Skip to main content
Use memory when your agent needs to remember context across turns, sessions, or long-running tasks. A customer-facing agent might store a user’s preferred language after the first interaction; a research agent might record intermediate findings across a multi-step pipeline. Without memory, every invocation starts from zero. Memory lets an agent write and read information over time. Unlike knowledge stores (reference content you ingest), memory is what an agent accumulates as it works, scoped by how long it should persist.

Key concepts

Memory scopes and TTL

Items expire automatically at the end of their scope’s TTL.

Operations

Create a memory store

string
required
Display name for the store.
string
required
One of TASK, SESSION, AGENT, WORKSPACE.

Response

string
Unique store identifier.
string
Display name.
string
The scope this store uses.
string
ISO 8601 creation timestamp.

Write a memory item

string
required
The information to store. Keep items concise and self-contained.

Search memory

string
required
Search query. Length: 1–10,000 characters.
integer
default:"10"
Number of results to return. Range: 1–100.

Reset a memory store

Clears all items from the store without deleting the store itself.

Delete a memory store

Deleting a store removes all its items permanently.

Errors

Security

Do not store secrets, credentials, or regulated personal data in memory. Long-lived scopes (AGENT, WORKSPACE) persist for up to 1 year and are accessible to any agent with read access in that workspace.

Scope selection guide

Use for scratch data that should not outlive the current task: intermediate results, step outputs, flags set during execution. Automatically cleaned up after the task TTL expires, so you never accumulate stale data from past runs.
Use for anything that needs to persist across multiple user turns within a session window: conversation history, user preferences stated in this session, in-progress context. Expires when the session ends or after 8 hours, whichever comes first.
Use when an agent should accumulate knowledge over days, weeks, or months: learned user preferences, recurring patterns, notes the agent builds up over time. Scoped to this agent — other agents cannot read it.
Use for facts that every agent in the workspace should know: team conventions, shared ontologies, cross-agent context. All agents with data-platform:memory:read in this workspace can access it.

Rate limits

Knowledge Stores

Reference content versus accumulated memory.

Context Bundles

Combine knowledge and memory for a run.

Knowledge & Memory

Conceptual overview.

Orchestration

How memory supports running tasks.

Next steps

Continue to Connectors.