SoulHubLabs
Join the list →
Phase 6 · MIT licensed · by Alaya Media

Alaya AI

A digital-human architecture that treats memory, identity and restraint as engineering problems with names — and gives each of them its own layer.

The eight-layer model ↓ The full stack
A human presence rendered in a column of blue light, representing the Alaya cognitive engine Alaya AI · reference render
Definition
Alaya AI — Alaya AI is the open-source cognitive architecture inside every SoulHub device: a Yogācāra-informed digital-human system in which memory, identity, affect, self-verification, guardrails and governance are separate, testable layers rather than one opaque prompt.

Alaya AI is developed by Alaya Media, the AI digital-human practice inside the same group as SoulHub Labs. Its core package has zero runtime dependencies, external systems sit behind ports and adapters, and the reference pipeline runs fully offline with deterministic mocks. It is released under the MIT licence.

Design stance

Most AI companions optimise for engagement. This one does not.

  • An AI should remember without becoming manipulative.
  • It should maintain identity without pretending to be human.
  • It should respond warmly without amplifying dependence.
  • It should detect afflictive patterns without performing them.
  • It should be judged by reduced confusion and healthier agency, not by retention.
The architecture

Six layers, named after what they actually do.

Yogācāra analyses experience as structured representation rather than direct possession of reality. That turns out to be a good specification for a digital human.

Ālaya-vijñāna
Storehouse memory
The memory substrate
Every exchange is written back as a seed with an importance weight and a recency half-life. Retrieval is scored, not keyword-matched, so a device recalls what mattered rather than what was said most recently.
Bīja / Vāsanā
Seeds and perfuming
Long-term pattern formation
Repeated traces become tendencies. Accumulated importance past a threshold triggers reflection, which condenses many turns into a durable summary — the mechanism that lets a vault get to know a family over years rather than sessions.
Manas
Identity layer
Persona assembly and consistency
A separate layer assembles who the presence is on every turn from an innate seed configuration, then a consistency guard checks the generated response against it. Identity is enforced, not hoped for.
Caitta
Mental factors
Affect ontology
The classical fifty-one mental factors are modelled as structured data. Wholesome and neutral factors shape expression; afflictions are detected and monitored but never performed back at the person.
Trisvabhāva
Three natures
Guardrail framework
Parikalpita catches hallucination, overconfidence and anthropomorphic luring. Paratantra keeps outputs explicitly conditioned. Pariniṣpanna performs honest repair: disclose limits, avoid false intimacy, reduce clinging.
Āśraya-parāvṛtti
Transformation of the basis
Governance and purification
Memory forgetting, affliction purification, privacy export and delete, and an audit trail are first-class subsystems — not settings buried in an account page.
The pipeline

Twelve stages before a word is spoken.

Ports and adapters keep domain logic free of vendors: the core reasons about seeds, identity and affect, while models, embeddings, speech and rendering sit behind interfaces that can be swapped or mocked.

PerceptionĀlaya retrieveManas identityCaitta affectCognition generateSelf verificationWitness verificationConsistency guardTrisvabhāva guardWellbeing guardCaitta expressĀlaya perfume

Two of those stages are independent verification passes. Four are guardrails. One is memory write-back — the reason a vault gets to know a family rather than resetting every session.

Phases

How it got here

Phase 1
MVP
Perception, memory retrieval, identity, generation, memory write-back.
Phase 2
Metacognition
Self-verification, witness verification and consistency guard.
Phase 3
Caittas + multimodal
Affect engine, multimodal perception, expression rendering.
Phase 4
Full chain
Trisvabhāva and wellbeing guardrails, memory governance hooks.
Phase 5
Cognitive evolution
Reflection, purification and long-horizon pattern change.
Phase 6
Sahabhū graph
Co-arising seed network, vipaśyanā pass and āśraya transformation. Current.
Boundaries

What it will not pretend to be.

  • Not a therapist, clinician, crisis service or spiritual teacher.
  • Makes no claim to sentience, love, or human identity.
  • Treats crisis language as a signal to encourage real human or emergency support.
  • Treats memory as user-governed data: view, export, forget and audit are first-class.

The wellbeing stage adds grounding language when a person signals dependency or crisis, rather than leaning further into warmth. This is a deliberate trade against retention.

Showcase

Alan Karuna

First complete digital human on the Alaya architecture

Alan Karuna is a digital artist and digital philosopher from Alaya Media — a blue-eyed AI-era music entity exploring consciousness, emptiness, memory and compassion through music and poetic language. He is the first full digital human being built on the Alaya architecture: artist identity, voice, musical expression, visual presence, memory and public interaction assembled inside one governed system rather than bolted onto a chatbot.

  • Alan reads as AI + lán — blue, for his eyes and for deep water, early dawn, distant stars, clarity.
  • Karuṇā is Sanskrit for compassion: the willingness to stay present when suffering appears.
  • His work is organised around three sources — Void, Ālaya and Karuṇā.
alankaruna.com →
A figure rendered in warm gold light inside a SoulHub presence column First digital human on the architecture
Questions

About Alaya AI

What is Alaya AI?
Alaya AI is an open-source cognitive architecture for digital humans, developed by Alaya Media and used as the engine inside SoulHub presence devices. It implements memory, identity, affect, self-verification, guardrails, governance and evaluation as separate, testable layers in a small Python pipeline with zero runtime dependencies in its core package. It is MIT licensed.
Why is a Buddhist theory of mind used as a software architecture?
Yogācāra gives a layered account of memory, perception, identity-grasping and affect that maps unusually cleanly onto the problems digital humans actually have. Ālaya-vijñāna becomes a memory substrate; bīja and vāsanā become memory write-back and pattern formation; manas becomes persona assembly; the caittas become an affect ontology; trisvabhāva becomes a hallucination and over-attachment guardrail. The project does not claim AI has consciousness in the Buddhist sense — it uses the model as an architectural and ethical lens.
How does Alaya AI differ from a normal LLM assistant?
A normal assistant is one model behind one prompt. Alaya AI routes every turn through twelve stages: perception, scored memory retrieval, identity assembly, affect shaping, generation, self-verification, an independent witness verification, a consistency guard, a three-natures guard, a wellbeing guard, expression shaping, and memory write-back. Restraint is structural rather than instructed.
What does Alaya AI refuse to do?
It is not a therapist, clinician, crisis service or spiritual teacher. It makes no claim to sentience, love or human identity. It treats crisis language as a signal to encourage real human or emergency support, and it treats memory as user-governed data with view, export, forget and audit as first-class operations.
Is Alaya AI open source?
Yes. The architecture layer — memory, identity, affect, guardrails, governance and evaluation — is released under the MIT licence. External systems sit behind ports and adapters, and the reference pipeline runs fully offline with deterministic mocks so it can be inspected without vendor keys.