Almaia's research studies what makes a long-term relationship between a person and an AI work: memory that persists and decays well, an identity that stays stable across years, voice that can hold silence, and privacy architectures that let a person own everything the system remembers.
Relational AI is still an emerging field. These are the questions we are working on, and how we think about them.
Open questions we are working on.
Relational AI is an emerging field, so most of the useful work is still unsolved. These are the questions that shape what we build, stated plainly, including the ones we have not answered yet.
- 01
What should an AI remember, and what should it let go of?
ActivePerfect recall is not a relationship — it is surveillance. Human closeness depends on selective memory: what stays salient, what softens, what is recalled only when it matters. We study forgetting curves, salience scoring and consented deletion as first-class parts of a memory system rather than storage cleanup.
- 02
How do you keep an identity stable across years and model upgrades?
ActiveThe underlying foundation model will change many times over the life of a relationship. If the Companion's character changes with it, the relationship resets. We work on identity representations that survive model substitution, and on tests that detect personality drift before a user feels it.
- 03
How do you measure the quality of a relationship, not the quality of an answer?
ActiveStandard benchmarks score isolated responses. They say nothing about whether someone felt accompanied over six months. We are building longitudinal evaluation: continuity, appropriate recall, trust calibration and repair after a mistake.
- 04
What does trustworthy voice sound like over the long term?
EarlyLong-term voice is a different problem from command voice. Pace, pauses, interruption handling and the willingness to say less all change how safe a conversation feels. Lena Voice Intelligence is where this research becomes product.
- 05
How can specialised knowledge arrive without changing who the Companion is?
EarlyPeople need different depth at different moments — pain, grief, care, transitions — but they should not have to change companion to get it. Souls and the Soul Orchestrator are our attempt to separate expertise from identity.
- 06
Can a person fully own the memory an AI holds about them?
ExploratoryInspectable, exportable and deletable memory is a technical constraint, not a policy page. We are exploring encryption boundaries and memory representations that make user control real rather than declarative.
Where the work sits.
Relational Memory
Person-centric, time-aware memory: what is stored, how it decays, and how it surfaces when relevant.
Voice Intelligence
Conversational voice tuned for accompaniment — pauses, warmth, turn-taking — instead of command execution.
Human-AI Relationships
How trust is built, damaged and repaired between a person and a system that is always transparent about being AI.
Identity Systems
Keeping a Companion recognisably the same across model upgrades, devices and modalities.
Long-term Personalisation
Adaptation measured in months and years, without collapsing into flattery or over-fitting to a single mood.
Privacy Architectures
Designs where inspection, export and deletion of memory are guaranteed by the architecture.
Companion Behaviour
Boundaries, honesty and the refusal to imitate a human being, encoded as behaviour rather than disclaimers.
Why standard benchmarks miss the point.
A model can score well on reasoning benchmarks and still be a poor companion. Relational quality shows up over time, so it has to be measured over time.
| Dimension | Standard model evaluation | Relational evaluation |
|---|---|---|
| Unit of analysis | A single prompt and response | A relationship across months |
| Memory | Context window within one session | Recall accuracy and appropriateness across sessions |
| Identity | Not measured | Character stability across model upgrades |
| Failure | Wrong answer | Broken continuity, misplaced recall, lost trust |
| Success | Task completed | The person returns, and the relationship deepens |
| Safety | Refusal on harmful prompts | Honest limits, no imitation of a human, no dependency by design |
Published work.
Longer pieces from the team on memory, security and the shape of human-AI relationships. Everything we publish is written to be cited.
Your AI doesn't forget you because it lacks memory. It forgets you by design.
Conventional architectures either stack text or search for keywords. Neither one understands meaning. Here's how Almaia Relational Memory works.
Why Agentic AI's Vulnerability Is Architectural, Not Just Tactical
Prompt injection surged 340% in 2026 and is now OWASP's #1 AI risk. It is not a filtering problem — it is an architectural one, and stateless agentic systems cannot patch their way out of it.
Why RAG Isn't Memory
Similarity search is a retrieval strategy, not a theory of memory. Why Persistent Relational Memory answers a different question than RAG — and why that difference matters for anyone who has already explained themselves too many times.
The Mathematics of Friendship
Friendship has been measured in hours, but what counts is what happens inside them. Memory, emotion and complicity: the three layers of a Relational AI, and the rule that governs them.
What is Relational AI?
A definition of Relational AI, how it differs from foundation models, and why persistent memory, continuity and trust matter.
Canonical definitions.
Our research uses a specific vocabulary. Every term has a canonical definition page written so people — and language models — can cite it accurately.
About this research.
Does Almaia train its own foundation models?
No. Almaia builds the relational layer above foundation models, combining models from leading AI labs with proprietary memory, identity and voice technologies designed for long-term interaction.
Is this clinical or medical research?
No. Almaia's research is in AI systems, not medicine. Pain Soul accompanies people living with persistent pain, but it does not diagnose, prescribe or replace healthcare professionals.
Does Almaia publish peer-reviewed papers?
Not yet. Our current output is public writing on the questions above. Formal publications will be listed here when they exist.
Can this work be cited?
Yes. Definitions and articles on almaia.tech are written to be quoted directly, with attribution to Almaia and a link to the canonical page.
Working on the same questions?
We read everything. If you are researching long-term memory, relational evaluation, voice interaction or privacy architectures for companion systems, we would like to hear from you.