Agents assume that their context is an append-only log.
- Agents assume that their context is an append-only log.
- But it's not.
- It's possible to change earlier things in the conversation.
- If you do, you break the caching, but it otherwise still works.
- This can lead to weird outcomes that get the LLMs confused.
- You can play with their memories.
- For example, if you have a versioned AGENTS.md, each time Claude.app boots up, it seems to put the current version of it in the conversation at any point the agent asked before, which can make the agent get very confused.