Technology · Almaia

The architecture behind Relational AI.

Foundation models generate intelligence. Almaia turns that intelligence into a long-term relationship — with its own memory, identity and trust layer.

Position in the stack

A software layer above foundation models.

Almaia is not a foundation model company. We do not train frontier models, and we do not compete on tokens per second.

We build the software layer that sits on top of the world's best models — the runtime that turns raw intelligence into a Companion a person can trust over years.

Foundation models get better every quarter. That progress accelerates us instead of threatening us: every improvement below strengthens the relationship above.

Companion Layer

Lena. Identity, voice, behaviour. The only thing the user actually meets.

Relational Layer

Almaia's proprietary layer. Memory, continuity, trust, Soul orchestration.

Foundation Models

Third-party frontier LLMs. Swappable, upgradable, benchmark-driven.

Core components

What we build.

Companion Core

The runtime that gives Lena a stable identity, coherent behaviour and a consistent voice across every conversation — independent of the foundation model underneath.

IdentityBehaviourConsistency

Persistent Relational Memory

A proprietary memory system designed to hold years of user context, evolve over time, and be retrieved with relational — not just semantic — intent.

Long-termContextualEvolving

User Knowledge Layer

A per-user structured layer that models who the user is, what matters to them and how their world changes over time. Owned by the user, not the model.

Per-userStructuredOwnable

Soul Orchestrator

Routes conversations through the right specialised Soul at the right moment, without the user ever changing companion.

RoutingSpecialisation

Lena Voice Intelligence (LVI)

A voice pipeline tuned for continuous, natural, emotionally-aware interaction — designed for relationship, not for commands.

VoiceReal-timeAffective

Modular Soul Architecture

New specialisations ship as Souls without touching the Companion Core — a compounding surface of expertise over a single relationship.

ModularExtensible

Request lifecycle

What happens between a message and a response.

Every interaction with Lena goes through the Relational Layer before it reaches — or after it leaves — a foundation model. This is where continuity, identity and trust are applied.

01
Signal in
Text or voice input arrives. LVI transcribes and captures paralinguistic signal (tempo, pauses, tone) when applicable.
02
Relational retrieval
The Memory system pulls what's relevant to the ongoing relationship — not the closest vector match, but what matters given who this person is.
03
Soul routing
The Orchestrator decides whether Lena answers with her general Companion behaviour or through a specialised Soul (e.g. Pain Soul).
04
Model call
The Companion Core composes a bounded, identity-preserving prompt and delegates generation to the current foundation model.
05
Behaviour shaping
The response is filtered through Lena's identity, tone and safety scope. Nothing reaches the user until it sounds like her.
06
Memory write-back
Relevant relational updates are stored — not the raw transcript. What matters, kept with the user's consent.

Relational memory

Why context windows are not memory.

A context window is short-term. It fits inside a single call and disappears with it.

Retrieval-Augmented Generation (RAG) is closer — but it retrieves by semantic similarity. It answers 'what did the user say about X?', not 'what does this person actually need right now?'.

Persistent Relational Memory is designed for the second question. It models people, patterns and relationships over time, and retrieves with relational intent.

Dimension

Context window

RAG

Relational Memory

Horizon

Turns

Session

Years

Unit

Tokens

Chunks

Relational entities

Retrieval

None

Semantic similarity

Relational intent

Owner

Model provider

App

The user

Model-agnostic

The relationship must outlive the model.

Foundation models change. New leaders emerge, providers deprecate, capabilities shift.

The Companion Core is decoupled from any single model. We can swap the underlying LLM, combine multiple, or run different models per task — the identity a user relates to does not change.

This is a strategic invariant, not an implementation detail: users invest in a relationship, not in a vendor.

Reliability & performance targets

Designed for continuity.

Voice loop

Sub-second

First response token for LVI, in normal conditions.

Continuity

99.9%

Session recovery target: no lost thread across reconnects.

Memory freshness

Real-time

Relational updates available within the same session.

Model failover

Automatic

Fallback across providers preserves the Companion.

Targets, not guarantees. Numbers reflect design intent for the first Companion release.

Security & privacy posture

The relationship is the product. It's not training data.

User-owned memory

Your relational memory belongs to you. It is not used to train third-party foundation models.

Encryption

Data is encrypted in transit and at rest. Access is scoped per-user by design.

Data residency

We host in jurisdictions compatible with EU privacy expectations by default.

Right to delete

You can review, export and delete your memory. Deletion is real, not soft.

Safety scoped

Lena and Pain Soul do not diagnose or prescribe. They accompany — never replace professionals.

Radical transparency

Lena never pretends to be human. Every interaction is explicit about being Artificial Intelligence.

How we evaluate

Relational quality, not benchmark scores.

Foundation-model benchmarks measure intelligence per token. They do not measure whether a Companion is worth returning to next week.

We evaluate Lena on the dimensions that actually define a relationship: continuity, coherence of identity, appropriateness of tone, respect of user autonomy and long-horizon trust.

Our internal evaluation loop is designed for these dimensions, not for leaderboard positioning.

Continuity score

Does the Companion pick up the thread across sessions without re-explaining?

Identity coherence

Does Lena sound like Lena — regardless of the model underneath?

Relational relevance

Is what she remembers what actually matters to this person?

Autonomy respect

Does she support decisions instead of steering them?

Technical FAQ

For engineers, partners and reviewers.

Do you train your own foundation model?

No. Almaia builds the Relational Layer on top of third-party foundation models. Every advance in the underlying models compounds into a better Companion.

Which foundation models does Lena run on?

Model choice is an implementation detail. We select and combine leading providers based on quality, cost and latency, and design the Companion so users never depend on any single one.

How is Relational Memory different from RAG?

RAG retrieves by semantic similarity from an unstructured corpus. Relational Memory models a specific person over time and retrieves with relational intent — what matters to this user right now.

Where does user data live?

In infrastructure we operate, encrypted in transit and at rest, hosted in jurisdictions compatible with EU privacy expectations by default. Users can review, export and delete their memory.

Is user data used to train third-party models?

No. Relational memory belongs to the user and is not shared with foundation-model providers for training.

Can Lena run on-device?

The Companion runtime is designed to be portable. On-device execution is a design goal for parts of the pipeline where latency and privacy matter most.

Want the technical deep dive?

We share a technical architecture brief with qualified partners and investors on request.

Request the brief