Give Your Agents Long-Term Memory

When to Use Long-Term Memory

Use long-term memory for stable context an agent should remember across future conversations and runs. Good memories include team preferences, recurring customer-handling rules, reporting formats, escalation paths, product terminology, and constraints the agent should apply every time.

Do not use memory for one-time task details, temporary drafts, passwords, API keys, private credentials, or information that changes frequently. Put short-lived context in the current prompt, template, file, or connected app instead.

Add a Memory

Open the agent that should remember the information, then add the memory from the agent memory area or ask your Super Agent to add it for that agent. Write the memory as a durable instruction or fact, not as a vague note.

Include the scope and trigger for the memory. For example: "For support summaries, group issues by billing, access, bug, and feature request, then list urgent items first."

Keep each memory small enough to review later. Several focused memories are easier to edit and delete than one long paragraph that mixes unrelated rules.

Edit a Memory

Edit a memory when a team process, preferred format, customer policy, or product detail changes. Replace the outdated rule with the new rule instead of adding a second memory that conflicts with it.

If the agent keeps making the same mistake, inspect the latest trace, the step-by-step run log, first. Then update the memory with the missing source, exception, approval rule, or output format that would have prevented the mistake.

Delete a Memory

Delete memories that are no longer true, no longer useful, too broad, duplicated, or based on sensitive information the agent should not retain.

After deleting a memory, rerun a small task that used to depend on it. This confirms the agent no longer applies the old rule and still has enough context from its instructions, prompt, or connected tools.

Safety and Privacy

Treat memories as workspace context that may influence future agent behavior. Only store information your workspace is allowed to retain and reuse for the agent role.

Avoid storing secrets, raw personal data, payment details, health information, or confidential customer records in memory. Keep sensitive source data in the system of record and let the agent retrieve only what it needs through approved integrations and permissions.

Use approval rules for actions that could affect customers, billing, public posts, or production systems. Memory can guide the agent, but it should not replace human review for sensitive decisions.

Verify the Memory Works

Run a narrow test prompt after adding or editing a memory. Ask the agent to complete a low-risk task where the memory should clearly change the answer or format.

Review the response and open the trace if the behavior is unclear. Confirm the agent used the expected instruction, did not apply it outside its intended scope, and asked for approval before sensitive actions.

If the memory should support a reusable prompt or scheduled job, test the manual run first. Then update the template or schedule only after the agent follows the memory consistently.

Troubleshooting

If the agent ignores a memory, make the memory more specific, remove conflicting memories, and rerun the same prompt so you can compare behavior.

If the agent applies a memory in the wrong situation, add scope such as the workflow, customer segment, channel, or output type where the memory applies.

If the agent uses outdated information, edit or delete the old memory and confirm the connected app or source of truth has the current data.

If a memory appears to expose sensitive context, delete it, move the data back to the appropriate system of record, and review who can access the agent.