logger for session 36a30e.
+/// Debug-mode NDJSON logger for session 7abc09.
 /// Posts to the local ingest server (use `adb reverse tcp:7704 tcp:7704`).
 class AgentDebugLog {
   static const _endpoint =
       'http://127.0.0.1:7704/ingest/b12c34e9-d2e1-4829-bc46-cbd81f2bb686';
-  static const _sessionId = '36a30e';
+  static const _sessionId = '7abc09';
 
   static Future<void> log({
     required String hypothesisId,
diff --git a/pubspec.yaml b/pubspec.yaml
index 8b2119a..1fac20a 100644
--- a/pubspec.yaml
+++ b/pubspec.yaml
@@ -64,6 +64,8 @@ dev_dependencies:
   build_runner: ^2.4.13
   dart_code_linter: ^4.1.6
   coverage: ^1.7.0
+  path_provider_platform_interface: ^2.1.2
+  plugin_platform_interface: ^2.1.8
 
 flutter:
   uses-material-design: true
\ No newline at end of file
diff --git a/test/services/chat_history_export_test.dart b/test/services/chat_history_export_test.dart
index edfc01b..7b06fca 100644
--- a/test/services/chat_history_export_test.dart
+++ b/test/services/chat_history_export_test.dart
@@ -1,21 +1,47 @@
+import 'dart:io';
+import 'dart:typed_data';
+
 import 'package:flutter_test/flutter_test.dart';
 import 'package:nova_assistant/models/chat_message.dart';
 import 'package:nova_assistant/models/conversation.dart';
 import 'package:nova_assistant/services/chat_history_service.dart';
+import 'package:path_provider_platform_interface/path_provider_platform_interface.dart';
+import 'package:plugin_platform_interface/plugin_platform_interface.dart';
 import 'package:shared_preferences/shared_preferences.dart';
 
+class _FakePathProvider extends Fake
+    with MockPlatformInterfaceMixin
+    implements PathProviderPlatform {
+  _FakePathProvider(this.root);
+  final Directory root;
+
+  @override
+  Future<String?> getApplicationDocumentsPath() async => root.path;
+}
+
 void main() {
   TestWidgetsFlutterBinding.ensureInitialized();
 
+  late Directory tempDir;
+
   setUp(() async {
     SharedPreferences.setMockInitialValues({});
+    tempDir = await Directory.systemTemp.createTemp('nova_chat_hist_');
+    PathProviderPlatform.instance = _FakePathProvider(tempDir);
     await ChatHistoryService.clear();
   });
 
-  test('exportAsText returns content without writing a file path', () async {
+  tearDown(() async {
+    await ChatHistoryService.clear();
+    if (await tempDir.exists()) {
+      await tempDir.delete(recursive: true);
+    }
+  });
+
+  test('persists conversations to a file, not SharedPreferences', () async {
     final convo = Conversation(
       id: 'c1',
-      title: 'Test chat',
+      title: 'Test',
       createdAt: DateTime(2026, 1, 1),
       updatedAt: DateTime(2026, 1, 1),
       messages: [
@@ -29,11 +55,37 @@ void main() {
     );
     await ChatHistoryService.saveConversations([convo]);
 
-    final content = await ChatHistoryService.exportAsText();
-    expect(content, isNotNull);
-    expect(content!, contains('Hello'));
-    expect(content, contains('Nova Assistant'));
-    expect(content.contains('/data/user/'), isFalse);
-    expect(content.endsWith('.txt'), isFalse);
+    final prefs = await SharedPreferences.getInstance();
+    expect(prefs.getString('conversations'), isNull);
+
+    final file = File('${tempDir.path}/conversations.json');
+    expect(await file.exists(), isTrue);
+    expect(await file.readAsString(), contains('Hello'));
+
+    final loaded = await ChatHistoryService.loadConversations();
+    expect(loaded, hasLength(1));
+    expect(loaded.first.messages.first.text, 'Hello');
+  });
+
+  test('does not persist screenshot bytes', () async {
+    final convo = Conversation(
+      id: 'c2',
+      messages: [
+        ChatMessage(
+          id: 'm2',
+          text: '',
+          isUser: true,
+          timestamp: DateTime(2026, 1, 1),
+          imageData: Uint8List(1024)..fillRange(0, 1024, 7),
+        ),
+      ],
+    );
+    await ChatHistoryService.saveConversations([convo]);
+
+    final file = File('${tempDir.path}/conversations.json');
+    final raw = await file.readAsString();
+    expect(raw, contains('"imageData":null'));
+    expect(raw, contains('[Screenshot]'));
+    expect(raw.length, lessThan(2000));
   });
 }
diff --git a/test/services/free_ram_gate_test.dart b/test/services/free_ram_gate_test.dart
index 4dd5240..f3276c5 100644
--- a/test/services/free_ram_gate_test.dart
+++ b/test/services/free_ram_gate_test.dart
@@ -13,11 +13,11 @@ void main() {
 
       final message = service.freeRamGateMessage(
         model: NovaModel.gemma4E2b,
-        availMemMb: 2000,
+        availMemMb: 1000,
       );
       expect(message, isNotNull);
       expect(message!, contains('Gemma 4 E2B'));
-      expect(message, contains('2000 MB'));
+      expect(message, contains('1000 MB'));
     });
 
     test('allows Gemma 4 when enough free RAM', () {
@@ -27,7 +27,7 @@ void main() {
       expect(
         service.freeRamGateMessage(
           model: NovaModel.gemma4E2b,
-          availMemMb: 4000,
+          availMemMb: 2000,
         ),
         isNull,
       );
@@ -55,5 +55,14 @@ void main() {
         isNull,
       );
     });
+
+    test('minFreeRamMbFor matches size tiers', () {
+      debugDefaultTargetPlatformOverride = TargetPlatform.android;
+      addTearDown(() => debugDefaultTargetPlatformOverride = null);
+
+      expect(service.minFreeRamMbFor(NovaModel.gemma4E2b), 1800);
+      expect(service.minFreeRamMbFor(NovaModel.gemma3_1b), 700);
+      expect(service.minFreeRamMbFor(NovaModel.smollm), isNull);
+    });
   });
 }