- fixed a bug where Kokoro TTS was crashing - fixed a bug where Chat Debug Info cannot be scrolled - fixed a bug where Long-term memory knowledge graphs were not being displayed on iOS

Layla 7.0 adds the Layla SDK and user-made mini-apps, rebuilds long-term memory around a new on-device pipeline, and brings a new set of voices. This release also includes every fix from 7.0.1 through 7.0.4. NEW - Layla SDK: build your own mini-apps as a single HTML file, a zip folder, or a self-hosted site. Mini-apps can use the LLM, characters, image generation, agents, memory and voice. - Browse Apps: install mini-apps made by other users, upload your own to Layla Cloud, manage versions in the Mini-Apps Manager, and share mini-app links. - Agent skill for Codex, Claude Code and OpenCode, plus a mini-app template for larger projects. - Real-time long-term memory, processed as you chat rather than ingested overnight in the background. - New default memory pipeline using a custom-trained model that builds summaries and knowledge graphs together. - Temporal knowledge graphs, with a full-screen view. - Dream rebuilt as a downloadable mini-app, with evolving character impressions, self-reflection and smarter conversation continuation. - Kokoro v1.0 voices, including multilingual options. - File Manager can import arbitrary files, including assets and draft models, and now shows image previews. - Expanded SDK APIs: chat sessions and history, sentiment detection, file save and read, image-to-image, TTS voice listing and generation, message scheduling, persona lookup, top-memory recall, and getting or setting inference engines. IMPROVED - Deleting or regenerating a message now updates or removes the memory it created. - Memory recall skips the most recent same-session memories to avoid feedback loops. - A mini-app can replace the default memory processor through the SDK. - SDK chat responses return reasoning and message content in separate fields. - Faster character selection in very large libraries. - Faster Browse Apps loading, with mini-app icons cached for offline use. - Better handling of large mini-apps that bundle image and audio into their HTML. - File Manager shows total folder size and is faster and more stable with many files. - Backups now include character folders and sorting. - Reference documents and agents are easier to select during character creation. - Sentiment detection aggregates across sentences and normalises pronouns. - Importing a mini-app that already exists asks for confirmation and overwrites it instead of blocking. - Benchmark mini-app redesigned to match the current visual style. FIXED - Long-term memory generation getting stuck while building knowledge graphs. - Knowledge graphs opening to a blank page. - Long-term memory recording messages for characters with LTM turned off. - Removing long-term memory with the References mini-app installed failing with an unknown error. - Scheduled future messages appearing on the chat history page. - Dreams skipping scheduling when using cloud models. - Image-to-image not applying the prompt prefix. - Image generation starting while the model was still reasoning about the image prompt. - Mini-app image generation reading default settings instead of saved settings. - Custom sampler presets not saving additional llama.cpp arguments. - Speak Responses reporting that TTS was not initialised. - Layla Cloud voices not stopping when requested, and the SDK stop-speaking call doing nothing. - PocketTTS not working through the SDK. - Voice chat minimising text with the setting turned off. - Mini-apps sending a system prompt Layla did not use. - Chat history appearing when swiping left in a mini-app. - Importing a Roleplay from a RoleplayPNG not saving its characters. - Roleplay not loading long conversation histories properly. - Auto-translated messages not showing. - Prompts being processed again unnecessarily. - A crash when loading a very large conversation history. Layla runs on your device: no account, no subscription, no data leaving the phone.

The Roleplay mini-app has been completely redesigned! Choose from 50+ templates to start your roleplay sessions: romance, adventure, isekai, comedy, interview practice, language learning, and more! - Roleplay now supports local models (GGUF), connecting to your PC/OpenAI, Layla Server, or Layla Cloud. - edit/create your own scenarios with an all new editor - create temporary characters directly within the Roleplay mini-app, or attach your existing characters to the scenario. (Your temporary characters will be automatically put into a folder with the name of your Roleplay in the character list) - Visual Novel mode: centres the character image as they speak/act - OOC Instructions: you can add out-of-character instructions at any time during the roleplay, useful to keep the AI on track - TTS: text-to-speech for individual characters in their own voices (configuration in the character edit) - Roleplay supports custom agents and image generation - Roleplay automatically switches LLMs based on your saved inference engines depending on who is talking, making it finally possible to have multiple LLMs chat with each other and see how they interact! Improvements: - improved long-term memory summarisation speed - added Gemma4 (thinking) preset prompt - redesigned the Saved Inference Engine section to easier filter by attached characters - improved TTS consistency by removing unnecessary stopping mid-sentences - updated Local Dream to add image editing features - added new schedulers (Kerras versions) in Local Dream Bug fixes: - fixed a bug where the model selector disappears after searching for available models when configuring OpenAI API - fixed a bug where chat bubble styles stops showing when performance counter is enabled - fixed a bug where changing the prompt but not the seed in image generation does not generate a new image - fixed a bug where "Speak Responses" in chat actions cannot be toggled off - fixed a bug where "Push to talk" had stopped working

