> ## Documentation Index
> Fetch the complete documentation index at: https://docs.formation.texturehq.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent perspectives

> How each agent keeps its identities, conversations, and knowledge separate.

A perspective is the set of people, conversations, and facts known to **one Formation
agent**. Hosting two agents in the same process does not give them a shared mind.

For example, if Scout and Atlas both see the same Slack thread, each records its
own observation. Scout may be addressed while Atlas merely observes. Each can
have different memories about the same person, and neither can read the other's
private conversations.

## Agent identity and participant identity

A **RuntimeAgent** is an agent the operator configures: its name, persona,
models, connector installs, and capabilities. An **Actor** is a participant as
known within that agent's perspective. These are separate concepts.

Actors do not carry a canonical human/agent/service classification. Connector
metadata may identify a bot for transport behavior such as reply-loop prevention,
but that is not a global identity or a link to another RuntimeAgent's private
configuration.

[Identity linking](/runtime/identity-linking.md) can associate one person's accounts
inside a perspective after proof of possession. Names alone do not link accounts,
and two agents establish their links independently.

## Knowledge has an audience

Perspective isolation establishes whose knowledge is available. It does not mean
every fact the agent knows is suitable for every conversation.

A preference learned privately remains subject to its source audience, even when
its subject joins a public channel. [Memory privacy](/runtime/memory-privacy.md)
applies those restrictions before retrieval and again before delivery. Summaries
and agent replies retain the restrictions of the sources they used.

## The operator boundary

The dashboard and operator API manage agents and inspect canonical records.
They are trusted administrator surfaces, not private end-user accounts. Their
ability to inspect several agents is not permission to copy those records into
model context. See [access control](/access-control.md) for the current trust boundary.

## What is not shared or implemented

There is no deployment-wide conversational directory, shared memory pool, or
implemented organization-profile container. Agents currently represent
organization facts through ordinary attributed memories and conversations.
Proposed organization and WorkGraph models must preserve the same perspective
boundary if implemented later.

For the physical entities, read the [data model](/data-model.md). For what a model
receives, read [context resolution](/runtime/context-resolution.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.