AI Systems Architect · Independent Researcher · São Paulo
AI systems should retain context, adapt over time, and remain useful beyond a single interaction.
I design memory architectures, cognitive frameworks, and long-running AI systems — and I build the products that prove they work. One thesis, pursued for years, across research and production.
Why memory
Most AI systems are amnesiac by design. Every session starts from zero: no accumulated context, no evolving understanding of the person on the other side, no continuity of purpose. We've built remarkable reasoning engines and given them the memory of a goldfish.
That's not a model problem. It's an architecture problem — and architecture is where I work.
My research and systems all attack the same gap: how machines retain, structure, decay, and retrieve context over months and years, not minutes. Memory, persistence, and personality are not features to bolt on. They are the substrate.
The Lab
Ongoing research programs. Each one is a different cut at the same question: what does it take for an AI system to genuinely persist?
Enter the lab →LoomDB
A graph-based temporal memory architecture.
Context as a living graph: activation spreads, attention decays, relevance is a function of time. Memory that behaves less like a database and more like a mind — recency, frequency, and association shaping what surfaces and what fades..
Active research · Proprietary
agent memory · retrieval · temporal graphs
EPCG
An emergent personality framework for AI agents.
Personality as structure, not system prompt. EPCG composes agent identity from three interacting layers — Behavior, Belief, and Biography — drawing on category theory to make the composition lawful, inspectable, and stable over time..
Active research · Proprietary
cognitive architectures · agent identity · category theory
Ayvu-Talian
Language modeling for minority language preservation.
A decoder-only transformer trained for Talian, a Venetian-derived language spoken in southern Brazil. What language modeling looks like when the training corpus is a community's living memory, not the internet — and the failure mode is cultural loss..
Active research · Open work
low-resource NLP · cultural persistence · from-scratch transformers
The Systems
Research that doesn't ship is speculation. These are the systems where the thesis runs in production.
All systems →Flagship · Case study
Pixie
An embodied AI learning companion.
Pixie talks, listens, remembers, and grows with the child using it. Built neurodivergent-first: predictable interaction patterns, sensory-considered design, and a memory that lets the companion know you on day ninety the way it couldn't on day one.
- 1.Voice-native interaction with real tool use
- 2.Long-term memory built on the lab's architecture research
- 3.Embodied presence
Next.js · Convex · Three.js · Tauri
Read the case study →
Open source
MultiverseState management with Git-style branching semantics.
Fork application state, explore alternate timelines, merge what works. Multiverse treats state history as a first-class data structure — the same thinking that drives my memory research, applied to frontend state.
Open source · MIT
$ npm install multiverse
Writing
Essays and research notes on memory, context engineering, and the architecture of AI systems that last.
All writing →RSSFirst essays arriving soon. Subscribe via RSS.
Work with me
I take on a small number of engagements where the memory problem is the hard problem.
If you're building agents that need to persist, retrieval that needs to mean something, or AI products that should still be useful in month six — this is what I do all day.