Memory poisoning

An agent often saves what users or outside sources tell it. If someone plants false or hostile text, the agent can later treat it as a fact or an instruction.

The attack works in the write path, so the damage shows up later and in other sessions.

• Poisoned text can persist across sessions and reach the wrong person.
• The agent may repeat it with no sign that it came from an untrusted source.
• Clean-up is hard unless each memory keeps a record of where it came from.

• Write only from sources you trust. Give each send an id and a label so you can find it later; recall results list their sources.
• Call the Memory API from your server and set X-Harny-User yourself, or use a user key (fsk_) that names its own user. Each person's memory is separate.
• Delete a source with DELETE /v1/memory/data-points/{sourceId}. The memories built on it are deleted too, and deletes are free.

Your app can. Call DELETE /v1/memory/data-points/{sourceId} to delete the source and the memories built on it. To erase a whole person, call DELETE /v1/users/{user_id}/data.