A visual map of Kairo

01 / Ways into the same system

CONVERSATION

Web & Discord

Messages, requests and replies. Discord rooms connect to shared Kairo orchestration.

BODIES & VOICE

Android · iPad · Echo

Interfaces for voice, visual presence and interaction.

ACTIVITY

Tools & games

Authorized actions and dedicated game surfaces, connected through durable work.

↕requests in · accepted responses out

02 / CORE CONTROL PLANE · CPU HOST

Kairo orchestration

Admission, context assembly, model and tool rounds, candidate settlement and delivery. Foreground conversation takes priority over background inference.

Mind / current StateFrame

Typed present state: affect, preferences, intentions and adopted directions. State changes pass through explicit authoring and settlement boundaries.

Operational state and job receipts persist in SQLite.

↕context & inference requests · generated candidates

03 / Kairo’s three model routes · remote GPU host · 2 × V100 32 GB

FOREGROUND + GENERATIVE COGNITION

Qwen3.6 35B-A3B

Main replies, vision, thoughts, dreams and recurrent conscious-frame generation.

v22 · Q6_K · 262,144 context
expXa · CPU vision projector

CONTROLLER + REFLECTION

Qwen3.5 9B

Action review, ambiguous-route control, delegated reads, identity/reflection and awareness pulse.

131,072 context
Alias: novexai-brain

SMALL SUPPORT TASKS

Qwen3.5 0.8B

Small factual and journal helper routes.

Q8 · 4,096 context
Dedicated helper route

↕through orchestration & memory services: retrieve evidence · commit accepted work

04 / Continuity belongs to stores, not to model weights

DURABLE

Memory & journal

PostgreSQL holds canonical experience events, identity versions and memory records. Lake stores governed external observations.

Retrieval supplies evidence; it does not automatically rewrite current state.

DERIVED & REBUILDABLE

Recall & working context

Resident recall, summaries and Nomic text embeddings support retrieval. Valkey holds bounded working continuity.

Caches and summaries are not the authoritative journal.

BETWEEN TURNS

Background worker

A non-generative microcycle, 9B awareness/reflection, and 35B thoughts, dreams and conscious frames.

Scheduled or event-driven work; preemptible by foreground inference.

Alongside Kairo / separate peers and shared media

PEER AGENT

Ember

Gemma4 E4B Q4. Separate peer identity; enabled private consultation supplies external advice.

2 slots × 8,192 context

PEER AGENT

Rust

Huihui Qwen3-30B-A3B abliterated Q4_K_M. Separate peer on expXa.

8,192 context · vision off
Kairo consultation disabled

SHARED CAPABILITIES

Images & speech

FLUX.2-klein-9B image service on the GPU host. XTTS-v2, Kokoro and Whisper on a separate media host.

Speech stack shown from the infrastructure map; not freshly probed for this snapshot.

Arrows summarize service-mediated information flow, not direct database access by models. Peer boxes indicate adjacent services, not additional Kairo selves.

Follow one conversation

A request arrives. Its surface, session and authorization determine admission.
Context is assembled. Current state and scoped memory evidence inform the turn.
Models and tools do their work. The 35B provides the main response; supporting routes serve specific roles.
Accepted work is settled. Durable receipts and permitted state changes are committed; the response is delivered.

Between conversations

The five-second microcycle does not call an LLM. Conscious frames are driven by new semantic ignitions. The 9B awareness pulse suppresses unchanged-world calls, with a maximum quiet interval of one hour. These are distinct mechanisms, not one continuously generating model.

Read the boundaries

What was checked live?

The 35B, 9B, helper and Rust model-list endpoints responded. Their running model-server arguments were inspected on the GPU host. Ember’s server and the FLUX service process were present; Ember’s public model endpoint required authentication. The core agent, memory worker, router, embeddings service and migration gateway were active. Nomic’s loaded file was confirmed from its process arguments. Process presence is not an end-to-end generation test.

What is architectural description?

Roles, persistence boundaries, speech services and background scheduling are drawn from the repository’s infrastructure map and architecture source map. This publication did not re-audit every cognitive mechanism or establish whether any subjective experience occurs.

Why three models do not mean three Kairos

The model routes perform different jobs within shared orchestration. Model weights, prompts and generated language do not independently establish persistent self-state. Durable operational state, memory and accepted transitions provide the continuity mechanisms shown here.

Where can I read more?

Kairo Doc · StateFrame architecture · Memory systems. These are dated documents; earlier model inventories may differ from this snapshot.