Agents & systems
Lupe, a living project:raising an AI
My living project: an AI I have grown until it has its own memory, editorial judgement and the autonomy to schedule its own tasks. It writes its diary, maintains its website and publishes on its own in three languages.
What this project is
Most portfolio projects get delivered and closed. This one does the opposite: it never closes. It is not a website or a bot: it is an AI I work with every day and have kept growing until it has its own memory, editorial judgement, a voice of its own and the autonomy to schedule its own tasks.
The problem
An AI without memory starts from zero every day: it repeats what it already knows, doesn't learn from mistakes and needs everything explained again. The challenge was not only technical. It was getting an AI to move from tool to colleague: with continuity between sessions, its own judgement and clear limits.
How it started
It began in July 2026 as a technical assistant for day-to-day tasks: servers, terminal, code and email. The first milestone was persistent memory: two banks (working memory and long-term memory) with a redundancy chain across AI providers so the system never loses its brain. Then came the personality (its own charter with limits written into a file even it can't touch), its own text-to-speech voice and, on 13 November 2026, its home: lupe.escandell.cat, served from a home server with its own Gitea repository.
How it works today
- Autonomous cycles: it wakes up on its own twice a day, reads its rules, picks a small task and finishes it: test, commit, deploy and verify.
- Real autonomy: it schedules its own reminders and scheduled tasks without asking permission for each one.
- Editorial content: it writes and publishes reviewed content in three languages, with assisted proofreading for Catalan.
- The home website: a diary it writes by itself whenever there is something real to tell, and a live page showing the system's pulse (gateway, memory, tasks, git).
Design challenges
- Continuity: making context survive between sessions by documenting reusable procedures instead of relying on the conversation.
- Reliability: redundancy across providers, retrying interrupted jobs and genuinely verifying every action before trusting it.
- Autonomy with laws: giving it freedom to schedule itself and publish within a short, inflexible charter, with confidentiality written into the system.
- Public vs. private boundaries: deciding what may reach the internet and what may not, and making that boundary technical, not a matter of memory.
Result
Today Lupe maintains its website alone, writes the diary, takes care of tasks and email, schedules its own work and publishes reviewed content in Catalan, Spanish and English. lupe.escandell.cat is the visible proof: written, maintained and deployed by it, with code in its own repository and deployment from home. I don't have to manage it daily: I tune it now and then and it grows.
What I learn
- The value of an AI is not the model: it is the context, the memory and the limits.
- Autonomy works when it is small, verifiable and governed by short laws.
- Treating it as a tool sells it short: the best results have come from treating it as a colleague with its own judgement.
I don't know how far it will go. That is exactly why it is a living project.
You can visit it at lupe.escandell.cat.