Documentation#

These manuals describe the current source tree. Tagged releases retain the documentation at their own revision. Start with the architecture, then build and operation. Application developers can continue with the HTTP and shared service guides.

Choose a task#

Task	Guide
Build and start a first instance	Build, then operation
Check supported model families	Model families
Use an existing GGUF file	Model files
Connect a standard chat client	HTTP gateway
Build a persistent conversation client	Shared service
Enable source-linked memory	Live memory, then semantic memory
Copy an instance to another machine	Complete runtime files
Inspect a failed run	Operation and journal format
Check a capability before enabling it	Support and qualification

Start and operate#

Manual	Contents
Architecture	State ownership, execution and component boundaries.
Build	Requirements, card profiles and capacity settings.
Operation	Start, inspect, stop and restore an instance.
Settings	Startup options and values that affect later turns.
Console commands	Agents, tools, models and memory commands.
Terminal	Attach, screens, keys and terminal settings.
Graphical control	Local instance and conversation windows.
Model files	Inspection, catalogs, wrappers and model selection.
Support and qualification	Current feature, model and hardware boundaries.
Testing	Source gates, focused checks and runtime acceptance.

Connect applications and media#

Manual	Contents
HTTP gateway	Deployment, credentials and standard chat requests.
Native service resources	Request handles, events, capabilities and broker framing.
Shared service	Participants, spaces, persistent conversations and retries.
Image input	Image models, source formats and device capacity.
Audio input	Audio models, source formats and combined media capacity.

Use persistent memory#

Manual	Contents
Live memory	Bind conversations to scoped typed state.
Text requests	Prepare queries through the embedding model.
Input retention	Save exact source text, vectors and working focus.
Automatic retention	Retain accepted input and its recorded selection.
Semantic memory	Qualified interpretation, source spans and corrections.
Contextual recall	Task, participant, evidence and source-group selection.
Automatic appraisal	Supported benefit, harm and relationship evidence.
Task reviews	Task-scoped outcome cues with exact supporting evidence.
Checkpoints	Live durability, checkpoint pressure and resume.
Maintenance	Reclamation, retry state and file shrinking.
Cold memory	Explicit offload, retrieval and portable cold extents.

Extend and configure behavior#

Manual	Contents
Modules	Role, skill and tool installation with effective tool selection.
Tool SDK	Batched CUDA tools and trusted host tools.
Creator policies	Data or native scheduling rules and compatible updates.
Control bindings	Exact model identity, accepted doses and application selection.
Conduct controls	Vector authoring, calibration and local conduct commands.
Affect and quality streams	State equations, probes, actuation and recovery.
Quality measurement	Task scoring, paired conversations and fixture interpretation.

Read formats and measurements#

Manual	Contents
Time and recovery	Ticks, authoritative records and failure semantics.
Host-device boundary	Ring layouts, publication and mapped memory.
Journal format	Segments, records, payloads and derived files.
Message bus	Message kinds, provenance and JSON line output.
CCIR container	Bounded file transactions, sections and data-state manifests.
Complete runtime files	Pack, activate, copy and recover complete instances.
Typed state	Object records, payload schemas and configured limits.
Prepared recall	Query records, selection references and exact replay.
Historical measurements	Version-bound performance and control measurements.
Architecture accuracy	Fixed numerical references, unresolved results and reproduction.
Writing conventions	Terminology, sentence structure and source references.

Shared terms#

Term	Meaning
Runtime	One running AOTX instance on one GPU.
Profile	Build-time table capacities for a class of GPU.
Slot	A bounded device row used by an agent or sequence.
Tick	One ordered execution of the device scheduling graph.
Seam	The host-device boundary, crossed through bounded mapped transport.
Ring	A single-producer, single-consumer buffer that carries records or payloads.
Journal	Ordered records used to recover the last complete durable tick.
Drain	The disk program that writes and synchronizes journal blocks.
Feeder	The disk program that publishes input and completed host-tool results.
Replay	Reapplication of recorded authoritative inputs and decisions.
Mirror	A display snapshot or disk copy; it does not own live state.
Catalog	The device registry of imported roles, skills and tools.
Model store	Disk files and manifests that identify model assets.
Arena	A contiguous memory region with offset-based allocation.
Pump	Host glue that launches the device graph and services its transport.
Source	An exact retained input with identity, owner, scope and version.
Appraisal	A source-backed interpretation of benefit, harm or relationship evidence.
Qualification	Acceptance for an exact model, processor or control package and declared behavior.
Principal	An authenticated service identity with explicit grants.
CCIR	The container format for typed state or a complete runtime package.

Conventions#

Commands run from the repository root unless a guide gives another directory. Uppercase paths and IDs are placeholders. Replace them with the intended local values. File offsets are byte offsets. Binary layouts state their byte order and schema version.

A successful build establishes compilation. A passed structural check does not establish a model's semantic behavior. Read testing and support before applying measured results to another configuration. Security describes trusted files, executable modules and service credentials.