- fixed a bug where custom vision models were not being detected correctly - fixed a bug where translation quick actions was not available - fixed a bug where Stable Diffusion models were stuck on generating

Layla supports one-shot Voice Cloning with PocketTTS mini-app! Create your own voices by recording or uploading an audio sample. Layla will create a custom voice out of it for which you can use for your characters! You can also set your custom voice as the the global TTS voice for Layla, or configure it individually for each character. Record audio right in Layla to create custom voices: you can record an audio snippet and use it as a base to clone a voice Voice Cloning tips: - PocketTTS is very sensitive to background noise. Record your audio in a quiet room or use an online editor to remove background noise to get better results - only supports 16-bit PCM wav files, use an online converter if you need to convert mp3's etc. - try not to have blanks/silences in the sample, PocketTTS will try to learn that and generate lots of silences - exaggerated voices work better New features: - added ability to configure chat-bubble colours in UI settings Improvements: - gallery of generated images in Layla are now backed up - flipped the user/AI position of chat bubble styles to match the standard for other social media/chat apps - image gallery now blocks other apps from taking a screenshot or recording of it's contents, and will also automatically hide its contents when Layla goes into the background - image gallery in Layla allows multi-selection for deleting images - improved TTS latency by implementing audio streaming - you can re-order characters inside folders and drag them outside to remove them from the folder - improved character impression generation by limiting the max output and cutting off incomplete sentences - Stable Diffusion prompts are saved in generated images gallery Bug fixes: - fixed a bug where Layla Cloud image generation fails to read the generated images after a few attempts - fixed a bug where default Automatic1111 json body stopped working - fixed a bug where setting a max response length does not trim incomplete sentences - fixed a bug where previewing native TTS using Google/iOS was not using the selected voice properly in other languages - fixed a bug where sometimes asterisks were not being skipped in TTS when skip asterisk setting on

Improvements: - advanced settings allows adjusting LLM and VLM hardware acceleration separately, allowing you to put the LLM on the CPU while the VLM on the GPU - supports Gemma-4 tool calling agents in both GGUF - allow setting additional llama.cpp command-line arguments in Advanced Settings - additional llama.cpp cmd arguments can be saved as Advanced Settings presets Bug fixes: - fixed a few "Failed to eval" bugs on long context usage - fixed a few UI issues that happens when the screen is horizontal - fixed a bug where viewing a memory that is too long will soft-lock you without any way to exit the chat - LTM now removes think tags from Gemma 4 if used as summariser

New features: - added new Choose Your Own Adventure characters - supports the latest Gemma-4 model family - added German translations Improvements: - added ability to send progress messages during python execution, you can view this progress in the Agent execution card to give you more insight into what the agent is doing - Layla Python can now generate files - new Python agents will automatically install needed dependencies when added for the first time - you can configure long-term memory to use a saved inference engine as summariser - added UI setting to display chats in a bubble-like interface instead (similar to Facebook Messenger or WhatApp etc.) - preset characters are now organised into folders Bug fixes: - fixed a bug where sometimes character expressions will not be saved - fixed a bug where some memories cannot be processed

New features: - added new Choose Your Own Adventure characters - supports the latest Gemma-4 model family - added German translations Improvements: - added ability to send progress messages during python execution, you can view this progress in the Agent execution card to give you more insight into what the agent is doing - Layla Python can now generate files - new Python agents will automatically install needed dependencies when added for the first time - you can configure long-term memory to use a saved inference engine as summariser - added UI setting to display chats in a bubble-like interface instead (similar to Facebook Messenger or WhatApp etc.) - preset characters are now organised into folders Bug fixes: - fixed a bug where sometimes character expressions will not be saved - fixed a bug where some memories cannot be processed

