Skip to main content
Memory is what an Agent remembers on its own — as opposed to knowledge, which you author. Memory ranges from very short-lived (a chat session) to very long-lived (facts an Agent has confirmed hold over months). Every Agent can have zero or more memory spaces attached. Each space is a container with its own retention, scope, and content model.

​
The four space types

Wiki

Notes you or the Agent write, retrievable by keyword or topic. Like a team wiki — human-editable, semi-permanent.

Index

Vector-indexed facts. Store any structured or unstructured content; find by similarity.

Session

Per-conversation memory. Auto-summarized as conversations grow. Cleared per session by default.

Evolution

Diagnostic memory — patterns that emerged from successful improvements. Read-mostly; auto-populated.
See Memory spaces for how to pick.

​
Scoping

Every memory read and write carries a scope key. The same Agent can have separate memory per user, per tenant, per conversation. Different runs never see each other’s memory unless you want them to. Scope is a first-class safety property — no accidental leaks across customers. See Scoping.

​
What lives in memory

  • Notes (short pieces of text with metadata).
  • Facts (structured claims about entities).
  • Relationships between entities.
  • Procedural rules (behavioral guards).
  • Session summaries.
Every type is retrievable via the same semantic search interface, so an Agent doesn’t need to know which kind it’s looking for.

​
When to use memory vs. knowledge

Many Flows use both. Knowledge for reference material; memory for personalization and accumulated context.

​
Forgetting

Nothing is permanent unless you mark it so. Memory items can be pinned (never forgotten), have TTLs, be superseded (like knowledge), or be pruned by importance decay. See Forgetting.