How do I know an agent really remembered a correction?

Before you start

• Check current account access, plan and task requirements.

Use a falsifiable example

Use a synthetic preference such as “draft reports in English”. Correct it to Chinese, open a fresh conversation and ask for a report without repeating the language. Keep a second unrelated task to check that the correction does not leak across scopes.

Inspect each stage

Check the stored record, the retrieved evidence and the resulting action separately. A stale record being retrieved alongside the correction is a conflict to resolve, not extra context to celebrate.

Export, deletion and disconnect are separate product questions. A text file export does not prove complete task-state recovery.

Record the result

1. Save old/new preferences, scope, session boundary and expected behavior.
2. Record the actual next action and any manual reminder.
3. Repeat after restart and test forgetting separately.

Evidence and limits

Official documentation checked

• Source-backed facts plus BotClaw editorial interpretation. No product account, inference or agent performance test was run.

Sources

• https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
• https://docs.openclaw.ai/concepts/memory
• https://help.openai.com/en/articles/20001530-getting-started-with-your-dot

Continue the decision

• Hermes: a practical decision dossier
• Check that a memory update changes the next session