New features: This update fleshes out Layla's Agent Framework. **Agent System Prompts** Agents now have their own system prompts. These will update your character's system prompt when they are activated. For example, you don't need to update both your system prompt and the agent for status cards; the system prompt to guide the model into generating status cards will automatically be injected when you attach the corresponding agent. **Introducing Short-term Memory** Why do we need a short-term memory when there is the LLM context itself? The main difference lies in the fact that Layla's short-term memory is **structured**. Both the LLM context and LTM is unstructured, and lookup is done purely via heuristics (via the Attention layer in LLMs, and embeddings in LTM). The short-term memory is structured a layer that lives between the LLM's context and Layla's LTM. Information flows from the LLM's context -> short-term memory -> LTM (as demonstrated by one of the example agents below). This can be updated by both the LLM and Layla's Agents. Because short-term memory is structured, you can use code and logic to manipulate the contents, not just via prompts. This opens up the possibility to *program* information flow. An example could be writing an agent that updates a character's Hit-Points or inventory based on triggering conditions. This layer will also serve as a sort of *RAM* for all agents. Any agent can read/write into Layla's short-term memory, allowing a horizontal layer of communication between agents (work-in-progress). This is now possible because... **Layla supports Python** Layla now comes with her own embedded Python interpreter as one of the mini-apps. Install it to enable writing and executing arbitrary python scripts directly in your phone! You can install packages via `pip`, and Layla comes with a small console and code area to run your scripts. All agents in Layla can now execute python scripts. Which means you can program arbitrary logic within an agent. You can download stuff from the internet, scrape the web, generate images, solve math problems, even program small games within Layla's Agentic Framework. Improvements: - added setting to disable system prompt in Inference Settings -> Custom Prompts: this will disable all character and user information. This is useful if you want to chat with the "raw model" without any instructions. Note this also stops most apps in Layla from working properly. - redesigned the model selection area, now with the ability to export your imported models - Claude API now allows sending images in chat - added Japanese and Vietnamese translations - added ability to organise characters into folders - you can tap on "memories" in your character card to immediately view their memories in the LTM app - added small UI sounds to notify you when Layla is listening or finished speaking - allow adding custom JSON fields to OpenAI endpoints Bug fixes: - fixed a bug where long prompts will crash LTM ingestion process with a custom LLM - fixed bug where Layla as default assistant on phone does not work with multiple inference engines attached to the character - fixed bug where setting default character was not working with using Layla as default assistant on phone - fixed a bug where custom SD models cannot be imported via the "Already downloaded" link

Improvements: - improved download stability of new model downloads in the welcome screen and stable diffusion model downloads - added TTS setting for a global default voice - added the ability to given your character/LLM a custom instruction during chatting in Chat Actions - added UI setting to control text area expansion in voice chat - think content is removed when reloading the chat (a compromise between removing think content on every message causing a reload, which is too slow for mobile) - added "copy" character button for user created characters - TTS now skips speaking unpronounceable characters (such as ASCII art, tables etc.) - added a button in Diagnostics popup that force restarts Layla Bug fixes: - fixed a bug where Dreams keep scheduling messages for the same character more than once - fixed a bug where blockquotes were not being rendered correctly - fixed a bug where Layla character cannot be duplicated - fixed a bug where different model options in Inference Settings cannot be scrolled sometimes - fixed a bug where "Speak Responses" chat actions were not working - fixed a bug where searching chat history does not search through all chat histories - fixed a bug where GPT-OSS model was not generating responses - fixed a bug where if you cancel voice chat before it initialises the UI dark cover is gone - fixed a bug where edited memories are still using old embeddings, causing recalled content to be different to queried content - fixed a bug where reloading chat sometimes do not read from cached session and instead loads everything from scratch - fixed prompt format for GPT-OSS - fixed a bug where Layla as your phone's default assistant was not working properly

