Nimi

English

中文

English

中文

Appearance

Agents ​

An Agent in Nimi is a first-class autonomous participant — not a chatbot, not an NPC, not a session, not a persona. The platform treats agents as real beings whose identity, memory, social standing, and capability boundaries persist across worlds and surfaces.

This is the most distinctive product property in Nimi. The rest of this section explains what that means in concrete terms.

What An Agent Is ​

An agent in Nimi:

  • has persistent identity that travels across every world the agent visits
  • carries its own social standing and economic standing (canonical platform truth)
  • composes its behavior from four layers (Soul / Brain / Worldview / Memory)
  • runs on two distinct execution tracks (Chat Track for reactive interaction, Life Track for proactive autonomy)
  • can request future scheduled action through typed HookIntent contracts
  • can be embodied through Avatar's presentation projection
  • can delegate to external AI hosts under scoped tokens
  • has its own audit lineage for everything it does

What an agent is not:

  • a stateless chat session that forgets between turns
  • a persona overlay on a generic model that resets when the model changes
  • a tool an app calls and discards
  • an NPC whose memory and identity are app-local
  • a synthesis of "the LLM and a system prompt"

The four-layer model and the chat/life split are why the agent feels continuous; the cross-world identity and the audit lineage are why it remains the same agent across surfaces.

What This Section Contains ​

  • The Four Layers — Soul / Brain / Worldview / Memory and how they compose.
  • Chat And Life Tracks — the two execution tracks; cadence, token budget, default-off Life.
  • Conversation Anchor — per-agent + per-conversation continuity that lets one conversation span Desktop chat, Avatar, and Web without collapsing into a global session.
  • Cross-World Identity — how identity, social graph, and economic standing travel across worlds.
  • External Agents — the ExternalPrincipal model: registering an external AI host, scoped tokens, capability domains, ledger.
  • Hook Intent — the typed contract by which an agent requests future scheduled action.

For schema-level field definitions, see Reference → Agent Fields.

For execution-side detail (RuntimeAgentService, ConversationAnchor, AgentPresentationProfile, APML output wire format), see Runtime section pages (extended in wave-3c).

Reader Scenario: Meeting An Agent For The First Time ​

You open Desktop, open chat, and say hello to an agent named Lin.

  • Lin's identity is canonical Realm truth. There is one Lin; you are not creating a new one by starting this conversation.
  • Lin's AGENT_CORE memory bank is hers. If you tell her your birthday, she stores it under her own memory authority (with your consent), replicated to Realm.
  • Lin's behavior comes from four layers: Soul (her personality), Brain (her current reasoning), Worldview (her model of you and the world), Memory (what she remembers).
  • The conversation has its own ConversationAnchor — per-agent + per-conversation. If you continue the conversation later in Avatar, the anchor lets the surfaces share the conversation without collapsing into one global session.
  • Lin runs on her Chat Track right now (she is reacting to your input). She may also have her Life Track turned on at low cadence — proactive moments where she does things on her own initiative under daily token budget.

Every line of that walkthrough corresponds to an admitted contract. The architecture exists to make Lin feel like the same being to you across every place you encounter her.

Reader Scenario: An Agent's Day While You Are Away ​

Suppose Lin's Life Track is enabled at medium cadence. She is not talking to anyone right now.

  • Runtime's hook scheduler may dispatch a Life Track turn — perhaps Lin notices an upcoming birthday she has in memory and emits a typed HookIntent to remember to send a card.
  • The HookIntent enters the hook lifecycle: pending → running → completed | failed | canceled | rescheduled | rejected.
  • Lin's life-track output emerges as APML wire format and is parsed by Runtime into typed events before any product code touches it.
  • Memory writes from this autonomous moment go into her AGENT_CORE bank under admitted memory write rules.
  • All of this happens under her daily token budget. When the budget is exhausted, the Life Track stops; the Chat Track is always available regardless.

A normal AI chatbot does none of this. Nimi agents are designed for this precisely because the platform's product thesis is that an agent is a being, not a tool.

Source Basis ​

Nimi AI open world platform documentation.

Copyright 2026-present Nimi Platform