Most of the "AI memory" conversation is about getting things in: chunking, embeddings, which vector database, how to compress a long chat into something you can recall later. Almost none of it is about what happens when one of those memories is wrong on purpose. That is the part we cannot stop thinking about while building a local-first personal memory layer, because an assistant that remembers you is also an assistant that can be talked into remembering a lie.
Key takeaways
Do not try to detect a poisoned memory. Gate what an unproven memory is allowed to do.
"Two independent sources" is the retrieval spelling of an old, unsolved problem: permissionless membership. Independence is not free.
Standing earned from outcomes beats standing bought with identity, but the agent cannot grade its own homework.