Improvements: - improved download stability of new model downloads in the welcome screen and stable diffusion model downloads - added TTS setting for a global default voice - added the ability to given your character/LLM a custom instruction during chatting in Chat Actions - added UI setting to control text area expansion in voice chat - think content is removed when reloading the chat (a compromise between removing think content on every message causing a reload, which is too slow for mobile) - added "copy" character button for user created characters - TTS now skips speaking unpronounceable characters (such as ASCII art, tables etc.) - added a button in Diagnostics popup that force restarts Layla Bug fixes: - fixed a bug where Dreams keep scheduling messages for the same character more than once - fixed a bug where blockquotes were not being rendered correctly - fixed a bug where Layla character cannot be duplicated - fixed a bug where different model options in Inference Settings cannot be scrolled sometimes - fixed a bug where "Speak Responses" chat actions were not working - fixed a bug where searching chat history does not search through all chat histories - fixed a bug where GPT-OSS model was not generating responses - fixed a bug where if you cancel voice chat before it initialises the UI dark cover is gone - fixed a bug where edited memories are still using old embeddings, causing recalled content to be different to queried content - fixed a bug where reloading chat sometimes do not read from cached session and instead loads everything from scratch - fixed prompt format for GPT-OSS - fixed a bug where Layla as your phone's default assistant was not working properly

- fixed a bug with llama.cpp fix params setting

New features: - redesigned character selection page to give you detailed information of your character before starting the chat - added ability to save custom sampler settings in the Advanced Settings screen - added ability to attach saved sampler presets as part of your custom Inference Engine Improvements - supports sending images to OpenAI compatible API inference settings Bug fixes: - fixed a bug where some app data was not deleted when removing a mini-app from Layla - fixed a bug where LTM sometimes crashes when handling very very long messages - fixed a bug where images generated during chat were not being backed-up - fixed a bug where tags from TavernPNG were not being imported - fixed a bug where backup cannot copy references with spaces or special characters in their name - fixed a bug where automatically converting JPGs when saving as TavernPNG was giving an error - fixed a bug where personas were not being saved as part of your Inference Engine - fixed a bug where upscaling anime images caused a crash

- updated llama.cpp to support the latest iPhones

New features: - visualisation of temporal relationships in Layla's long-term memory A while ago, Layla introduced temporal memories: memories that evolve in time. This update brings a whole new way to visualise your character's memories! Your memories are now laid out in a 3D space; past memories are "deeper" into the z-axis. A slider at the bottom allows you to navigate through your memories in time, jumping from past to present. You will notice that memories naturally form "clusters" in time. These are memories that occur within a relatively short timeframe. Each cluster has it's own central node(s), with auxiliary memories branching out around them. These clusters are segmented via a time-dilated algorithm: recent memories are finely-segregated, while older memories are more clustered together. This mimics how human memories work; you often remember recent events clearly, while the longer the events are in the past, the more they combine into one big cluster. Important nodes in each cluster are linked together in time. These are usually "common entities" between temporal clusters. Intuitively, this should be important characters, topics, etc. The graph reflects this. The "anchors" in each cluster are usually the two main participants of the conversation (you and your character). These nodes glow brighter in the memory network, and you can see faint "temporal links" between them. These temporal links are annotated by the time elapsed between each cluster, which is visible in the memory graph. There are a lot more ways in which this marshalled data and relationships can be used during the recall process, and this is something I will continually improve on over time! Improvements: - significant improvements to the performance of chatting, especially for very long conversations spanning hundreds/thousands of messages - long-term memory will now recall earlier memories from the same conversation that has been cycled out through context shifting - fixed a few conditions in triggering new search agent - supports the new Ministral-3 family [THINK] tags Bug fixes: - fixed a bug where invalid filename in References mini-app causes Layla to crash - fixed a bug where characters list will sometimes show nothing when updating a filter - fixed a bug where character list resets to the first page after edit/saving a character - fixed a bug where regenerate messages leaves a large amount of empty space when the length of the regenerated message varies largely - fixed a bug where automatically searching for models in OpenAI compatible API endpoints was not working - fixed a bug where Layla does not stop the server from generating messages even after pressing "stop" - fixed a bug where voice chat does not speak aloud responses - fixed a bug where audio is only played through speakers and not AirPods

