Layla 7: Custom Mini-Apps and the Layla SDK, Real-Time Long-Term Memory, and Multilingual Voices

Custom mini-apps and the Layla SDK

The centrepiece of Layla 7 is that mini-app development is open to everyone!

Mini-apps are where Layla's features live and how the app has grown over time. Until now they were always built in-house and shipped as first-party updates; v7 changes that, so anyone can create a mini-app and extend Layla however they want.

The Layla SDK lets custom mini-apps integrate with Layla's language model, characters, image generation, agents, memory, and more, all running on-device. A mini-app can be a single HTML file, a zip folder with its own assets, or a self-hosted website, and mini-apps can be organized into folders and searched like characters.

If you would rather have a tool build the mini-app for you, the SDK ships an agent skill compatible with Codex, Claude Code, and OpenCode, plus a mini-app template for larger projects.

The SDK surface grew steadily across the alphas and now covers the language model, characters, and image generation (including image-to-image from a base64 source); chat sessions, chat history, and sentiment detection; reading and saving files and chat messages; reading, creating, and operating on memories so a mini-app can run its own memory ingestion; text-to-speech voice listing and generation; message scheduling, persona lookup, and top-memory recall; and getting and setting the inference engines saved in Inference Settings. Chat responses now return reasoning and message content in separate fields, matching the OpenAI API convention.

The point of all this is customization. You can build a long-term memory app that decides how a character's memories are coalesced and served, a note-taking app that reads from and writes to that memory, a budgeting tool, a mini-app that pulls data from a smart watch, or a custom inventory system for roleplay scenarios. Your characters and their memories are already yours; the SDK extends that to how you use the app itself.

Mini-apps are not confined to your own device. You can upload a custom mini-app to Layla Cloud so it appears in Browse Apps for other users, manage published versions from the Mini-Apps Manager, and share mini-app links directly from Layla. User-created mini-apps are already showing up in Browse Apps and the community showcase, including ones the team would not have thought to build.

This is also a change in how Layla itself is developed. Going forward, most new features will be built as mini-apps on the same SDK you have access to, and they will be open source, so you are free to take any of them and adapt it to your own use. The SDK will keep expanding alongside Layla, adding access points and APIs for more complex builds.

Real-time long-term memory

Long-term memory has been rebuilt for v7. Memories are now processed in real time during a conversation, so Layla no longer needs to ingest them in the background overnight.

The default memory pipeline uses a custom-trained BART model that generates summaries and knowledge graphs together. It was trained on roleplay data distilled from popular uncensored roleplay models, including Cyndonia and Bloodmoon, and is built to be fast enough for live conversations while still producing useful recall for most chats. LLM-based memory generation is still supported for anyone who prefers it, and it is now integrated with the SDK so a mini-app can fully replace the default processor.

Because memory is handled on-device, your conversation history and everything derived from it stay private to your phone. The knowledge graph display was redesigned to represent temporal nodes and time links, with a new full-screen view. Deleting or regenerating a message now removes or updates the memory entry tied to it, and recall was refined to avoid pulling the most recently created memories from the same session, which helps prevent feedback loops.

Dream, redesigned as a mini-app

Dream is now a downloadable custom mini-app with deeper character self-reflection. It combines sentiment analysis, recalled long-term memories, and per-character prompts to generate an impression of where a relationship stands, and those impressions evolve over time and can be reviewed or edited before you start a chat.

Dream can also continue a conversation more intelligently by drawing on past memories, the current context, and previous impressions.

For an offline AI companion, this means a character carries a sense of history between sessions rather than starting cold each time. Three SDK APIs support the mini-app: top-memory recall weighted by frequency and recency, future message scheduling, and persona lookup that falls back to the global persona when a character has none. In this release Dream generation is triggered manually inside the mini-app; automatic scheduling is planned as part of the v7 task and scheduling overhaul.

Multilingual voices and a text-to-speech SDK

Layla 7 adds Kokoro v1.0 voices, including multilingual options, alongside the existing PocketTTS one-shot voice cloning. Custom mini-apps can now reach Layla's speech system through the SDK: they can list available voices, generate spoken audio, pass a null voice ID to use the default voice configured in Layla, and stop the current audio on demand. Speak Responses and voice chat both received reliability fixes covered below.

More model formats and inference control

Beyond GGUF models on llama.cpp, Layla 7 adds LiteRT model support to both the Roleplay mini-app and custom mini-apps.

It also adds support for Gemma 4 MTP draft models: import the draft model through the File Manager, open its details to copy the full path, and set that path in the llama.cpp command-line arguments. Custom sampler presets in Advanced Settings now correctly save additional llama.cpp arguments, and mini-apps can read and set the saved inference engines through the SDK.

Improvements

• The File Manager mini-app can import arbitrary files, including mini-app assets and draft models, shows total folder sizes in folder details, and adds an image preview icon. It is faster and more stable with large numbers of files.

• Custom mini-app icons are cached so they stay available offline, and mini-apps on the Browse Apps page load faster.

• A wallpaper can be removed without uninstalling its mini-app.

• Backups now include character folders and character sorting.

• The Benchmark mini-app was updated to match Layla's current visual style.

• Sentiment detection aggregates across multiple sentences and normalizes pronouns.

• Character selection is faster for very large character libraries.

• Importing an existing mini-app now asks for confirmation and overwrites, rather than blocking the import.

Bug fixes

• Fixed LLM-based long-term memory generation getting stuck while building knowledge graphs, and knowledge graphs that could open to a blank page.

• Fixed long-term memory that could still record messages for a character with LTM turned off.

• Fixed scheduled future messages appearing on the chat history page.

• Fixed the LiteRT engine failing to load certain models without vision encoders.

• Fixed Dreams occasionally skipping scheduling when using cloud models.

• Fixed models linked from an SD card that could not be deleted.

• Fixed image-to-image generation not applying the prompt prefix, and image generation starting during a model's thinking time when it reasoned about a Stable Diffusion prompt.

• Fixed custom mini-app image generation reading default settings instead of saved generation settings.

• Fixed Speak Responses reporting that TTS was not initialized, Layla Cloud voices not stopping when requested, PocketTTS not working through the SDK, and the SDK stop-speaking call not working.

• Fixed voice chat minimizing text even when that setting was off.

• Fixed custom mini-apps sending a system prompt Layla did not use, and chat history still appearing when swiping left in custom mini-apps.

• Fixed importing a Roleplay from a Layla RoleplayPNG not saving its characters.

• Fixed prompts that could be processed again unnecessarily, and a crash when loading a very large conversation history.

Layla 7 runs the same way Layla always has: fully on your Android device, offline, with no account or subscription required to use your own models. The difference is that the assistant is now a platform you can extend. And because all of it runs locally, the way you build and use Layla stays yours and can't be taken away.