Core concepts

Read this once after Your first hour. Come back when a later chapter uses a word you do not recognise. This page is the mental model, not a screen-by-screen guide.

EYAS is not a single chatbot window. It is a personal AI operating system: named agents, durable memory, a work board, automation, and multi-channel I/O on your machine.

Building blocks

Concept	What it is	Where in UI
Agent	Named AI actor with model, tools, skills, voice, workspace files, optional channels	Agents
Colleague	Primary or team agent you talk to — home thread, SOUL. Setup creates two (Personal Assistant + System Engineer).	Sidebar Colleagues / Agents
Specialist	Narrow worker from a shared pool; any colleague can spawn an enabled one (`run_specialist`)	Agents (tier: Specialist)
Conversation	Message thread with one or more agents; tool calls, runs, context rail	New Conversation / chat
Board card	Trackable work item; often linked to a conversation	Board
Project / stage	Delivery structure; conversations can sit on stages	Projects
Skill	Reusable markdown procedure pack agents can load	Skills
Tool	Invokable capability (shell, browser, APIs, MCP…) with permissions	Tools / agent config
Memory	Hybrid recall: working → episodic → semantic/procedural → archive + vault files	Memory
Knowledge page	Explicit wiki page you edit (not automatic memory)	Knowledge
Document	Uploaded file indexed for retrieval	Documents
Channel	External inbox/outbox (e.g. Telegram) bound to an agent	Communication
Provider	LLM backend (cloud API, host CLI, or local runtime)	Providers
Prompt chain	master → project-type → project → conversation layers	Prompts / Settings
Security gate	Policy checks before dangerous actions	Security
Forge	Human-approved proposals to evolve agent soul/identity	Forge

Typical flow

flowchart LR
  User --> Conv[Conversation]
  Conv --> Agent
  Agent --> Tools
  Agent --> Memory
  Agent --> Board
  Agent --> Channel[External channel]
  User --> Board
  Board --> Conv

Setup creates owner, primaries, provider
You open a conversation or create a board card
The colleague may use tools/skills, write memory, hand off / spawn a specialist, or answer on a channel
You review results in chat, board, documents, or outbound messages

Agent vs conversation vs card

	Agent	Conversation	Board card
Lifetime	Long-lived configuration	Thread of messages	Work tracking unit
“Who”	Persona + tools + memory	Session of talk	Task state
Change often?	Settings, forge, workspace	Every message	Status, assignee, due date

Memory vs knowledge vs documents

Store	Who writes it	Best for
Memory tiers	System / agents during work	Automatic recall, episodes, procedures
Vault markdown	Import / agents / you / auto-capture after a conversation turn (on by default since 0.8.16-beta)	Long-lived semantic & procedural notes
Knowledge base	You (editor)	Curated wiki
Documents	Upload	PDFs, office files, source dumps

A durable fact stated in chat can become a vault note without anyone asking. Capture runs after the reply is delivered; a failed capture costs a note, never the answer. Details: Memory.

Orchestration (conversation fields)

When chatting you may see controls such as:

Control	Meaning
Model	Which model answers: a fixed model, Auto-routing, or the colleague’s default
Effort	Reasoning depth vs cost/speed. Lists only the levels the conversation’s model offers; Auto inherits (Deep → Max, the colleague’s effort, the delegating conversation’s) or uses the model’s default
Orchestration: Solo	No specialists, hand-offs or team proposals
Orchestration: Auto	Model decides fan-out
Orchestration: Deep	Aggressive multi-agent fan-out; default effort Max

Details: Conversations.

Security mental model

Root owner — human admin account
Master password — encrypts Secrets store
CASL permissions — what each user/agent may do
Security gate — runtime checks on risky tool use
Autonomy flags — how much agents may do without asking

Next reading

Your first hour
Getting started
Agents overview
Teams & delegation
Memory
Architecture pointer (deep technical specs in the repo)