Knowledge

Agents can read from and write to a persistent knowledge store using the Memory workflow node. This enables retrieval-augmented generation (RAG) and long-term personalization across executions.

Memory Node Modes

The Memory node operates in two modes:

Mode	Trigger	Effect
Recall	Before an LLM node	Retrieves relevant context and injects it into the LLM prompt
Learn	After an LLM node	Queues an episode for async indexing into the vector store

Recall

In Recall mode, the Memory node performs a hybrid search — combining dense vector similarity with sparse keyword matching — against the workspace’s knowledge store.

Query input comes from execution variables, typically user_message from the Trigger node.

Retrieved documents are formatted and injected as a system-message segment before the LLM call. The number of retrieved documents and similarity threshold are configurable per node.

Learn

In Learn mode, the Memory node queues an episode for async indexing. The episode typically contains the user message and the agent’s response.

Indexing happens in the background — the execution does not wait for it. This keeps latency low for interactive agents that learn from every turn.

Scope

Memory is scoped at the workspace level. All agents in a workspace share the same vector store. Use execution variables to filter retrieval by customer ID or tag if you need per-user knowledge isolation.

Knowledge Sources

Beyond agent-generated episodes, you can index documents via the admin UI:

Plain text and Markdown
PDF files
Web pages (URL fetch + parse)

All sources are chunked, embedded, and stored in the platform’s vector index.

Workflows — Memory node in the full node type table
Create an Agent — add Memory nodes to a workflow
Intelligence — the embedding model used for indexing