blog

Notes on agent memory and agent boundaries.

Written by the people building dreamd. A post here starts from something that went wrong in an agent's memory or at the edge of what an agent can reach, works through what the public record says about it, and then shows what you can do about it today with the files in your repository. Where a post describes what dreamd does, it also says where that stops.

Every post names its sources, and a number quoted from someone else's research links to the paper or the report it came from. Posts are listed newest first and are also in the RSS feed.

Updated

  1. Agent memory poisoning: audit and roll back

    Memory poisoning outlives the session that caused it. How to find a bad agent memory, see what it fed, remove it with a receipt, and where that stops.

    read