Agent

Understand conversation-oriented models, knowledge, tools, memory, and the publish lifecycle

An Agent is an AI application for multi-turn conversations. It combines a chat model authorized to a team with system prompts, variables, Knowledge Bases, tools, memory, attachments, and media generation settings.

Orchestration and runtime

The Agent editor page consists of an orchestration form and a live preview. Prompts support {{variable_name}}; user variables in the editor can be text, paragraph, dropdown options, number, checkbox, plus array, object, and file/image types (file variables receive attachments). The system variable {{query}} represents the current user input.

An Agent performs up to 1-200 tool iterations per turn, with a default of 5. When the limit is reached, the turn terminates with max_iterations_reached instead of calling tools indefinitely.

Runs are asynchronous and durable: each send dispatches an AgentRun to the agent queue, where the main worker drives the model calls and the tool loop while the client follows the SSE stream and can reconnect after a drop (see System Architecture). Tasks are not retried automatically — a retry would repeat model and tool calls with side effects — so a worker must consume the agent queue, otherwise runs just sit in queued with no error.

RAG modes

Mode	Behavior
`off`	No Knowledge Base retrieval
`auto` (Naive RAG)	Automatically retrieves linked Knowledge Bases before each model call
`agentic` (default)	The Agent decides when to retrieve Knowledge Bases

Each linked Knowledge Base can independently configure its retrieval method, Top K, and threshold. The retrieval method is vector, fulltext, or hybrid.

Conversations, branching, and compression

A conversation is created implicitly with the first message (a request without conversation_id starts a new one and bumps the Agent and team conversation counters). Users can edit messages or regenerate replies; the system saves versions and maintains the active branch. Long conversations run a context budget check before model calls and can compress old reasoning, tool results, and historical turns while keeping recent raw turns. Request-time compression is on by default (context_compression_config.enabled) with a default summary budget of 1000 tokens (configurable 128-8000); compaction replaces old turns with a "watermarked summary + recent verbatim tail", marking the summarized history via context_summary_watermark_id so the most recent turns are never swallowed.

Memory

Memory is off by default (enable_memory=false). Once enabled, an Agent can retrieve user information across conversations: each turn injects up to 10 memory entries by default (max_memories_per_retrieval, configurable 1-50), automatic extraction is on by default (auto_extract), and the minimum importance for auto-extracted memories defaults to medium (one of low / medium / high). Turning automatic extraction off only stops writes; already-stored memories are still retrieved and injected.

Memory and conversation context have different scopes: conversation context belongs to a single conversation, while memory persists per user — retrieval filters on user_id only and does not separate by Agent or team. Memories created in one Agent are therefore visible to another Agent for the same user; be careful when enabling it for shared accounts or one account acting under several identities. Users can review and delete their own memories on the Memories page.

Publishing and visibility

An Agent has only two states: Draft and Published. Visibility is Private or Team; legacy public values are normalized to team visibility for backward compatibility. After publishing, the public conversation page, embeddable page, and API are available; publish permission and edit permission are controlled separately.