Skip to main content
Memory is what makes an agent get better over time instead of starting from zero on every message.

​
What memory holds

Memory keeps small, useful facts an agent learned during a run and might need again:
  • User preferences it discovered (“this account prefers email over SMS”)
  • Corrections you gave it (“we call this product ‘Pro Plan’, not ‘Premium’”)
  • Repeated context (“this customer’s org has three sub-accounts”)
  • Patterns from past problems it solved
Memory is per Agent and per scope — a support agent doesn’t accidentally use a research agent’s memory. Scope typically follows the natural boundary of the conversation: a user, a session, or an account.

​
What memory is not

Memory is not a chat log. It’s the distilled things worth keeping — not every message. Memory is not your knowledge base. Company documents belong in a DataSource, not in memory. Memory is per-interaction learning; DataSources are shared reference.

​
Inspecting memory

Open any Agent block and click Memory. You’ll see:
  • What’s currently remembered, grouped by scope
  • Where each fact came from (the run that added it)
  • When it will fade (older, unused facts naturally decay)
You can delete any memory entry manually — useful if a fact is wrong or a customer requests it be forgotten.

​
When memory is used

Every time the Agent runs, it pulls relevant memory for the current scope and includes it as context. This happens automatically — you don’t write any glue code.

​
Turning memory off

Some Flows don’t want memory. In the Agent block settings, set Memory to Off and each run starts fresh.