New features: - Layla supports Qwen3-VL family models: images are automatically down-sized before sending to the LLM to keep processing times reasonable on mobile, you can control how much down-sizing occurs in Advanced Settings - Layla supports animated PNGs and GIFs for your character cards Improvements: - added setting in Reference documents to control the threshold of how long a document needs to be before it is summarised instead of injected wholly into the context - prompt templates are now case insensitive - display character full name in chat instead of just the first name - automatically convert JPGs to PNGs when exporting your characters as a TavernPNG - you can share text with Layla to start a chat with that text Bug fixes: - fixed a bug where the document name was not detected correctly when adding a document from chat - fixed a bug where character sorting changes back to the default after exiting chat - fixed a bug where tool calls are included in long-term memory messages - fixed a bug where Layla crashes on importing a huge TavernPNG card

- fixed bug where option toggles are not showing

- improved stability in chatting - fixed a few bugs in edit/regen - fixed bugs related to the calendar agent

Layla 6 supports Agents! Agents are fully configurable, self-contained workflows that can be triggered and executed by Layla during chats. Their functionalities can range from injecting simple context, to fully automated workflows such as reading a webpage and saving it as a reference document. Each agent is also modular: you can attach them to your own characters; they will still complete their respective tasks, but under the personality of your attached character. An agent is comprised of two parts: triggers and tools. An agent can have multiple triggers and call multiple tools. You are free to mix and match them in any combination when creating your own agents! New features: - Long-term memory has been significantly improved! - LTM now implements a 4 stage retrieval process, to balance speed and quality: - “raw memories” are created immediately as you chat; they are available immediately when starting the next conversation, no processing time needed - “embeddings” are created in the background; this process is relatively fast, and will be available 5-10 minutes after a conversation ends - “summaries” are created as a background process, when your phone is idle; this process can either be done by the built in summariser, or a chosen LLM - “knowledge graphs” are created last, with all the information available in the previous step, and require an LLM to process - During recall, memories are retrieved starting from “raw memories”, and re-ranked by content in each level as they become available, so you will immediately get memories, but the injections will increase in quality as new memories are ingested Improvements: - support for Gemma 3n models - improved prompt format slightly for vision models - backup and restore data now prevents back button presses or navigation to avoid accidentally corrupting your data if you move away from the screen - added ability configure a character level image generation prompt prefix - LTM no longer processes think tags from reasoning models - performance improvements to rendering chat messages - Layla will now automatically select the best prompt format after you've selected your model - backups & restore will now also backup your OpenAI inference settings Bug fixes: - fixed a bug where favourite characters were not being saved - fixed a bug where searching for language names in the Offline Translator app was searching for the language's English name instead of the name in the UI language - fixed a bug in native speech-to-text where it stops listening after the first round - fixed a bug where horoscopes were not showing - fixed a bug where a character with an empty description stops all characters from showing up - fixed a bug where OpenAI inference settings was not backed up

- fixed bug where Metal GPU inference was not working

Improvements: - added support for Kroko ASR (speech-to-text model, this is now the default STT model in Layla) - ability to configure KV Cache quantisation settings in Advanced Settings - added French, Portuguese, and Spanish UI translations Bug fixes: - fixed bug where the correct number of characters in Roleplay was not showing - fixed bug where chat images cannot be expanded when tapped on - fixed bug where slightly moving your finger during hold-to-speak mode stops Layla from listening - fixed bug where sometimes phone native voices are not being shown or not working in text-to-speech

- fixed bug where edit and delete actions were not showing for previous messages - removed animated backgrounds temporarily due to a bug

New features - app UI has been translated to multiple languages! Currently supports: English, Chinese, Korean, Polish, Russian - Offline Transcription: transcribe audio such as lectures, audiobooks, or videos in real time - all preset characters now have animated backgrounds! Improvements: - allow configuring minimum silence length and other settings for speech-to-text - made stop generating slightly more responsive - improved UI performance when generating responses, should reduce speed inconsistencies where response times sometimes start slow and speeds up at the end of the message - added TopK in advanced settings Bug fixes: - fixed bug where DuckDuckGo news search was not working - fixed bug where multiple regens make the model increasing incoherent - fixed bug where text-to-speech sometimes pauses unnecessarily after the first sentence

- fixed bug where the AI sometimes repeats what you said before responding

- fixed a small bug where chat screen is not redirecting to Layla Cloud login if you are logged out