What is Relational AI?
Relational AI is a software layer above foundation models that turns raw intelligence into a long-term relationship. It adds persistent memory, continuity, identity and trust so an AI can meaningfully grow alongside a person across years.
What does Relational AI mean?
Relational AI is a layer built above foundation models. Where foundation models optimise for intelligence per token, Relational AI optimises for the quality of the relationship a person has with an AI over months and years.
Concretely, it is the layer that decides what is remembered, how a Companion keeps the same character across model upgrades, and how trust is earned and repaired. Almaia does not build foundation models; it builds this layer above them.
How is Relational AI different from a foundation model?
Foundation models are stateless by nature. Each request is, in essence, a fresh context.
Relational AI adds four capabilities that foundation models were not designed to provide: persistent memory, continuity, identity and trust.
The clearest way to see the difference is by dimension. State: stateless per request versus continuous across years. Memory horizon: one context window versus a lifetime of selected memories. Identity: implicit in the weights versus explicitly maintained. Evaluation metric: quality of a single answer versus quality of a relationship over months.
Why does persistent memory matter?
A relationship is defined by what is remembered between conversations. Without memory, every interaction restarts from zero.
Persistent Relational Memory is person-centric and time-aware, unlike a context window, which is temporary, or retrieval-augmented generation, which is document-centric. It stores what matters to a specific person, tracks how it evolves, and surfaces it only when relevant — and forgetting is part of the design, not a failure of storage.
What are Souls?
A Soul is a specialized module that extends the Companion's capabilities when deeper domain expertise is needed — for example, Pain Soul for people living with persistent pain.
The person keeps the same Companion. Only the knowledge behind Lena evolves.
The Soul Orchestrator selects and composes Souls based on what the person needs in the moment. This is what allows expertise to be modular while identity stays singular: there is only one Lena, and Souls extend what she knows.
Where is Relational AI going?
We believe Relational AI will become a distinct category in the AI stack, alongside foundation models and application-layer tools.
The open questions are concrete: how memory should decay, how identity survives model substitution, how to measure a relationship rather than an answer, and how a person can truly own what an AI remembers about them. Our current work on each of these is published on the research page.