The risk with always-on agents isn't that they remember too much. It's that you often can't correct what they remember.
OpenAI's new dots are a clean example. Their privacy FAQ says you currently can't view, delete or modify individual dot memories. To make a dot forget something, you delete the dot. Disconnecting an app doesn't remove what the dot already built into its context. That part is straight from their docs.
My read: once memory has no per-item edit path, a stale or wrong fact can't be fixed in place. The only reset is starting over, which gets more expensive the longer the agent has been useful. If you're shipping agent memory, "show me what you know about X" and "forget this one thing" belong in v1, not on the roadmap.
I compared dots, Muse, Team Bots and Claude Managed Agents on where memory lives, who can edit it, and what deletion actually removes:
https://pub.towardsai.net/the-memory-problem-openais-dots-makes-impossible-to-ignore-525e93363d64
Top comments (0)