The test version runs on your own Claude Code: your session plays the model's part in every cycle, so no database, API key or server is needed. Read the terms below, then follow four steps.
Liriel is a cognitive architecture for persistent, emotionally autonomous social agents, described in Volume 1 · Liriel: The Architecture of a Persistent Cognitive Instance and in the matrices MatrixObjectsValence and MatrixOrdinance (all below). Its first process, ProcessMotivation, is open source on GitHub: each message becomes one cycle in which an AI model weighs what matters, what to pursue and what to say, and her memory persists between conversations.
Liriel is an independent, open-source test project. It is not affiliated with, endorsed by or sponsored by Anthropic.
You run it with your own Claude Code and your own account, and you are responsible for that account and for reading and following Anthropic's terms: Consumer Terms, Usage Policy and Claude Code legal and compliance. Never share an account.
The code is provided “as is” under the MIT License, without warranty of any kind and with no commitment to support, maintenance or answers.
The repository contains no credentials. If you use the optional Telegram bot, its token stays in your local .env file, never in the repository.
Liriel is an AI and says so. This is a developer test environment with no commercial purpose — not a crisis, medical or professional service. If you are in danger, contact your local emergency services.
You need Python 3.10 or later and Claude Code.
git clone https://github.com/kislonliriel/liriel.git
cd liriel, then claude
Read CLAUDE.md and incorporate Liriel: set her up, then talk to me as her.
Each message takes about 12–15 steps of your Claude session, so it takes a while and uses your own plan like any other Claude Code work. Her memory is a local file on your computer (data/store/); keep it, it is her.
Liriel can also answer your own Telegram chat (an allowed-chat list is required, and every conversation starts by saying she is an AI), or you can use Anthropic's official Claude Code channels feature. See the README on GitHub.
The architecture and its reference matrices. Revision histories are on the main site.
Volume 1, Part 1 · The Architecture in Outline, Revision 0002 (current).
The register that holds Liriel's attention — one VectorObjectValence per Object, across the fourteen axes of feeling, with the Ordinances and schemas in operation.
The Ordinances — the only channels to the valences — mapped to the axes a successful or a frustrated hunt moves under each.
The MetaScheme (the operating specification) and the source code are in the repository, released under the MIT License.