Skip to main content
Memory stores durable facts that agents can recall across threads. Thread-level conversation history is handled separately by the runtime checkpointer; memory is for long-lived knowledge.

​
How it works

Each memory is embedded on save. Recall embeds the query and returns the closest matches by cosine similarity. Memories are workspace-scoped, with optional per-agent scoping.

​
From the CLI

​
From an agent

Two built-in tools:
  • memory_search - recall relevant facts for a query.
  • memory_write - persist a new fact.
Reference them from an agent node to give the agent read/write memory.

​
API

  • POST /v1/memory - store a memory
  • POST /v1/memory/search - recall